Quick Read

• Digital English is English used inside search, messaging, collaborative documents, platforms, interfaces, multimodal media and online research; AI English adds systems that can generate or transform language before the learner has necessarily built the underlying capability.
• The central distinction is tool-assisted performance vs independent capability. A polished output does not prove the learner can reproduce the reasoning, writing, verification or source use when support is removed.
• Use the right tool for the question: search for discovery, dictionaries/corpora for language verification, primary sources for authoritative facts, AI for provisional generation/critique where appropriate, and humans for judgement that requires responsibility, context or care.
• Every AI-assisted English workflow should protect authorship, source verification, privacy, ownership boundaries and transfer.
• Verification should match consequence. Low-stakes brainstorming needs less checking than an examination claim, medical statement, official rule, research citation or published factual assertion.
• Strong AI use keeps the learner cognitively active: compare, critique, justify, revise, verify and then perform a fresh task without the tool.
• Because digital and AI systems change quickly, tool-specific guidance must be dated and reviewed; the durable principles are evidence, transparency, learner ownership and independent transfer.

Use this owner when the controlling question is how English changes inside digital or AI-assisted work. Use Reference Literacy to verify language and usage, Evidence, Data & Multimodal English to verify claims and mixed media, and Academic English when the task is university study or research.

Digital and AI English: language when the interface can read, write and answer back

Digital English is not simply English typed on a screen, and AI English is not simply English produced by a chatbot. Digital environments change how people search, read, write, collaborate, revise, cite, verify, present identity and decide what deserves attention. Generative AI changes the system again because the interface can now produce fluent language, summaries, explanations, arguments, images, code, questions and feedback before the learner has necessarily built the underlying capability.

That creates a new educational problem. A student can submit a polished paragraph without being able to construct the argument independently. A reader can receive an instant answer without knowing which sources support it. A writer can improve style while quietly losing ownership of the reasoning. A researcher can search faster and still build a weaker evidence set if generated summaries replace source-level reading. A learner can receive unlimited feedback and become more dependent on feedback.

The key distinction is therefore between performance with a digital system and capability that remains when the system is removed.

OECD’s 2026 Digital Education Outlook makes this distinction unusually explicit: general-purpose generative AI can improve students’ immediate task performance without necessarily producing learning gains, while educational uses designed around clear pedagogical purposes can support more durable learning. UNESCO’s student AI-competency framework emphasises human-centred judgement, ethics, understanding, application and creation. In September 2026, the Council of Europe adopted a recommendation on AI literacy that places human dignity, agency, privacy, equity, critical thinking and democratic participation alongside technical understanding.

SETC uses those developments to distinguish learning from immediate task performance. Evaluate assistance against the stated learning or participation goal. When independent English performance is the goal, look for learner-owned decisions on fresh tasks under the permitted conditions. When access or AI-assisted communication is the goal, evaluate those capabilities directly rather than treating all continued tool use as failed learning.

This is SETC’s guide to Digital & AI English. It maps digital reading, search literacy, hyperlink and interface reading, AI-assisted writing, prompt design, source verification, hallucination risk, authorship, citation, privacy, bias, digital genres, multimodal communication, platform rhetoric, AI feedback, tutoring, learning transfer and the human skills that remain essential when language can be generated instantly.

It connects several existing owners. Corpus, Dictionary & Reference Literacy verifies language usage. Evidence, Data & Multimodal English verifies claims, charts, images and provenance. Rhetoric & Persuasion explains attention, framing and platform influence. Academic English owns research and scholarly communication. This page owns the digital/AI environment in which all of them now operate.

The first rule: product quality is not learning evidence

A learner uses AI to improve an essay and the final draft becomes clearer. What has been demonstrated?

At least one thing: the combined human-plus-AI system produced a better document.

What has not yet been demonstrated?

• that the learner can identify the weakness independently;
• that the learner understands why the revision is better;
• that the learner can reproduce the improvement on a new task;
• that the learner can detect when the AI revision is wrong;
• that the learner can perform without the tool.

This is the core diagnostic distinction for AI-era education.

The Digital & AI English operating loop

Need → search or prompt → retrieve/generate → inspect → verify → compare → decide → transform → publish/use → reflect → retest without support.

Every arrow can fail.

A weak search query can retrieve poor sources. A fluent answer can invent evidence. A correct answer can be copied without learning. A verified paragraph can still violate authorship rules. A useful draft can leak sensitive information. A good explanation can be accepted passively instead of retrieved later.

The system therefore needs both digital literacy and language-learning discipline.

Digital English existed before generative AI

AI did not create digital language. Email, search engines, forums, blogs, messaging, social media, collaborative documents, online video, subtitles, hyperlinks, comments, digital forms and learning platforms had already changed English before generative AI became widely accessible.

Those environments created new genres and conventions:

• subject lines;
• search queries;
• usernames and handles;
• hyperlink anchor text;
• threaded replies;
• reaction language;
• hashtags;
• captions;
• interface microcopy;
• FAQ structures;
• collaborative comments;
• version histories.

Generative AI enters that existing ecology rather than replacing it.

The medium changes the English job

Writing a search query is not the same genre as writing a question to a teacher. Writing a message inside a team chat is not the same as writing an academic paragraph. A prompt to an AI system is not the same as the final document produced from the interaction.

Strong digital writers understand which language is for:

• retrieval;
• instruction;
• collaboration;
• public communication;
• record-keeping;
• persuasion;
• learning;
• verification.

Genre still matters even when all of it happens on one device.

Search literacy is language literacy

A search engine does not receive the learner’s intention directly. It receives language. Query formulation therefore shapes the evidence landscape.

Compare:

good vocab words essay

with:

verbs for qualifying causal claims academic writing corpus examples

The second query specifies the language relationship and likely evidence type more clearly.

Search queries are compressed problem representations

Good search often depends on identifying:

• the key concept;
• the target population;
• the relevant date;
• the evidence type;
• the official authority;
• the language variety or register;
• the exact phrase where wording matters.

Search is iterative

The first query can reveal terminology needed for the second. Search literacy includes reformulation rather than expecting one perfect query.

Discovery and verification are different

A search engine is excellent for discovering candidate sources. The search-result ranking is not itself proof that the top result is most authoritative.

AI literacy is not prompt cleverness

Prompting matters because instructions shape output. But a learner who can produce elaborate prompts while being unable to verify the result is not yet AI-literate.

A durable AI-literacy model needs at least:

• understanding what the system can and cannot be assumed to know;
• clear task formulation;
• critical evaluation of output;
• source verification;
• privacy and data judgement;
• authorship awareness;
• bias awareness;
• ability to revise or reject output;
• ability to perform the underlying task independently where learning is the goal.

AI is a sociotechnical system

The Council of Europe’s 2026 AI-literacy recommendation explicitly frames AI as more than software: people, institutions, policies and social practices determine how systems are designed and used. That is important for English education.

A chatbot response is shaped by model design, training, interface rules, user prompt, conversation context and deployment environment. The English on screen is therefore part of a larger system.

The human-centred rule

UNESCO’s AI competency framework places a human-centred mindset and ethics alongside technical understanding and creation. SETC translates that into a language-learning rule:

The learner remains the final owner of meaning, evidence, judgement and responsibility.

AI can assist formulation. It cannot assume the learner’s responsibility for whether the claim is true, whether the source is represented honestly, whether the language is appropriate or whether the submission follows the governing rules.

Digital & AI English is a fast-moving Atlas territory

The tools will change faster than the deep principles. SETC therefore separates:

• stable principles: verification, authorship, audience, evidence, transfer, privacy, source ownership, independent capability;
• changing practices: interfaces, model capabilities, institutional rules, platform conventions and available tools.

This owner should be reviewed more frequently than foundational grammar or reading pages, but its constitutional principles should remain stable.

Search literacy: finding information is a language-and-evidence task

Search is often treated as a technical skill—type a few words, scan the results, click. In reality, strong search depends on English comprehension, concept formation, vocabulary, source judgement and iterative reasoning.

A learner who does not understand the question cannot formulate a strong query. A learner who lacks the vocabulary of the field may not know the terms experts use. A learner who cannot distinguish an official specification from commentary may retrieve information but still fail to verify it.

Search literacy therefore belongs inside English learning, not outside it.

Step 1: define the information need

Before searching, decide what kind of answer is needed:

• a definition;
• a current fact;
• a historical fact;
• an official rule;
• research evidence;
• a language example;
• a comparison;
• a local service;
• a quotation;
• a statistic;
• a current price or schedule;
• an explanation of mechanism.

The information need determines the best source family.

Step 2: identify the source owner

For a current examination format, search for the examining authority. For a word meaning, use a strong dictionary. For a language-frequency question, use a corpus. For an institution’s policy, use the institution.

This applies the same owner principle used across the SETC Atlas: do not ask a secondary page to own a primary fact when the primary authority exists.

Step 3: build the query

A useful query can include:

• core concept;
• specific population;
• source type;
• date or period;
• country or institution;
• technical term;
• exact phrase in quotation marks;
• domain name where the authority is known.

The query should be specific enough to reduce noise but broad enough to discover better terminology.

Step 4: learn from the results page

Search results are not just answers. They reveal the vocabulary of the field. If several authoritative results use a term you did not know, that term may improve the next query.

Search therefore loops:

query → vocabulary discovery → refined query → stronger source set.

Step 5: separate navigation from research

Sometimes the goal is simply to find the official page. That is navigational search. Other times the goal is to understand a contested or complex question. That requires multiple sources.

Do not over-research a navigation problem, and do not treat one navigational result as a complete evidence review.

Step 6: inspect the result before clicking

Titles, domains, publication dates and snippets provide clues. They do not prove quality, but they can help prioritise:

• official organisations;
• primary documents;
• recent sources for changing claims;
• specialist sources for technical questions;
• stable canonical references for foundational concepts.

Step 7: read the source, not only the snippet

A search snippet is generated for retrieval. It can omit qualifications or combine text from different parts of a page. Do not cite or reason from the snippet when the full source is available.

Step 8: triangulate where the claim matters

One source may be enough for a simple official fact. A contested claim, research conclusion or high-stakes decision may require several independent sources.

Search-query language

Keyword queries

Keyword queries are efficient for discovery:

English collocation corpus academic writing

They omit ordinary sentence grammar because the genre is retrieval.

Natural-language queries

Modern search and AI interfaces increasingly accept full questions. Natural language can be useful when the relationship among terms matters.

Exact phrases

Quotation marks can search a phrase more precisely. This is useful for verifying quotations, formulaic expressions or exact wording.

Alternative terms

If one query fails, search synonyms or technical equivalents. A student who searches only AI cheating school may miss research framed through academic integrity, assessment validity or generative AI use.

Negative and domain filters

Search systems may support ways to restrict results to a domain or remove unwanted meanings. These operators change over time, so tool-specific syntax should be checked rather than memorised as permanent English knowledge.

Search results are rhetorically designed

Result pages rank, label and present information. Position creates salience. Titles compete for attention. Sponsored material can resemble ordinary results. Rich snippets can answer a question without the user opening the source.

The Rhetoric & Persuasion owner explains why presentation changes judgement. Search literacy applies that understanding to retrieval.

Ranking is not truth

A result can rank highly because it matches the query, has strong authority signals, is popular or satisfies many users. Ranking does not prove the claim inside the page.

Freshness is query-dependent

For a current syllabus, software feature, law, price or AI policy, recency matters. For the definition of a clause or a historical text, recent publication may be less important than authority and scholarship.

Searchers need to know which facts can change.

Personalisation and localisation

Search results can differ by location, language, account state and inferred intent. This means one person’s result page is not necessarily a universal ranking.

Digital reading: the page is not always the text

Printed reading usually presents a bounded object. Digital reading often presents a network. A webpage can contain navigation, sidebars, comments, advertisements, embedded media, hyperlinks, interactive charts and related articles.

The learner must decide what is part of the central text and what is interface.

Navigation literacy

Headings, menus, breadcrumbs, tables of contents and search boxes reveal information architecture. Strong digital readers use them to predict where information belongs.

Hyperlink literacy

A link has at least three components:

• anchor text;
• destination;
• relationship to the current page.

Useful links tell the reader why another page matters. SETC’s internal-link rule follows the same principle: links should express prerequisite, mechanism, evidence, application, practice or handoff.

Open-web reading

When a learner opens several tabs, the reading problem becomes comparative. Sources can conflict, repeat one another or cite the same underlying report.

Count independent evidence, not tab count.

Source loops

Online articles often cite one another. A claim can appear in ten pages but originate from one weak source. Follow the citation chain toward the primary evidence.

Publication date vs event date

A current article can report an old event. A new summary does not make the underlying evidence new.

Digital readers should track both.

Versioning

Webpages can change after publication. Policies and product documentation may be updated. When exact wording matters, record the date checked and use archived or versioned sources where appropriate.

Dynamic content

Interfaces can personalise or generate content dynamically. Two users may not see identical pages. This complicates citation and verification.

Reading comments and community discussion

Comments, forums and social posts provide evidence of user experience, language attitudes and practical problems. They are not automatically representative.

Community evidence is strongest when the question is about community experience. It is weaker for technical claims that have official specifications or controlled research.

Digital source hierarchy

A practical hierarchy for many questions is:

1. primary/official source for the controlled fact;
2. strong specialist or research source for explanation;
3. reputable secondary synthesis;
4. community experience where experience is relevant;
5. search/AI output as discovery rather than final authority.

The hierarchy can change with the question. A Reddit thread may be useful primary evidence about what Reddit users report experiencing. It is not an official specification.

Link literacy and citation

A hyperlink is not the same as a citation. A link gives access; a citation identifies the source relationship.

In academic or formal work, the writer may need author, date, title, publication and other details even when the reader can click.

Link rot

Web links disappear or move. Important digital research should preserve enough bibliographic information to identify the source beyond one URL.

URL credibility

A domain can provide clues but not certainty. Attackers can imitate names, use misleading subdomains or create professional-looking pages.

Read the actual domain and verify important organisations through known routes.

Reading PDFs, images and screenshots

A screenshot is not necessarily the original source. It can be cropped, old or edited. A PDF can be official, scanned, republished or modified.

Check provenance:

• where did this file come from?
• who published it?
• is there an official version?
• what date?
• does the page number or section match?
• is the screenshot complete enough to preserve context?

Digital note-taking

Copy-and-paste makes note collection easy and source confusion easy.

Keep copied text visibly marked as quotation. Store the source link and date. Separate your interpretation from imported text.

The Academic English owner uses the same principle for source notes because digital convenience does not change intellectual ownership.

Search summaries and AI summaries

Summaries reduce reading time by selecting. Selection can remove the qualification that controls the conclusion.

Use summaries for orientation. Verify load-bearing claims in the source.

Snippet dependence

If a learner can answer only from snippets, they may develop a habit of shallow extraction rather than model-building. Train full-source reading for complex questions.

Digital skimming

Skimming is a legitimate strategy when the task is orientation. It becomes a problem when the reader never switches into close reading at load-bearing evidence.

Tab overload

Opening more sources can become avoidance. Use a source-stop rule: another tab must add authority, mechanism, evidence, counterposition or updated information.

Digital reading under interruption

Notifications and links can fragment the passage model. Long-form comprehension benefits from periods of uninterrupted reading.

The solution is not moral panic about screens. It is deliberate attention management.

Prompting: instruct the system without surrendering the task

A prompt is a piece of functional English. It specifies a job, context, constraints, audience, format and sometimes evidence standard. Strong prompting therefore draws on the same English systems SETC already teaches: clarity, reference, genre, audience, scope, sequencing and evaluation.

But prompting is not the final educational goal. A learner who can prompt an AI to write an excellent paragraph may still be unable to write the paragraph independently. Prompting should therefore be evaluated according to the task.

Prompting for production

If the real-world job is to use AI as a tool, prompt quality matters directly. The user should be able to define:

• the purpose;
• the audience;
• the source material;
• the desired output;
• constraints;
• what must not be changed;
• what must be verified;
• what uncertainty should remain visible.

Prompting for learning

If the goal is learning, the prompt should preserve useful cognitive work for the learner.

Compare:

Write my essay on whether AI improves education.

with:

Ask me five questions that will help me define what “improves education” means. Do not write the thesis for me. After I answer, show me two weaknesses in my reasoning and ask me to revise.

The second prompt uses AI as a tutor rather than a substitute author.

The prompt should reflect the learning objective

If the learner is practising inference, do not ask the AI to infer for them. If the learner is practising sentence editing, it may be useful for AI to identify a problem but not rewrite every sentence. If the learner is practising source synthesis, AI can challenge the synthesis but should not replace source reading.

Prompt specificity

Specific prompts can improve output, but excessive specification can also hide the learner’s own weak task model. If the tutor writes the perfect prompt for the student, the student may succeed without understanding the task.

Teach learners to build prompts from the same questions they would use to plan independently.

Role prompts

“Act as a teacher” or “act as an editor” can influence output style, but role labels do not create expertise automatically. A system can imitate the language of an expert while generating weak or unsupported content.

Role prompting should never replace evidence verification.

Context windows

AI systems operate on available context. If important source material, definitions or constraints are missing, the answer can be based on incomplete information.

Users should learn to ask: What does the system currently have access to?

Prompt ambiguity

Ambiguous prompts produce ambiguous or over-assumed answers. If the user writes “make this better”, the system must infer what “better” means.

Useful dimensions include:

• clearer;
• shorter;
• more formal;
• more precise;
• more persuasive;
• better supported;
• more age-appropriate;
• more concise without losing qualification.

Define the improvement dimension.

Prompt constraints

Constraints can protect meaning:

Preserve the source’s uncertainty. Do not add claims. Keep technical terms unchanged. Explain every major edit.

These instructions make the editing process inspectable.

Prompt iteration

AI interaction is often iterative. The user can ask for clarification, correction, alternatives or a stronger critique.

Iteration is useful when each turn reflects judgement. It becomes dependence when the user keeps asking “make it better” until the system produces something acceptable without learning what changed.

AI-assisted reading

AI can support reading by explaining vocabulary, unpacking syntax, generating questions, summarising sections or comparing interpretations.

The educational risk is that explanation arrives before the learner attempts meaning.

Vocabulary assistance

A learner can ask AI to explain a word in context, compare senses or generate examples. The answer should still be checked against a dictionary when exact usage matters.

Syntax assistance

AI can split a dense sentence into propositions or identify clause relationships. This can be valuable scaffolding for academic reading.

After explanation, the learner should reconstruct the original sentence independently.

Summary assistance

AI can summarise long texts quickly. The summary can orient the reader but should not replace source-level reading when the exact claim matters.

Question generation

AI can create comprehension or retrieval questions. The learner should evaluate question quality: does the question test the intended skill or only surface recall?

Interpretation comparison

A learner can ask AI to produce two plausible interpretations of a passage and then test each against textual evidence.

This preserves the learner’s evaluative role.

AI-assisted writing

Writing involves task interpretation, idea generation, selection, structure, formulation, revision and editing. AI can enter at any of these stages. The educational effect depends on which stage the learner outsources.

Brainstorming

AI can expand the possibility space by suggesting angles or questions. The learner should decide which ideas are relevant, defensible and original enough for the task.

Planning

AI can propose outlines. A learner should inspect whether the outline actually answers the task and whether each section has a necessary job.

One useful exercise is to compare three AI outlines and explain why none should be accepted unchanged.

Drafting

Full-draft generation creates the greatest authorship and learning risks. If the assessment is intended to measure the learner’s writing, generated prose can invalidate the evidence.

Where AI-assisted drafting is permitted in real-world work, the user still owns accuracy, source representation and final meaning.

Revision

AI can be powerful during revision if it helps the learner see choices:

• identify an unclear reference;
• flag an unsupported leap;
• compare two versions;
• explain why one sentence is more concise;
• locate repetition;
• suggest where qualification is needed.

The learner should make the revision or at least justify accepting the suggestion.

Editing

Grammar and style tools can correct errors quickly. But automatic correction can conceal recurring weaknesses.

For learning, record error patterns. Ask the system to classify rather than merely fix. Then write a new sentence using the same relationship.

Rewriting

“Rewrite this better” can change voice, meaning and claim strength. Strong instructions specify what must remain.

For example:

Rewrite for clarity without changing the factual claim, level of certainty or source attribution. Explain any change that affects meaning.

Tone transformation

AI can transform a message from informal to formal or warm to concise. This is a useful register exercise if the learner compares which linguistic features changed.

The Register and Tone owners provide the underlying language system.

Vocabulary enhancement

Requests such as “make this more sophisticated” can generate unnatural or inflated diction. Ask for precision instead.

A better prompt is:

Identify vague words and suggest alternatives only where the replacement adds a more precise distinction. Check collocation and register.

Paraphrase

AI can produce fluent paraphrases that subtly alter source meaning. If the source matters, compare proposition by proposition and preserve attribution.

Translation

AI can translate across languages rapidly. Technical terms, cultural concepts and pragmatic meaning still require verification.

The Multilingual English & Mediation owner supplies the translation principles.

Feedback

AI can provide immediate feedback at scale. Feedback quality varies and the system may apply generic criteria or invent rubric interpretations.

Give the actual rubric when permitted and ask the system to separate:

• observed evidence;
• inference;
• suggested change.

Then verify the feedback against the task.

AI as tutor

A tutor does more than provide answers. A good tutor diagnoses, selects the next task, adjusts difficulty, waits, asks questions, notices error patterns and reduces support over time.

AI can simulate some of these behaviours, but the user or teacher must design the interaction carefully.

Socratic prompting

Ask the system to question rather than answer:

Do not give me the solution. Ask one question at a time to help me identify the evidence gap.

Hint ladders

Provide progressively stronger hints only after the learner attempts the task.

Error diagnosis

Ask the learner first to explain why the answer may be wrong. Then use AI to compare the explanation with likely failure points.

Retrieval practice

AI can generate varied questions from a concept. The learner should answer without notes, receive feedback and later retest on fresh material.

Spacing

AI can help create spaced review schedules, but actual scheduling discipline matters more than the sophistication of the generated plan.

Variation

Generate new examples that preserve the target mechanism while changing surface features.

This supports transfer better than repeated practice on one familiar example.

Interleaving

Ask AI to mix several problem types so the learner must decide which method applies.

Reflection

After a task, ask:

• What did you decide?
• What evidence did you use?
• Where did AI help?
• What could you now do without it?

The AI assistance ladder

SETC can order assistance from lower to higher cognitive substitution:

1. Environment support: timing, formatting, reminders.
2. Retrieval support: questions, flashcards, examples.
3. Metacognitive support: planning questions, self-check prompts.
4. Diagnostic support: identify possible errors or gaps.
5. Explanatory support: explain a concept or relationship.
6. Revision support: suggest ways to improve existing learner work.
7. Planning substitution: AI designs the structure.
8. Formulation substitution: AI writes sentences or paragraphs.
9. Task substitution: AI completes the assessed task.

The higher the substitution, the weaker the evidence about independent learner capability unless the task is specifically about AI-assisted production.

The release test

After AI-assisted practice:

1. remove the AI;
2. change the topic;
3. change surface wording;
4. delay the task;
5. reduce prompts;
6. measure the same target capability.

If performance survives, the assistance may have supported learning. If quality collapses, the system was carrying more of the task than the learner.

AI and learner confidence

AI can make learners feel more capable because the output looks better. Confidence should be calibrated against independent performance.

Overconfidence is dangerous when the learner stops checking. Underconfidence is also possible when AI output appears impossibly polished. Show learners that fluent output is cheap; judgement remains valuable.

AI and voice

Repeated AI rewriting can homogenise language. A learner may lose awareness of their own syntactic and stylistic choices.

Protect voice by asking AI to diagnose rather than replace, compare alternatives and preserve distinctive phrasing where it is clear and appropriate.

AI and creativity

Creativity can involve generating possibilities, combining ideas, selecting, transforming and refining. AI can increase idea quantity. The human creative job increasingly shifts toward selection, direction, evaluation and integration.

For learners, however, idea generation itself can be a skill worth practising. Do not outsource every blank page.

AI and writer’s block

AI can reduce initiation friction. Use it to ask questions, generate constraints or propose directions. Then require the learner to create the actual material.

AI and language models as examples

Generated examples can be useful, but they should not be assumed to represent frequency, authenticity or community norms. Verify with corpus or reference evidence where that matters.

AI and explanation quality

An explanation can be fluent and pedagogically poor. Check:

• Does it define the concept?
• Does it show the mechanism?
• Does it distinguish nearby concepts?
• Does the example genuinely illustrate the rule?
• Does it state boundaries?
• Can the learner apply it afterward?

Clarity is necessary but not sufficient.

Verification: fluent language is not evidence

Generative AI can produce confident, coherent and grammatically polished language even when the underlying claim is incomplete, unsupported or false. This is not an unusual edge case to treat as a curiosity. It is a structural reason that AI literacy must include verification.

A fluent sentence answers one question: Can the system produce language that looks usable?

Verification answers another: Is the claim supported by the right evidence for the decision we are making?

Deep application route: use How Digital Source Verification Works for the full source-level workflow across lateral reading, claim tracing, image/video context, provenance, authenticity and AI-generated content. This Digital & AI English owner keeps the wider system; the verification page carries the focused application.

Hallucination is a source problem, not only a factual-error problem

An AI system can invent:

• citations;
• authors;
• publication titles;
• statistics;
• quotes;
• laws;
• technical details;
• product capabilities;
• historical claims;
• URLs;
• institutional rules.

Even when the factual sentence happens to be correct, an invented source remains a serious failure because the evidence trail is false.

The verification rule

Verify the load-bearing claim at the source level.

If a paragraph depends on one statistic, verify the statistic. If an argument depends on one quotation, verify the quotation in context. If an AI system says an examination changed, verify the current official specification. If a generated language rule sounds plausible, verify it against the appropriate dictionary, grammar, corpus or style authority.

Source provenance

Provenance asks where information came from and how it reached the current representation.

A claim may travel:

research paper → news article → blog → social post → AI summary → student essay.

Each transformation can introduce simplification or distortion. Strong digital readers move backward toward the source owner when the claim matters.

AI citations

Never assume a generated reference exists. Check:

• author;
• title;
• journal or publisher;
• year;
• DOI or stable identifier where relevant;
• whether the source actually contains the claimed evidence.

The last step is essential. A real paper can still be cited for something it does not say.

Quotation verification

AI can produce plausible quotations or alter wording. Search the exact wording in the primary source where possible. Preserve the surrounding context needed to interpret it.

Statistics

Ask:

• What is the source?
• What population?
• What year?
• What denominator?
• What measure?
• Was the statistic directly observed or estimated?
• Does the AI wording preserve uncertainty?

The Evidence, Data & Multimodal English owner supplies the deeper evidence framework.

Current facts

AI models and interfaces can have knowledge cutoffs, retrieval limitations or stale information. Current facts about laws, product features, exam formats, prices, schedules or public policy should be checked against current sources.

Language rules

AI often converts usage variation into rigid rules. A model may say “never use X” because one style tradition discourages it. Reference literacy tests the claim across dictionary, corpus, grammar and target context.

Source comparison

When two sources disagree, do not ask AI to choose by confidence. Compare source scope, authority, date, population and evidence.

Verification depth should match consequence

Not every casual use requires a research audit. Verification effort should scale with stakes.

Authorship: who made the intellectual decisions?

AI complicates authorship because several kinds of contribution can be distributed across human and machine.

A system may suggest ideas. The learner may select them. The system may draft. The learner may revise. The system may correct grammar. The learner may verify. The final document can therefore contain several layers of contribution.

Institutional rules differ, so students must follow the rules governing the actual task. But the deeper educational question remains: Which decisions does the learner own?

Idea ownership

If AI supplies the central thesis, structure and examples, the learner’s contribution may be primarily selection and editing. That can be appropriate in some real-world workflows and inappropriate in an assessment designed to measure independent argument.

Language ownership

If AI rewrites every sentence, the final language may no longer be good evidence of the learner’s productive English.

Evidence ownership

Even when AI proposes sources, the learner remains responsible for verifying and representing them accurately.

Decision ownership

The strongest educational use keeps final decisions with the learner:

• What claim should I make?
• Which source should I trust?
• Which revision preserves my meaning?
• Which uncertainty matters?
• What should I reject?

Disclosure

Some institutions require disclosure of AI assistance; others restrict certain forms of use. Disclosure rules change and vary by context.

A durable principle is transparency: if the rules require disclosure, state the assistance accurately enough that the reader can understand its role.

Academic integrity

Academic integrity includes:

• honest authorship;
• accurate source attribution;
• no fabrication;
• no false citation;
• compliance with assessment rules;
• truthful representation of process where required.

AI creates new ways to violate these principles, but it does not change the principles themselves.

AI detectors

Automated systems that claim to detect AI-generated writing can produce errors. Institutions should not treat a detector score as self-sufficient proof of authorship.

From a teaching perspective, stronger evidence comes from process: drafts, notes, oral explanation, revision history and the learner’s ability to reproduce the reasoning.

Process evidence

In the AI era, assessment can place more value on evidence of thinking:

• source notes;
• planning;
• draft progression;
• reflection;
• oral defence;
• unseen transfer task;
• explanation of key decisions.

These artefacts can make authorship and learning more visible without turning every task into surveillance.

Privacy

Users can expose personal, confidential or proprietary information when interacting with AI and digital platforms.

Before submitting material, ask:

• Does this contain personal data?
• Does it contain another person’s private information?
• Is it confidential school, workplace or client material?
• Does the service policy allow this use?
• Do I have permission to upload it?
• Can I remove identifying details?

Children and privacy

Schools and parents should be especially careful about student identity, images, voice recordings, health data, assessment records and other sensitive material.

Data minimisation

Give the system only the information needed for the task. Replace names with placeholders where possible.

Bias

AI output can reproduce biases present in data, evaluation processes or user prompts. Bias can appear in:

• stereotyped examples;
• uneven representation;
• assumptions about names, accents or identities;
• default cultural references;
• tone judgements;
• language-variety correction;
• recommendations;
• generated images.

World Englishes awareness is especially important. A model may “correct” legitimate regional English toward a dominant standard without explaining the choice.

Bias audit

Ask:

• Whose English is treated as default?
• Which examples repeat stereotypes?
• Are different names or identities receiving different assumptions?
• Is a local variety being labelled incorrect without context?
• Are there missing perspectives?

Fairness and accessibility

AI can lower barriers by providing explanations, translation, text-to-speech, speech-to-text and adaptive support. It can also create new inequalities if access, device quality, paid features or digital literacy differ.

Education should evaluate both capability and access conditions.

Copyright and intellectual property

Users should not assume that AI output is free of rights issues or that uploading copyrighted material is always permitted. Rules vary by jurisdiction, platform, licence and use.

For English education, the practical rule is to respect source ownership, quote within permitted boundaries, attribute where required and follow institutional policies.

Confidentiality

A workplace document can be linguistically harmless and institutionally confidential. Language tools do not remove confidentiality obligations.

Safety and inappropriate content

Digital literacy includes knowing that AI systems can produce harmful, biased or age-inappropriate content and that platform safeguards vary.

Learners need strategies for stopping, reporting or seeking adult support when content becomes unsafe.

AI-generated images, audio and video

Generative AI can create convincing multimodal material. English literacy now intersects with provenance literacy.

When media is used as evidence, ask:

• Who created it?
• Can origin be verified?
• Was it edited?
• Does metadata exist?
• Is there an original?
• Do independent sources confirm the event?

Deepfakes

Synthetic media can imitate people or events. A realistic appearance should not be treated as proof of authenticity.

Transcripts

AI-generated transcripts can mishear names, numbers, accents or technical vocabulary. Verify important passages against audio.

Captions

Automatic captions increase accessibility but can introduce errors. The error rate may vary by accent, audio quality and domain vocabulary.

Provenance signals

Technical standards and platform labels may help indicate how digital media was created or edited, but no single signal is infallible. Provenance information should be one part of verification.

AI confidence is not calibrated confidence

Language systems can present weak claims in the same polished tone as strong claims. Users should not interpret tone as probability.

Ask for evidence, not confidence theatre.

AI self-correction

Asking a model “Are you sure?” may cause a revision, but the revision can still be wrong. Self-critique is useful as another hypothesis, not as independent verification.

Cross-model agreement

Two AI systems agreeing does not equal two independent sources if they rely on overlapping data or reproduce the same widespread error.

Generated consensus

AI may summarise “what experts think” without showing how experts were selected. For contested questions, inspect real sources and current consensus documents.

The human verification checklist

1. What is the central claim?
2. What evidence would support it?
3. Which source owns that evidence?
4. Can I access the source?
5. Does it actually say what the AI says?
6. What population/date/context?
7. What uncertainty?
8. What changed through paraphrase?
9. What institutional rules apply?
10. Would I defend this claim without the AI wording?

Digital genres: English changes when communication becomes networked, persistent and interactive

Digital communication is not one genre. A private message, support ticket, collaborative comment, social post, forum reply, newsletter, product notification, search query and AI prompt all use English under different constraints.

Strong digital writers identify the genre before choosing tone.

Email

Email is asynchronous but persistent. It often carries more institutional weight than chat.

Effective email design includes:

• a subject line that identifies the job;
• an opening that matches relationship and formality;
• the decision, request or update early enough to find;
• context proportional to the reader’s need;
• clear actions and deadlines;
• attachment references that actually match attached files;
• a close that fits the relationship.

AI can draft email quickly, but the sender remains responsible for tone, facts, commitments and unintended implications.

Subject lines

A subject line is retrieval metadata. “Question” is weak. “Approval needed by 3 Oct: Secondary English workshop venue” tells the recipient what kind of action awaits.

Subject lines are a digital form of information architecture.

Chat and messaging

Chat rewards speed, context and turn-taking. Complete formal sentences are often unnecessary. But ambiguity becomes expensive when a short message controls action.

Use explicit language for:

• dates;
• times;
• ownership;
• deadlines;
• changes;
• decisions.

Threading

Replying inside the correct thread preserves context. Starting a new message can separate an answer from the question it resolves.

Digital literacy includes information placement, not only wording.

Reactions

Emoji and reaction buttons can acknowledge, agree, celebrate, soften or signal completion. Their meaning depends on community norms.

For high-stakes decisions, do not rely on an ambiguous reaction where explicit confirmation is needed.

Comments in collaborative documents

A useful comment identifies the problem and invites the next decision:

The claim is stronger than Source 3 supports. Could we change “causes” to “is associated with”, or add causal evidence?

“Unclear” is weaker because the writer does not know what to repair.

Track changes and revision history

Collaborative tools preserve process evidence. This is valuable for writing pedagogy because teachers can see development, not only final polish.

Revision history can also help establish authorship in AI-rich environments without requiring invasive surveillance.

Forum posts

Forums combine public writing with conversational interaction. Strong posts quote enough previous context to remain understandable, distinguish fact from experience and avoid assuming that one community’s norms are universal.

Community Q&A

Answers should identify whether they are based on documentation, personal experience or inference. This increases epistemic transparency.

Social posts

Social platforms compress context. A sentence may circulate far beyond the audience for which it was written.

Writers should anticipate:

• screenshots;
• quotation without context;
• algorithmic amplification;
• audience mixing;
• rapid emotional response;
• limited correction visibility.

Caption writing

A caption can frame an image before the viewer interprets it. Captions should not imply facts the image cannot establish.

Hashtags

Hashtags can classify, rally, signal identity or join a public conversation. They can also collapse distinct claims into one slogan.

Short-form video text

On-screen text and subtitles compete with visual and spoken information. Keep the cognitive load manageable.

Long-form digital articles

Digital longform benefits from headings, navigation, explicit definitions and purposeful links because readers can enter from search at any point and skim before committing.

SETC’s longform architecture should support both sequential reading and retrieval.

Newsletters

A newsletter creates an ongoing relationship. Subject lines, preview text, headings and link selection shape what readers notice.

Push notifications

Notifications are an extreme compression genre. They can inform, remind or manipulate attention. Urgency should reflect real urgency.

Interface microcopy

Buttons, labels, errors, confirmation messages and onboarding text guide user action.

Compare:

Submit

with:

Submit application

The second reduces ambiguity.

Error messages

A useful error message tells the user:

• what happened;
• what remains safe;
• what they can do next.

“Something went wrong” is low-information language.

Consent language

Digital systems often ask users to agree to terms, tracking or data use. Consent language should not hide consequence behind vague labels.

Dark-pattern rhetoric belongs in the Rhetoric owner; Digital English applies it to interfaces.

Platform rhetoric

Platforms are not neutral containers. Their design affects what language becomes visible and rewarded.

Character limits

Limits encourage compression and can reduce qualification. Writers may choose a stronger simpler claim because nuance does not fit.

Ranking metrics

Likes, views and shares create social proof. High engagement does not establish accuracy.

Recommendation systems

Recommendation systems can repeatedly expose users to similar language, narratives and sources, creating a false sense that one view is universal.

Virality

Content that triggers strong emotion can spread rapidly. Virality is a distribution property, not an evidence property.

Thumbnail and headline optimisation

Headlines and thumbnails can maximise curiosity or outrage before the source is read. Readers should compare the headline with the article’s actual claim.

Context collapse

One post can reach friends, teachers, employers and strangers simultaneously. Register decisions become difficult because the audience is unstable.

Persistence

Digital messages can survive after the speaker’s view changes. Correction and deletion do not guarantee disappearance.

Searchability

Public language can be retrieved years later by keywords. Writers should distinguish private conversational habits from durable public records.

Digital collaboration

Collaborative writing changes authorship. Several people can plan, edit and comment on the same document. AI may become another participant.

The team needs explicit rules:

• who owns final factual verification;
• who approves claims;
• which AI uses are allowed;
• how sources are stored;
• how revisions are accepted;
• how disagreements are resolved.

Shared terminology

Teams should define key terms early. Digital collaboration makes it easy for different writers to use the same word differently across sections.

Comment overload

Too many comments can hide priority. Label feedback by level: structure, evidence, source, paragraph, sentence, formatting.

Simultaneous editing

Real-time editing is efficient but can create collision. Decide when to draft separately and when to merge.

Version control

Save meaningful versions when major conceptual changes occur. “final-final-v7” is a symptom of weak version naming.

AI in collaborative documents

If AI produces or revises text inside a team document, contributors need to know which parts require verification and who takes responsibility.

Digital identity

English choices contribute to digital identity. Usernames, bios, posts, comments and public replies create a record of voice and affiliation.

Professional and personal boundaries

Different platforms blur these boundaries. A teacher’s public account, student portfolio and private group chat may have different audiences but can become connected.

Profile language

Short bios are high-compression identity texts. Every noun and verb carries weight.

Public disagreement

Digital disagreement is easily amplified. Write to the proposition, not the imagined character of the person.

Quote-posting

Quoting another person to criticise them can send attention or harassment toward the original speaker. Rhetorical responsibility includes distribution effects.

Digital civility

Civility is not avoidance of disagreement. It is a communication discipline that protects the possibility of continued reasoning.

Useful moves include:

• quote the exact claim;
• state what you agree with;
• identify the precise disagreement;
• provide evidence;
• avoid identity insults;
• correct publicly when your public claim was wrong.

Online harassment and abuse

English education should teach learners to recognise targeted abuse, dogpiling, threats and coercive language. The correct response may be to disengage, block, report, preserve evidence or seek support rather than “argue better”.

Digital humour

Memes, irony and sarcasm travel quickly but depend on shared context. They can be misread by outsiders or future audiences.

Emoji literacy

Emoji can modify tone, substitute for reaction or carry community-specific meanings. Interpret them as pragmatic resources rather than universal pictograms.

Digital punctuation

Punctuation can acquire interactional meanings in messaging. A full stop can feel final; repeated punctuation can signal intensity; ellipses can suggest hesitation or implication.

These meanings are community-dependent and should not be confused with formal punctuation rules.

Digital code-switching

Users may shift languages or varieties across platforms and audiences. Multilingual digital communication accelerates borrowing and hybrid language.

The Multilingual English & Mediation owner explains the broader repertoire.

Accessibility

Digital English should consider users who rely on screen readers, captions, transcripts, keyboard navigation or simplified layouts.

Alt text

Alt text describes the information function of an image for users who cannot see it. The best description depends on why the image appears.

A decorative image may need little or no descriptive content. A chart needs the relationship the reader must recover.

Captions and transcripts

Captions support Deaf and hard-of-hearing users, language learners and people in noisy environments. Transcripts support search and review.

Plain language

Plain language can make interfaces and instructions more accessible without reducing conceptual accuracy.

Multimodal composition

Digital writers combine words with charts, screenshots, audio, video and interaction. Composition therefore includes deciding which mode should carry which meaning.

Do not make every mode repeat the same content. A chart can carry the pattern while prose explains significance.

Screenshot annotation

When explaining an interface, annotations should point to the real control and remain current when the interface changes.

Video scripting

Video scripts should account for what the viewer can see. Spoken language can explain relationships rather than describe every visible object.

Podcast scripting

Audio requires stronger verbal signposting because the listener cannot see structure.

Digital English transfer

A learner should be able to move the same communication core across:

• email;
• chat;
• forum;
• presentation;
• social post;
• longform page;
• AI prompt.

The content remains related; genre decisions change.

AI and learning: assistance should make the learner stronger after assistance ends

The deepest educational question is not whether AI can help complete a task. It can. The question is what part of the capability remains with the learner afterward.

OECD’s 2026 Digital Education Outlook synthesises emerging evidence that general-purpose generative AI can improve task performance while failing to produce durable learning gains when cognitive work is simply offloaded. That distinction matches SETC’s transfer architecture exactly: performance supported by a tool is not the same as independent capability.

The learning equation

A useful SETC formulation is:

Assisted performance + explanation + retrieval + variation + delayed retest + support removal = evidence of learning.

Without the later stages, we know only that the assisted system performed well.

When AI helps learning

AI can support learning when it:

• increases useful practice;
• provides timely feedback;
• generates varied examples;
• asks questions instead of supplying answers;
• helps the learner compare options;
• makes hidden relationships visible;
• adapts difficulty;
• supports retrieval and spacing;
• reduces irrelevant friction while preserving the target cognitive work.

When AI weakens learning

Risk increases when it:

• completes the target operation;
• removes productive struggle too early;
• lets the learner skip source reading;
• replaces idea generation every time;
• revises without explanation;
• creates polished language that the learner cannot reproduce;
• encourages passive acceptance;
• removes the need to remember or retrieve.

Productive struggle

Learning often requires effort because effort reveals what is not yet automatic. This does not mean every task should be difficult. It means assistance should not remove the exact operation being learned.

If the target is thesis formation, allow struggle with thesis formation. If the target is sentence editing, do not auto-correct the sentence before the learner sees the problem.

Scaffolding with AI

AI scaffolds can be faded deliberately.

A sequence might be:

1. AI models one example.
2. Learner explains what the model did.
3. AI gives a partially completed example.
4. Learner completes it.
5. AI gives a checklist only.
6. Learner performs independently.
7. Fresh delayed task without AI.

The scaffold is successful when it becomes unnecessary.

Feedback design

AI feedback should prioritise diagnosis over correction.

Instead of:

Here is the corrected paragraph.

use:

Your paragraph has one unsupported causal leap and one unclear pronoun reference. Identify both before I suggest revisions.

The learner performs the repair.

Feedback hierarchy

AI feedback can follow SETC’s existing levels:

1. task interpretation;
2. reasoning/evidence;
3. structure;
4. paragraph;
5. sentence;
6. grammar/vocabulary;
7. formatting.

Do not spend AI capacity polishing punctuation while the argument answers the wrong question.

Retrieval practice with AI

AI can generate questions, but the learner should answer before seeing the model response.

For vocabulary:

• definition from memory;
• collocation;
• contrast with near-synonym;
• new sentence;
• delayed retest.

Variation and transfer

AI is especially useful for creating surface variation:

• new passage;
• different topic;
• new genre;
• new audience;
• new vocabulary context;
• different evidence format.

Variation should preserve the same underlying skill while preventing memorisation of one example.

Interleaving

AI can mix grammar, inference or writing problem types so the learner must diagnose which method applies.

That is harder than blocked practice and often more useful for transfer.

Delayed retesting

Immediate correction can create an illusion of mastery. Retest later, after memory of the solution has faded.

Unseen-task performance

The strongest evidence is a fresh task that does not resemble the prompt history too closely.

AI and assessment

Assessment becomes difficult when AI can perform the same visible product as the learner.

Schools need to ask what the assessment is intended to measure.

If the construct is independent writing

AI-generated drafting should be restricted or removed during the assessment because it would replace part of the construct.

If the construct is AI-assisted professional communication

AI use may be part of the task. Assessment should then measure prompt design, verification, judgement, revision, source use and final accountability.

If the construct is reasoning

Oral defence, process notes and unseen follow-up can reveal whether the learner owns the argument.

If the construct is research

Assess source selection, verification and synthesis, not only polished prose.

Assessment redesign

AI does not make assessment impossible. It makes construct definition more important.

Useful redesign options include:

• in-class unaided components;
• oral explanation;
• process logs;
• draft comparison;
• source verification tasks;
• AI critique tasks;
• new-task transfer;
• personalised evidence requiring course-specific knowledge;
• reflection on why revisions were accepted or rejected.

Do not redesign everything into surveillance

Attempts to prove authorship can become invasive. Stronger pedagogy reduces the need for suspicion by making process visible through ordinary learning activities.

A teacher who has seen the learner plan, discuss, draft and revise has richer evidence than an automated detector score.

Three-student tutorials in the AI era

SETC’s three-student model becomes more valuable when AI can generate polished individual work privately. The tutorial makes reasoning visible socially.

One student explains the claim. Another challenges the evidence. A third tests the transfer.

Maren may bring an AI-generated paragraph. Iona asks which source supports the central claim. Leonie asks whether Maren can reproduce the reasoning on a new prompt without the paragraph.

The AI output becomes an object of analysis rather than a substitute for thinking.

Shared prompt critique

Give all three learners the same AI prompt and compare outputs or interpretations. Ask:

• What assumption did the prompt contain?
• What did the model add?
• Which answer is better and why?
• What needs verification?

AI debate partner

AI can produce counterarguments, but the students must test them rather than treating them as authoritative.

Error hunt

Generate an explanation containing deliberate or suspected weaknesses. Students identify errors and verify.

Scaffold comparison

One student receives a hint, one receives a worked example, one receives only the task. Compare what each support type does to learning.

Primary learners and AI

Younger learners need stronger adult mediation. Core priorities include:

• distinguish human from generated content;
• understand that AI can be wrong;
• protect personal information;
• ask before uploading work or images;
• use AI for explanation and practice rather than replacing reading/writing;
• retain substantial handwriting, oral explanation and independent composition.

Primary reading

Do not allow AI summaries to replace actual reading. Use AI after reading to ask comprehension questions or explain a difficult concept.

Primary writing

Let children generate ideas orally and write independently before comparing with suggestions.

Primary vocabulary

AI can produce examples, but dictionaries and teacher-selected authentic examples should remain part of verification.

Secondary learners and AI

Secondary students can take more responsibility for verification and authorship.

They should learn to:

• identify when AI completes the assessed skill;
• check citations;
• compare AI claims with sources;
• analyse bias;
• use AI for feedback without accepting every edit;
• maintain process evidence;
• perform unseen tasks independently.

JC and pre-university learners

Advanced students should use AI as an intellectual opponent, research assistant or language tool while preserving source-level reasoning.

For General Paper, useful AI roles include:

• generating counterarguments;
• testing definitions;
• challenging examples;
• simulating an examiner question;
• proposing alternative structures.

Weak uses include generated essays, unverified examples and source-free “knowledge banks”.

University learners

University use depends strongly on institutional rules and discipline.

Students should understand:

• permitted AI use;
• disclosure;
• source verification;
• data privacy;
• research integrity;
• disciplinary conventions;
• how to preserve independent scholarly judgement.

OECD’s 2026 higher-education work highlights these same system-level concerns: reliability, academic integrity, equity, data protection and development of student knowledge and skills.

Teachers and AI

Teachers can use AI for:

• example generation;
• question variation;
• drafting practice materials;
• differentiation ideas;
• rubric-language explanations;
• lesson planning support.

Teacher expertise remains responsible for:

• curriculum alignment;
• accuracy;
• age appropriateness;
• diagnostic interpretation;
• safeguarding;
• assessment validity.

Tutors and AI

A tutor should not use AI to hide weak subject knowledge. The tutor must be able to explain, verify and adapt beyond the generated output.

AI is most useful when it increases variety and reduces administrative friction while human attention remains focused on diagnosis.

Parents and AI

Parents need practical questions rather than panic:

• What is my child using AI for?
• Does the school permit that use?
• Can my child explain the work?
• Can they do a similar task without AI?
• Are they sharing personal information?
• Are they checking sources?
• Is AI replacing reading time?

Recovery from AI dependence

A learner can become dependent when the first response to difficulty is always to ask AI.

Recovery should not necessarily mean banning the tool. Rebuild the missing operations.

Step 1: diagnose what has been outsourced

Ideas? Planning? vocabulary? sentence formulation? source search? editing?

Step 2: remove AI from that operation temporarily

Practise it directly at manageable difficulty.

Step 3: reintroduce AI at a lower substitution level

Use questions, hints or critique instead of full generation.

Step 4: retest without AI

Confirm that capability returns.

AI-free zones can be useful

Some tasks should remain deliberately unaided because the cognitive operation itself is the training target.

Examples include:

• first-draft idea generation;
• retrieval practice;
• oral explanation;
• timed examination practice;
• handwritten planning;
• source-reading sessions.

The purpose is not nostalgia. It is construct preservation.

AI-rich zones can also be useful

Other tasks can deliberately teach AI-assisted performance:

• critique a generated answer;
• verify citations;
• compare prompt versions;
• edit for audience;
• identify bias;
• build a transparent source trail;
• test whether the final output survives human review.

The assessment principle

Assess the capability you actually care about, under conditions that make that capability visible.

AI literacy in 2026: understand, evaluate, use and remain responsible

By 2026, AI literacy is no longer being framed only as “how to use a chatbot”. International education frameworks increasingly treat it as a combination of understanding, critical evaluation, ethical judgement, human agency and practical capability.

UNESCO’s AI Competency Framework for Students organises learning around four dimensions: a human-centred mindset, ethics of AI, AI techniques and applications, and AI system design, progressing from understanding to application and creation. OECD and the European Commission’s 2026 framework for primary and secondary education similarly frames AI literacy as knowledge, skills and attitudes needed to understand how AI works, evaluate outputs critically and use systems ethically and creatively. The Council of Europe’s September 2026 recommendation adds a strong human-rights, dignity, democratic and social-agency dimension.

SETC does not reproduce those frameworks as a separate curriculum. It translates their durable principles into English learning.

Dimension 1: understand the system

Learners should know that generative AI predicts and constructs outputs from patterns rather than consulting a perfect store of truth. They should understand that systems can use retrieval, tools or external data in some settings and not others.

The exact technical architecture can change; the user-level question remains: What evidence does this answer have access to, and what should not be assumed?

Dimension 2: evaluate the output

Critical evaluation includes:

• factual accuracy;
• source quality;
• claim strength;
• missing context;
• bias;
• regional language assumptions;
• genre fit;
• authorship and originality;
• privacy and safety.

Dimension 3: use AI intentionally

The learner chooses whether AI is the right tool at all. Sometimes the strongest action is to read the source, ask the teacher, use a dictionary or think without external assistance.

Dimension 4: create responsibly

When AI participates in production, the human should retain responsibility for the final representation, especially where the work affects other people.

Dimension 5: understand social consequence

AI changes labour, access, information flow, education, authorship and public discourse. AI literacy therefore includes asking who benefits, who is excluded, whose data is used and who is accountable when systems fail.

Digital citizenship

Digital citizenship is broader than online safety. It includes responsible participation in networked communities: communication, rights, responsibilities, privacy, source judgement, respectful disagreement and awareness of how platforms shape behaviour.

Participation

Digital users are not only consumers. Comments, likes, reposts, ratings and prompts can shape what becomes visible and what systems learn from interaction.

Responsibility

Sharing false information after being corrected is different from making an honest initial mistake. Digital citizenship includes correction and repair.

Rights

Users have legitimate interests in privacy, access, expression and fair treatment. These can conflict, so digital citizenship includes judgement rather than one simple rule.

Community norms

Every online community develops language conventions. Joining responsibly requires noticing norms without assuming that community practice overrides wider ethical or legal constraints.

Misinformation, disinformation and uncertainty

Digital English education should distinguish several problems.

Misinformation

False or misleading information shared without necessarily intending deception.

Disinformation

False or misleading information deliberately created or spread to deceive.

Malinformation

Sometimes used for genuine information shared in a harmful or misleading context, such as private information exposed publicly.

The exact classification matters less than understanding source, truth status, intent and harm separately.

Why false information spreads

False information can be:

• novel;
• emotionally intense;
• identity-confirming;
• easy to repeat;
• visually vivid;
• algorithmically amplified.

Rhetorical strength can outrun evidential strength.

Familiarity effect

Repeated claims can feel more familiar and therefore more plausible. Verification should not depend on how often the learner has seen the sentence.

Confirmation bias

People notice and accept information that fits prior beliefs more easily. Search queries can magnify this if they are written to prove a conclusion.

Compare:

evidence AI destroys student learning

with:

generative AI student learning outcomes independent performance systematic review

The second query is less conclusion-loaded.

Motivated reasoning

People can apply stricter evidence standards to claims they dislike than to claims they prefer.

A useful practice is symmetrical scrutiny: what evidence would I demand if the conclusion were reversed?

Lateral reading

When a source is unfamiliar, strong digital readers do not always stay on the page and inspect its design. They open new sources to learn who publishes it, how others describe it and whether claims are corroborated.

This technique is especially useful for polished websites that create authority visually.

Who is behind the source?

Search the organisation name independently. Look for ownership, expertise, funding, editorial policy and correction history where relevant.

What is the evidence?

Follow citations rather than trusting the source’s summary.

What do other strong sources say?

Independent comparison can reveal whether the claim is mainstream, contested or isolated.

Reverse image and media verification

Where tools are available, image search or provenance tools can help identify earlier appearances of an image. But the user still needs to check context and date.

Old image, new claim

A genuine image from one event can be reused to represent another. Authentic pixels do not guarantee authentic context.

Screenshot literacy

Screenshots are persuasive because they look like direct evidence. They can be:

• cropped;
• edited;
• taken from a parody account;
• outdated;
• generated;
• missing the surrounding thread.

Find the original page or account where possible.

AI-generated misinformation

Generative systems reduce the cost of producing plausible text and media at scale. The verification principles do not change, but the quantity and fluency of questionable content can increase.

Quantity is not corroboration

Ten generated pages repeating one false statement do not create ten independent sources.

Style is not authenticity

A document can imitate academic, legal, journalistic or official language without belonging to the institution it resembles.

Information laundering

A weak claim can become more credible-looking as it moves through layers:

anonymous post → blog → generated summary → presentation → student essay.

At every layer, the source can become less visible while the language becomes more polished.

Provenance notebook

For important digital research, record:

• claim;
• source;
• original source if different;
• date;
• evidence type;
• verification status;
• uncertainty;
• where the claim is used.

This is a digital version of the evidence dossier in Reference Literacy.

Search literacy and AI search

AI-enhanced search interfaces can synthesise several results into one answer. This can reduce navigation cost while making source boundaries less visible.

The reader should identify:

• which claims are linked to which sources;
• whether sources are independent;
• whether the summary preserves disagreement;
• whether source dates match the requested time period;
• whether important primary sources are missing.

Answer engines and source compression

When a system answers directly, the user may never visit the underlying page. This changes reading behaviour.

SETC’s solution is not to reject summaries. It is to teach escalation:

summary for orientation → source for verification → synthesis for judgement.

AI literacy and democracy

The Council of Europe’s 2026 recommendation places AI literacy inside human rights, democracy and the rule of law. For English learners, this means language judgement has civic consequences.

Citizens increasingly encounter:

• generated campaign material;
• synthetic media;
• automated public-service communication;
• algorithmic recommendations;
• AI summaries of policy;
• generated commentary.

They need enough English, evidence and rhetorical literacy to inspect those messages rather than merely consume them.

Social agency

AI literacy should not make learners feel powerless in front of complex systems. Social agency means knowing when to question, verify, opt out, seek help, appeal a decision or demand a human explanation.

Digital well-being

Digital English performance can be degraded by attention fragmentation, compulsive checking and endless feed exposure.

Well-being is not outside literacy. A reader who cannot sustain attention long enough to understand a complex argument has reduced functional access to language.

Attention design

Use environments that match the task. Long-form reading may need notifications off. Collaborative chat may need them on. The goal is intentional attention rather than permanent disconnection.

AI conversational pull

AI systems can make continued interaction frictionless. Learners should know when another prompt adds value and when it replaces independent thinking.

Digital balance

A healthy English learning environment can include:

• books and long-form reading;
• handwriting;
• face-to-face conversation;
• digital search;
• AI tutoring;
• collaborative documents;
• audio/video;
• independent offline practice.

The question is not screen versus no screen. It is which medium strengthens the target capability.

Worked Digital & AI English workflows

The strongest way to teach digital and AI literacy is through repeated decisions. The workflows below show how the same principles operate across reading, writing, research and communication.

Workflow 1: AI explains a difficult paragraph

A learner cannot follow a dense academic paragraph.

Weak workflow:

1. paste paragraph;
2. ask AI to simplify;
3. read simplification;
4. move on.

Stronger workflow:

1. attempt the paragraph first;
2. mark the exact sentence or relationship that failed;
3. ask AI to unpack only that relationship;
4. compare the explanation with the original;
5. reconstruct the paragraph in your own words;
6. check against the source;
7. later read a similar paragraph without AI.

The AI is used to restore the reading system, not bypass it.

Workflow 2: AI edits an essay paragraph

Maren writes a paragraph and asks for feedback.

She requests:

Identify the two highest-level problems. Do not rewrite. Separate evidence problems from language problems. Ask me to revise first.

The system flags an unsupported causal claim and a paragraph that drifts from the thesis. Maren revises. Then she asks AI to compare versions and explain what changed.

The final test is a new paragraph on a different topic.

Workflow 3: AI generates a thesis

The learner asks for three theses. Instead of choosing the most polished, they evaluate:

• Does it answer the exact question?
• What terms need definition?
• What evidence burden does it create?
• Is the claim too strong?
• Could I defend it without AI?

Then the learner writes a fourth thesis independently.

Workflow 4: AI provides a citation

Iona receives a plausible journal reference.

1. search the title;
2. verify author and journal;
3. open the original source;
4. locate the claimed finding;
5. check sample and date;
6. record the citation from the source, not from AI memory.

If the source does not exist, the reference is discarded and the claim is re-researched.

Workflow 5: AI summarises an article

The summary says a study “proves” that AI improves learning.

Iona checks the paper and finds that the study measured immediate task quality under AI access, not delayed unaided performance.

The repair is not only factual. It is conceptual: product performance was incorrectly transformed into learning.

Workflow 6: AI proposes vocabulary

A learner asks for “advanced words” for an essay. The output contains words that are formal but poorly collocated.

Use Reference Literacy:

1. identify the intended distinction;
2. check dictionary sense;
3. check collocation;
4. check academic register;
5. write an original sentence;
6. retain only words that increase precision.

Workflow 7: AI creates comprehension questions

Leonie checks whether the questions target literal recall, inference, evidence or evaluation. If all questions ask surface facts, the tool has not created the intended practice.

She asks for revised questions with explicit skill labels, then audits each label.

Workflow 8: AI helps with oral practice

A student asks AI to simulate an examiner. The learner responds orally, then transcribes only selected turns for review.

Useful feedback categories:

• answer development;
• evidence/example;
• interaction;
• clarity;
• pronunciation where the system can assess it reliably;
• recovery after challenge.

Human review may still be needed for pronunciation and nuanced interaction.

Workflow 9: AI translation

A bilingual learner translates a school explanation for a family member.

The learner checks:

• technical terms;
• institution names;
• dates/numbers;
• tone;
• whether culturally specific ideas were flattened.

Meaning preservation is the primary criterion.

Workflow 10: AI and creative writing

A learner is writing fiction.

Useful AI roles may include:

• generate constraints;
• ask character questions;
• identify continuity errors;
• simulate a reader response.

If AI creates every plot turn and sentence, the learner is practising curation more than authorship. That may be a valid creative workflow, but it is not the same learning target.

Workflow 11: search + AI + source synthesis

A student researches “Does social media harm attention?”

They begin with AI to identify terminology: attention, task switching, problematic use, longitudinal study, adolescents.

Then they search authoritative databases and reports, read selected sources, construct notes and return to AI only to challenge their synthesis.

AI helps map the field; sources own the evidence.

Workflow 12: current exam information

A student asks AI for the current format of an examination. The answer sounds confident.

The student opens the current official syllabus/specification and verifies paper structure, year and changes.

Current institutional facts should not remain AI-only.

Workflow 13: AI gives a grammar rule

AI says a phrase is “incorrect”. The learner checks a dictionary and corpus and discovers it is a regional or register variant.

The answer becomes:

This form is attested, but your school’s formal target prefers X.

The learner has converted false universality into context-sensitive English.

Workflow 14: AI-generated feedback conflicts with teacher feedback

Do not choose the more flattering answer. Compare criteria.

The teacher may be applying a specific rubric the AI has not seen. Give the AI the rubric if allowed and ask it to explain the difference. The official assessment criteria control the mark.

Workflow 15: AI suggests a source-free statistic

Delete it from the draft until verified.

Do not keep an attractive number in working memory and search only for confirmation later. That creates anchoring.

Workflow 16: AI rewrites local English

A model changes Singapore English or another regional form to American English without explanation.

Ask:

• Was the original actually wrong?
• Was the task formal Standard English?
• Was the change regional rather than grammatical?
• Did meaning or identity change?

World Englishes literacy protects against default-standard bias.

Workflow 17: AI helps a parent understand a school comment

AI can paraphrase a teacher comment, but the parent should preserve the exact original wording and ask the teacher when the interpretation affects action.

AI is a mediator, not the teacher’s mind.

Workflow 18: AI and study planning

AI creates a detailed revision schedule. The learner checks whether the schedule includes retrieval, spacing, practice under exam conditions and realistic time.

A beautiful schedule that is impossible to follow is not useful.

Workflow 19: AI and model answers

Instead of reading a generated model passively:

1. predict what a strong answer should contain;
2. compare with the model;
3. identify decisions;
4. critique weaknesses;
5. close the model;
6. write a new answer.

Workflow 20: AI and correction

When AI corrects a sentence, ask for:

• the error type;
• why it changes meaning or convention;
• one contrast example;
• a fresh test sentence.

Correction becomes learning evidence.

Digital & AI English failure states

Answer addiction

The learner asks before thinking. The repair is an attempt-first rule.

Prompt addiction

The learner believes better prompts will solve weak subject understanding. The repair is concept study and source reading.

Polish addiction

The learner repeatedly rewrites for style without improving reasoning. The repair is to freeze sentence editing until claim/evidence structure is checked.

Summary addiction

The learner consumes summaries instead of texts. The repair is summary for orientation followed by direct reading of load-bearing sections.

Citation illusion

The presence of references creates trust without verification. The repair is source-level checking.

Confidence illusion

Fluent AI tone is mistaken for certainty. The repair is explicit evidence labels.

Personalisation illusion

The learner assumes an AI response is personalised because it uses their name or prior context. Personalisation language does not guarantee pedagogical diagnosis.

Tutor illusion

The system behaves conversationally, so the learner assumes it understands their learning history as a human tutor would. The repair is to make evidence of performance explicit.

Memory outsourcing

The learner stops remembering because everything can be asked again. The repair is regular closed-tool retrieval.

Search outsourcing

The learner asks AI for sources without learning how to search independently. The repair is alternating AI-assisted and independent search.

Judgement outsourcing

The learner asks, “Which is better?” before defining criteria. The repair is criteria-first comparison.

Voice outsourcing

The learner accepts AI style as default. The repair is preserve-and-explain editing.

Responsibility outsourcing

The learner says, “The AI told me.” The repair is explicit final-owner responsibility.

Verification fatigue

Checking every trivial sentence becomes exhausting. The repair is risk-based verification: verify load-bearing and changing claims first.

Tool hopping

The learner repeatedly changes tools instead of learning the task. The repair is tool stability long enough to compare learning outcomes.

Automation creep

A tool introduced for one low-level task gradually begins doing planning, drafting and decision-making. Periodically audit which cognitive operations remain human.

AI perfectionism

The learner continues prompting because another revision is always possible. Use a stop rule tied to the task standard.

Output accumulation

Generating ten versions creates more decision load than value. Generate alternatives only when the comparison serves a question.

Feedback overload

AI can produce more feedback than the learner can use. Limit each cycle to one or two high-value repair targets.

Rubric hallucination

The system invents what an examiner “wants”. Use official rubric language.

Source homogenisation

AI synthesis can flatten disagreement. The repair is source-by-source claim tracking before synthesis.

False originality

Generated text may feel original because it is newly worded while relying on common patterns or untraceable source influence. For academic work, originality is not simply novel wording.

Diagnostic matrix

The first repairable cause in AI dependence

Do not respond to overuse with one generic ban. Trace the dependency:

What did the learner stop doing because the tool started doing it?

That is the first operation to rebuild.

Digital & AI English curriculum: from safe use to independent judgement

Digital and AI literacy should develop progressively. A Primary learner needs different responsibilities from a JC student or undergraduate, but the underlying trajectory is stable:

recognise → question → verify → use deliberately → create responsibly → transfer independently.

Phase 1: digital orientation

Teach the learner to distinguish:

• webpage from advertisement;
• source from search result;
• human-written from generated where disclosed;
• public from private information;
• link from citation;
• fact from opinion;
• message from evidence.

Exit evidence: the learner can explain what kind of digital object they are looking at and what it can reasonably prove.

Phase 2: search literacy

Teach information-need definition, query refinement, official-source discovery and source comparison.

Exit evidence: the learner can locate the correct source owner for a simple current fact without relying on an AI answer.

Phase 3: AI output scepticism without cynicism

Teach that AI can be useful and wrong. Learners should identify claims requiring verification and understand that fluent language is not source evidence.

Exit evidence: the learner can take one AI answer, mark which parts need checking and verify them through appropriate sources.

Phase 4: prompting as task representation

Teach purpose, audience, constraints, source boundaries and desired output.

Exit evidence: the learner can explain why one prompt is better specified than another and can complete the underlying task without the prompt template.

Phase 5: AI-assisted revision

Use AI to diagnose learner-created work while preserving learner decision-making.

Exit evidence: the learner can accept, reject or modify suggestions with reasons.

Phase 6: source verification

Teach citation checking, provenance, source hierarchy and claim-to-source matching.

Exit evidence: no central AI-generated factual claim enters final work without a verifiable source trail.

Phase 7: digital genre control

Move one communication task across email, chat, formal page, social post and presentation.

Exit evidence: the learner changes register, structure and explicitness while preserving meaning.

Phase 8: bias and representation

Compare outputs across names, regions, dialects or perspectives. Analyse what the system assumes.

Exit evidence: the learner can identify at least one representational bias without assuming every difference is discriminatory.

Phase 9: multimodal verification

Work with screenshots, charts, generated images, transcripts and videos.

Exit evidence: the learner can distinguish content, caption, provenance and interpretation.

Phase 10: AI-assisted research

Use AI for field mapping and question generation, then conduct source-level research independently.

Exit evidence: the evidence set remains traceable without the AI conversation.

Phase 11: AI-assisted production under explicit rules

Complete a task where AI is deliberately permitted. Assess prompt quality, verification, revision, authorship disclosure and final accountability.

Exit evidence: the learner can explain every major AI-assisted decision.

Phase 12: independent release test

Remove AI and change the surface task.

Exit evidence: the learner retains the underlying reading, writing, reasoning or research capability.

Curriculum principle: teach the non-AI skill first enough to judge AI

A learner cannot reliably evaluate an AI-generated paragraph if they have no model of strong paragraph reasoning. They cannot verify a paraphrase if they do not understand the source. They cannot identify hallucinated citations if they do not know what a real citation looks like.

AI literacy therefore depends on ordinary literacy.

Curriculum principle: AI capability should be added, not substituted

The learner should gain:

independent writing + AI-assisted writing, not AI-assisted writing instead of independent writing.

source reading + AI summarisation literacy, not summarisation instead of source reading.

human discussion + AI simulation, not simulation instead of interaction.

Curriculum principle: keep transfer visible

Every AI-assisted unit should identify a later unaided or differently aided task.

Digital & AI English standards by stage

Assessment governance

Institutions should decide AI conditions before students complete the task.

Possible states include:

• AI prohibited;
• AI allowed only for defined support;
• AI permitted with disclosure;
• AI required as part of the assessed workflow.

Ambiguous rules produce accidental misconduct and weak assessment evidence.

Assessment condition should appear in the task

Students should know:

• which tools are permitted;
• which stages are permitted;
• whether generated text can appear in final submission;
• how use should be disclosed;
• whether sources suggested by AI must be independently verified;
• what part must be completed independently.

AI and feedback policy

Schools should distinguish learner use from teacher use. A teacher using AI to draft comments still owns the feedback. Personalised assessment data may raise privacy issues. Generic comments generated without reading the learner’s work are not diagnostic feedback.

AI and curriculum materials

Generated worksheets, passages and questions require review for:

• accuracy;
• age appropriateness;
• answerability;
• bias;
• copyright/provenance;
• alignment with target skill;
• quality of worked answers.

AI and examination preparation

AI can generate endless mock items, but quantity without construct fidelity can train the wrong skill.

Examination materials should match:

• official syllabus;
• question type;
• difficulty;
• timing;
• scoring criteria;
• language quality.

Generated questions must be audited before use.

AI and marking

Automated marking can support preliminary feedback but should not be assumed equivalent to human scoring under every rubric.

Where consequential marks are involved, institutions need evidence of reliability, fairness and appropriate review.

AI and oral assessment

Speech systems can transcribe and analyse language, but accent and audio quality can affect performance. Human interpretation remains important where nuance matters.

AI and special educational needs

AI can support accessibility through text simplification, speech, translation and personalised practice. It can also create dependence or inaccurate accommodations.

Support should be designed around the learner’s real access needs and assessed capability.

Governance: a fast-moving owner needs review triggers

SETC should not rewrite this page every time a new AI product launches. It should review when:

• major education frameworks change;
• important institutional policies change;
• new evidence materially changes learning conclusions;
• platform capabilities alter verification or privacy risks;
• new forms of synthetic media become common;
• school/examination AI policies change.

Stable content

Keep stable:

• source verification;
• authorship;
• independent transfer;
• privacy;
• audience;
• claim/evidence relationships;
• human accountability.

Volatile content

Date and review:

• specific tool capabilities;
• detector claims;
• institutional AI rules;
• platform privacy settings;
• current model limitations;
• new legislation or regulation.

Federation boundary

SETC owns Digital & AI English where the controlling question is language, reading, writing, evidence, authorship and learning.

Broader professional automation, career workflows and adult workplace application should hand off to eduKateSG’s adult/career estate. SETC explains the English machinery; the federation applies it to wider life and work.

AI and the eduKate knowledge graph

The federation should use AI to strengthen canonical ownership, not generate duplicate content at scale.

Before publishing AI-assisted material:

1. locate the Atlas territory;
2. identify the canonical owner;
3. decide whether the new item is owner, application or directory;
4. check collision;
5. verify factual claims;
6. preserve internal links that express real relationships;
7. record volatility and review triggers.

Automation without ontology creates content debt faster.

AI and SEO

AI makes it cheap to generate keyword pages. That does not make them useful.

SETC’s rule remains: search demand can reveal an information need, but it does not define the educational ontology. A search phrase should route into the correct owner rather than justify another near-duplicate root.

AI and style consistency

AI can help maintain terminology across large sites, but it can also flatten distinctive voice. Use canonical definitions and owner maps, then allow genre-appropriate style variation.

AI and internal linking

Generated links should not be based only on keyword similarity. Link relationships should be typed conceptually:

• parent;
• prerequisite;
• mechanism;
• application;
• evidence;
• practice;
• federated handoff.

AI and content review

A machine can identify possible outdated claims or broken links, but promotion, deletion and canonical changes should follow the ownership rules.

The institutional AI question

Before automating a task, ask:

What capability, responsibility or evidence becomes less visible if the machine does this?

If the answer is important, redesign the workflow rather than automate blindly.

A complete human–AI English workflow

For substantial English work, SETC uses the following sequence.

1. Define the human job

What is the learner or writer meant to understand or produce independently?

If that is unclear, AI use cannot be judged properly.

2. Define the permitted AI role

Is AI being used for brainstorming, explanation, practice, feedback, source discovery, editing, translation, simulation or final production?

3. Attempt first where learning requires it

Create an initial model, paragraph, answer, search query or explanation before asking AI to help.

4. Use the smallest effective assistance

Prefer a hint before a full solution when a hint will do.

5. Inspect the output

What did AI change or add? Which claims are new? Which assumptions appeared?

6. Verify load-bearing content

Check citations, statistics, quotations, current rules, technical claims and language-usage claims against the correct owners.

7. Make the human decision

Accept, reject or modify with reasons.

8. Preserve authorship and source boundaries

Disclose use where required. Keep source identity visible.

9. Produce the final work

The human remains responsible for the final representation.

10. Reflect

What did the AI do that you could not yet do? What did you reject? What did you learn?

11. Retest without AI

Use a fresh task after a delay.

12. Decide whether support can reduce

If independent performance is stable, AI becomes optional rather than necessary.

Frequently asked questions

Is using AI cheating?

Not inherently. It depends on the task rules and what AI is doing. Using AI in a task that prohibits it can violate the rules. Using AI to replace the capability being assessed can invalidate the assessment even if the technology itself is not prohibited elsewhere. Always follow the governing institution’s current policy.

Can AI help me learn English?

Yes, when used for explanations, varied practice, feedback, questioning, simulation and retrieval in ways that preserve learner thinking. The strongest evidence comes from later independent performance.

Can AI make my writing better?

It can often improve the immediate document. Whether it makes you a better writer depends on whether you understand, practise and later reproduce the decisions independently.

Should I ask AI to correct all my grammar?

For a final real-world document, comprehensive correction may be useful if permitted. For learning, full correction can hide recurring error patterns. Ask for diagnosis and explanation, then repair selected errors yourself.

Can I trust AI explanations?

Treat them as provisional. Simple explanations may be accurate; complex or current claims require stronger verification. If the exact rule matters, use Reference Literacy.

Can AI invent sources?

Yes. Always verify generated citations against real source records and check that the source supports the claim.

Can AI search the web?

Some systems can retrieve current sources; others cannot, and capabilities can change. Even with retrieval, the user should inspect source relevance, date and authority.

Can I cite AI?

Citation and disclosure rules vary by institution and publication. Follow the current rules. In most research contexts, AI should not replace citation of the underlying evidence when that evidence is available.

Can AI paraphrase a source for me?

It can, but you remain responsible for preserving source meaning and attribution. Compare the paraphrase against the original.

Can AI summarise a book or article I have not read?

It can generate a summary, but you should not represent yourself as having read the source based only on the summary. For important claims, read the relevant original material.

Can AI help me with vocabulary?

Yes. Use it to generate contrasts and examples, then verify important senses, collocations and register in dictionaries/corpora.

Can AI help with pronunciation?

Speech systems can provide models and feedback, but performance varies by accent and technology. Use multiple sources and human feedback where intelligibility matters.

Can AI help with speaking practice?

Yes. Simulated interviews, discussion questions and role-plays can increase practice. Include real human interaction because human turn-taking and social meaning are richer.

Can AI detect plagiarism?

Some systems can compare text or flag similarity, but plagiarism judgement depends on source use and authorship, not only phrase overlap. Follow institutional procedures.

Can AI detect AI writing?

Detection systems are imperfect and can produce false positives/negatives. Detector output should not be treated as definitive evidence of authorship by itself.

Should schools ban AI?

Different tasks need different conditions. Some learning operations benefit from AI-free practice; others benefit from explicit AI-literacy training. The core question is what capability the task is meant to build or assess.

Should young children use generative AI?

Use should be age-appropriate, supervised and aligned with platform/institution rules. Younger learners need stronger privacy protection and should still spend substantial time reading, writing, speaking and thinking without generative assistance.

What is the best AI prompt?

There is no universal prompt. The best prompt represents the task accurately while preserving the work you still need to do.

Should I tell AI to “act as an expert”?

That can shape style but does not confer expertise. Verify the content.

How do I stop AI from changing my meaning?

Specify what must be preserved: claim, certainty, technical terms, source attribution, tone or structure. Then compare versions.

How do I know if an AI answer is biased?

Compare outputs, inspect assumptions, check representation, test alternative prompts and use external evidence. Bias is not proved merely because you dislike the answer.

How do I protect privacy?

Do not upload sensitive or confidential information without appropriate permission. Minimise data and follow the service/institution’s current privacy guidance.

What if AI and my teacher disagree?

For assessment rules, the official rubric/teacher/institution governs. For language claims, verify with appropriate references. Ask which evidence each judgement uses.

Will AI make English learning unnecessary?

No. AI can produce language, but people still need to understand, judge, verify, communicate, interpret other humans, make decisions and know when generated language is wrong. In many contexts, AI increases the value of deep literacy because fluent output is easier to produce and harder to trust automatically.

Do I still need vocabulary if AI can suggest words?

Yes. Without vocabulary knowledge you cannot judge whether the suggestion fits meaning, register or collocation. Retrieval speed also matters in reading, speaking and unaided writing.

Do I still need grammar?

Yes. Grammar lets you interpret and construct relationships. A user without grammar awareness can accept an AI revision that subtly changes meaning.

Do I still need to read long texts?

Yes. Long-form reading builds sustained models, context, inference and knowledge that summaries cannot fully replace.

Will prompting become a core English skill?

Prompting is already a useful functional genre, but the deeper transferable skills are task definition, audience, constraints, evidence, clarity and evaluation. Those survive interface changes.

Current authoritative AI-literacy anchors

SETC should review this fast-moving owner against current authoritative sources. As of September 2026, useful anchors include:

• OECD Digital Education Outlook 2026: Exploring Effective Uses of Generative AI in Education — current synthesis of learning opportunities and risks, including the difference between task performance and learning.
• OECD / European Commission: Empowering Learners for the Age of AI — 2026 AI-literacy framework for primary and secondary education.
• UNESCO AI Competency Framework for Students — human-centred mindset, ethics, techniques/applications and design.
• Council of Europe Recommendation CM/Rec(2026)12 on AI literacy — human dignity, agency, privacy, equity, critical thinking and democratic participation.
• NIST AI Risk Management Framework — risk-management principles and generative-AI risk resources for organisations.

These are evidence anchors, not a complete Digital & AI English curriculum. SETC’s role is to connect their AI-literacy concerns to actual English reading, writing, research, rhetoric, verification and transfer.

A note on current evidence and causation

Current education studies and large-scale data are still developing. Associations between AI use and educational outcomes do not automatically establish causation. Students differ in prior achievement, task selection, access, motivation and use patterns.

SETC should therefore avoid headlines such as “AI makes students worse” or “AI improves learning” without specifying the design, population, task and outcome.

The more useful question is:

Which AI use, for which learner, performing which cognitive operation, under which support conditions, measured by which independent outcome?

The AI evidence ladder

The SETC AI-learning experiment

For any new AI learning workflow, SETC can run a small evidence cycle:

1. define the target skill;
2. measure baseline independently;
3. use AI under a defined role;
4. record what cognitive work remains human;
5. retest immediately;
6. retest after delay;
7. change surface task;
8. remove or reduce AI;
9. compare performance;
10. keep, redesign or reject the workflow.

This is more educationally useful than adopting a tool because it is impressive.

AI and the future of English assessment

As generated language becomes ordinary, assessment may increasingly separate:

• independent capability;
• AI-assisted capability;
• verification capability;
• collaborative human–AI production;
• oral defence and explanation.

Those are different constructs and can all matter.

AI and future English expertise

Some language tasks will become cheaper. Drafting routine text may require less time. That does not eliminate expertise; it moves value toward:

• question formation;
• judgement;
• source selection;
• verification;
• audience adaptation;
• ethical decisions;
• domain knowledge;
• original insight;
• interaction;
• responsibility.

Continue through SETC

• English Atlas — locate Digital & AI English inside the whole English system.
• How English Works — return to the language system beneath digital interfaces.
• Corpus, Dictionary & Reference Literacy — verify usage and choose the right authority.
• Evidence, Data & Multimodal English — verify claims, charts, images, sources and media.
• Rhetoric & Persuasion — analyse framing, attention, platform influence and manipulation.
• Academic English — apply AI and digital literacy to university research and writing.
• Multilingual English & Mediation — translation and meaning across languages and media.
• World Englishes & Variation — detect default-standard and regional-language bias.
• How Feedback Changes Performance — turn feedback into future performance.
• How English Learning Transfers — prove that supported learning survives new tasks.

The Digital & AI English standard

A digitally capable English user can define an information need, search intelligently, identify source ownership, read across links and media, communicate through digital genres, use AI deliberately, preserve privacy, verify claims, distinguish generated fluency from evidence, keep authorship honest, detect manipulation and adapt language to audience.

An AI-literate learner can do something more important: they can use assistance without losing the underlying human capability.

The final question is not:

Can AI help me do this?

It is:

After AI helps me do this, what can I now understand, decide and perform that I could not reliably do before?

The final release test

Give the learner a fresh task after a delay. Remove the AI or restrict it to a lower assistance level. Change the topic, source set or audience.

Then test whether the learner can:

• define the task;
• search or retrieve appropriately;
• read the sources;
• identify evidence;
• form a judgement;
• write or speak clearly;
• verify factual claims;
• protect source ownership;
• explain important decisions;
• recover from uncertainty.

If the capability survives, AI has become part of a stronger learning system.

If the capability disappears when the interface closes, the tool produced a better output—but the learner still needs teaching.

Use the machine. Keep the judgement. Verify the evidence. Preserve the human capability.

AI system literacy: understand enough of the machinery to ask better questions

English learners do not need to become machine-learning engineers in order to use generative AI responsibly. But they need a working model that is accurate enough to prevent false assumptions.

The most important principle is simple:

A language model generates language from patterns. Fluency does not mean the system has human understanding, direct experience or guaranteed access to the source of every claim.

Different systems may add retrieval, search, tools, files, memory, calculators, databases or other capabilities around the language model. Those additions can make answers more useful and current, but they do not remove the need to inspect what evidence was actually used.

Prediction and generation

Modern language models generate sequences by estimating plausible continuations from learned patterns and current context. The system can produce highly coherent prose because it has learned extensive regularities in language.

This explains both its strength and one of its risks. The same mechanism that allows a model to generate a plausible explanation can also generate a plausible falsehood if the context does not constrain the answer sufficiently.

Tokens

Language models process text in units often called tokens. A token is not always one word; it can be a word, part of a word, punctuation or another text unit depending on the system.

Learners do not need to count tokens manually for ordinary use. The useful insight is that the system processes representations of text, not human meanings in the same way a person experiences them.

Context

The system responds partly to the information available in its current context. That may include the user’s prompt, previous conversation turns, attached materials, retrieved sources or system instructions.

If important information is missing, the model may infer. Inference can be useful and wrong.

Context is not perfect memory

Users should not assume that a system remembers every previous conversation or every detail they have ever provided. Memory features vary by product and setting.

When a detail is essential, restate or verify it rather than relying on assumed memory.

Retrieval

Some AI systems can retrieve information from external sources such as web search, databases or connected files. Retrieval changes the evidence environment.

A retrieved answer can still fail if:

• the wrong source was retrieved;
• the source is outdated;
• the relevant passage was missed;
• the model misinterprets the source;
• several sources are merged inaccurately;
• the answer is stronger than the source.

Retrieval helps; it does not eliminate verification.

Tools

An AI system may use tools for calculation, code execution, search, databases, calendars or other actions. Tool use can improve accuracy for tasks that language prediction alone handles poorly.

The user should distinguish:

generated answer from tool-derived result.

For example, a calculation performed by a calculator tool is different from a number generated from language patterns.

Tool output can still be misused

A correct calculation based on the wrong inputs produces a wrong conclusion. A current search result can still be irrelevant. Tool accuracy does not replace task interpretation.

Training data

Language models learn statistical patterns from large datasets under specific training processes. Users usually do not know every source or exact influence behind a generated phrase.

This is one reason AI output should not be treated as a traceable citation to its training data.

Knowledge is not a library lookup

A model can generate factual material without retrieving a specific stored document. It may also blend patterns from many sources. This makes “Where did you get that?” an unreliable way to reconstruct provenance unless the system explicitly provides verifiable sources.

Fine-tuning and system behaviour

AI behaviour can be shaped after general training through additional data, feedback, system instructions, safety policies and product design.

This means two systems using similar underlying technology can behave differently.

Temperature and randomness

Some systems allow controls that affect output variability. Higher variability can increase creative diversity and inconsistency. Lower variability can make outputs more stable without guaranteeing truth.

Consistency is not the same as correctness.

Reasoning-like output

AI can produce stepwise explanations and solve complex problems. Users should evaluate the final reasoning and evidence rather than assuming that a detailed explanation is inherently reliable.

A long chain of plausible steps can contain one hidden false assumption that invalidates the conclusion.

Self-critique

Asking a model to critique its own answer can reveal problems because the new prompt creates a different generation context. But self-critique is not independent evidence.

Ensembles and multiple models

Comparing answers from several systems can reveal disagreement. Agreement can increase confidence modestly in some contexts, but overlapping training and common misconceptions mean agreement is not proof.

Memory systems

Some AI products can store user preferences or past information. This can improve continuity and create privacy considerations.

Users should know whether persistent memory exists, what information is appropriate to store and how to correct stale assumptions.

Connected applications

AI systems can sometimes connect to documents, email, calendars, databases or other services. Once connected, the language model may operate on information that was previously private to another system.

Access should follow least-necessary principles: connect only what the task requires and understand permission boundaries.

Agents and multi-step automation

AI systems can perform sequences of actions rather than one response. This creates a different risk profile because a mistaken interpretation can propagate through several steps.

Before allowing automation, define:

• what the agent may read;
• what it may change;
• what requires human approval;
• what should be logged;
• how to recover from error.

Human-in-the-loop

Human review is most valuable at decisions with high consequence or uncertainty. Review everything manually and automation loses value; review nothing and errors can scale.

Place human checkpoints strategically.

Model updates

AI systems change. A prompt that worked well before may behave differently after an update. Product-specific advice should therefore be dated and reviewed.

Benchmark performance

Models are often evaluated on benchmarks. Benchmark scores provide evidence about defined tasks, not proof of universal competence.

A model can perform strongly on one benchmark and fail on a novel classroom or community context.

Emergent behaviour and capability surprises

Users may discover capabilities or failure modes not obvious from a simple product description. This is another reason workflow-level testing matters.

Uncertainty estimation

Language models do not always communicate uncertainty in a calibrated way. “I am confident” should not be treated as a statistical confidence interval.

Users should infer confidence from evidence quality, not prose tone.

Boundary of model literacy

Students do not need to memorise every technical term. They need enough understanding to avoid these false assumptions:

• AI sounds confident, therefore it knows;
• AI provides a citation, therefore the citation exists;
• AI remembers everything, therefore it has complete context;
• AI searched, therefore the best source was found;
• AI used a tool, therefore the task interpretation was correct;
• AI agrees with me, therefore my claim is supported;
• AI wrote it better, therefore I learned.

AI English diagnosis: what kind of system error occurred?

The AI-source boundary

When an AI answer includes linked sources, distinguish three levels:

1. source exists;
2. source is relevant;
3. source supports the exact claim.

All three must hold.

The AI-tool boundary

When AI calls a tool, distinguish:

1. tool was appropriate;
2. inputs were correct;
3. tool output was correct;
4. AI interpreted output correctly;
5. final wording preserved the result.

This five-step chain is useful whenever automation becomes complex.

The model-literate learner

A model-literate learner does not need to distrust AI. They know which questions require stronger evidence, when the system is guessing, when context is incomplete and when the answer should be escalated to another source.

That is enough technical understanding to support good English judgement.

Digital & AI English investigation lab

The final stage is not memorising rules about AI. It is learning how to investigate a digital-language question with evidence.

Investigation 1: Does AI improve this student’s writing?

Do not compare one assisted essay with one unaided essay from months ago.

Design a cleaner test:

1. choose two comparable writing tasks;
2. collect baseline unaided writing;
3. use AI under a defined feedback role;
4. record which revisions the learner makes;
5. delay;
6. give a new unaided task;
7. compare the targeted mechanisms.

If reference clarity, paragraph reasoning or grammar control improves on the new unaided task, there is stronger evidence that learning transferred.

Investigation 2: Which AI prompt produces better learning?

Compare:

Prompt A: “Correct my paragraph.”

Prompt B: “Identify one evidence problem and one language problem. Ask me to repair them before you show alternatives.”

Measure not only final paragraph quality but later independent repair.

Investigation 3: Does AI summary reduce comprehension?

Use two matched texts. On one, allow a summary before reading. On the other, require orientation and direct reading first. Test:

• central idea;
• evidence;
• qualification;
• inference;
• delayed recall.

The result can show whether this learner uses summaries as scaffolds or substitutes.

Investigation 4: Which source errors does AI make?

Ask AI about a topic with known authoritative sources. Classify errors:

• invented source;
• real source, wrong title;
• real source, wrong claim;
• outdated source;
• missing qualification;
• accurate citation.

This turns “AI hallucinates” into observable source-literacy work.

Investigation 5: Does AI erase regional English?

Provide sentences from several legitimate English varieties. Ask for correction without specifying a target. Record which forms are changed.

Then repeat with a defined target:

Convert this to formal Standard English for an international academic audience. Distinguish regional variation from grammar error.

Compare the outputs.

Investigation 6: Can learners detect AI overclaiming?

Give students a generated paragraph containing:

• one accurate cautious claim;
• one true statistic with missing denominator;
• one correlation written as causation;
• one invented citation;
• one outdated fact.

The task is to classify and verify each problem.

Investigation 7: Search engine vs AI search

Give one current question. Half the group begins with ordinary search; half begins with AI-assisted search. Everyone must produce the same source dossier.

Compare:

• time to first strong source;
• source diversity;
• primary-source retrieval;
• misattribution;
• final accuracy.

The result evaluates workflow, not ideology.

Investigation 8: Does AI improve vocabulary transfer?

Teach ten advanced words with AI-generated examples, collocations and contrast questions. Retest later without AI in new reading/writing contexts.

Compare with a learner who only read definitions.

The target is usable vocabulary, not prompt engagement.

Investigation 9: Does AI feedback improve oral performance?

Record a short oral response. Use AI feedback on structure and clarity. Rehearse. Then answer a different unseen question.

If structure improves only on the rehearsed response, the learning has not transferred fully.

Investigation 10: Can learners resist polished misinformation?

Create two answers:

• a poorly written answer with strong sources;
• a polished answer with weak or false evidence.

Ask learners which deserves trust and why.

This reveals whether rhetoric is overpowering evidence literacy.

AI-English evidence notebook

For each significant AI interaction, learners can record:

The notebook turns invisible AI use into inspectable learning evidence.

Publication audit for AI-assisted English content

Before SETC or any eduKate site publishes substantial AI-assisted material, run this audit.

Ownership

• Which Atlas territory owns the concept?
• Does a canonical owner already exist?
• Is this page an owner, application or directory?
• Does it duplicate another federated site?

Evidence

• Which claims are factual?
• Which are current?
• Which are contested?
• Were primary sources checked?
• Were citations verified?
• Does paraphrase preserve qualification?

Language

• Are definitions stable?
• Is regional variation represented accurately?
• Has AI flattened register or voice?
• Are collocations natural?
• Are claims stronger than evidence?

Learning value

• What will the reader be able to understand or do after reading?
• Are mechanisms explained?
• Are there worked examples?
• Are diagnosis and transfer included?
• Does the article add information rather than volume?

Digital quality

• Are links purposeful?
• Do external links go to the correct authority?
• Are screenshots or interface instructions likely to age quickly?
• Does fast-moving material have a review trigger?

AI transparency

Where institutional or editorial policy requires it, record how AI contributed. Internally, preserve enough process evidence to audit errors later.

AI-assisted publication failure states

Duplicate owner creation

AI sees search demand and generates a new broad page despite an existing owner. Repair: route the intent into the owner or create an application.

Keyword inflation

The text repeats search phrases without adding concepts. Repair: delete repetition and strengthen mechanisms.

Source laundering

AI paraphrases secondary claims until the original source disappears. Repair: restore the primary source trail.

False comprehensiveness

Length creates the appearance of completeness while major mechanisms remain missing. Repair: audit against Atlas territory coverage.

Generic expertise

Polished explanations lack worked decisions. Repair: add diagnosis, examples and transfer tests.

Stale certainty

Product-specific claims are published without dates. Repair: date volatile material and set a review trigger.

Federation drift

One site starts owning another site’s specialist domain because AI can generate the content cheaply. Repair: restore the federation boundary.

The digital reading stress test

Give the learner a realistic online research task with:

• search results;
• one official source;
• one strong secondary source;
• one outdated page;
• one persuasive low-quality page;
• one AI summary;
• one relevant chart;
• one misleading screenshot.

The learner must:

1. identify the strongest owner;
2. verify current facts;
3. interpret the chart;
4. reject or qualify weak claims;
5. write a short answer with traceable sources.

This integrates digital English.

The digital writing stress test

Give a communication problem that must move through several genres:

1. research note;
2. AI prompt;
3. draft email;
4. short public update;
5. formal record.

The learner must preserve the same core facts while changing register, detail and structure.

The AI authorship stress test

Give a learner their own AI-assisted final piece and ask them, without looking at the conversation:

• state the thesis;
• explain each major source;
• defend the structure;
• identify one rejected AI suggestion;
• rewrite one section independently;
• answer an unseen objection.

If the learner cannot, authorship is weaker than the polished document suggests.

The AI verification stress test

Give five generated claims with mixed truth status. The learner must choose the appropriate evidence tool for each:

• current official website;
• dictionary;
• corpus;
• historical source;
• research paper;
• calculation;
• human expert.

The skill is source selection, not just fact checking.

The independent-transfer stress test

After several weeks of AI-supported learning, give a novel task with no AI.

Measure:

• initiation time;
• planning quality;
• source selection;
• accuracy;
• language control;
• self-correction;
• completion;
• confidence calibration.

Compare with baseline. This is the strongest practical answer to “Did AI help the learner?”

The institutional release condition

Digital & AI English is ready as a canonical owner when it can do four things simultaneously:

1. remain useful as tools change;
2. ground fast-moving claims in current authoritative sources;
3. teach practical digital and AI workflows;
4. protect independent English capability as the final educational measure.

Tool literacy without English literacy is fragile. English literacy without digital literacy is increasingly incomplete.

SETC therefore treats the two as one connected modern system.

AI-output evaluation rubric: judge the language before you use it

Generated English should not be accepted or rejected as a block. Evaluate it across separate dimensions. One answer can be grammatically excellent and evidentially weak; another can contain strong evidence but poor audience fit. A rubric prevents one attractive feature from hiding another failure.

Do not average the rubric blindly

Some dimensions are gating conditions. A beautifully written answer with fabricated evidence cannot become acceptable because it scores highly on tone and grammar. A private-data leak cannot be cancelled by excellent structure.

Use severity, not just totals.

The red–amber–green decision architecture

A simple classroom model can help learners decide how much verification an AI output needs.

Green: low-consequence generative support

Examples:

• brainstorming fictional character names;
• creating extra grammar-practice sentences;
• generating discussion questions;
• offering several ways to phrase a low-stakes personal note.

Check quality, but extensive source verification may not be necessary.

Amber: learning and factual support

Examples:

• explanations;
• summaries;
• grammar advice;
• vocabulary usage;
• historical claims;
• research brainstorming.

Verify important facts and compare with appropriate reference sources.

Red: high-consequence or institutional claims

Examples:

• current examination rules;
• medical/legal/financial information;
• formal academic citations;
• public institutional statements;
• safety instructions;
• sensitive personal data;
• high-stakes assessment submissions.

Use primary authority, strong verification and human accountability.

Digital decision architecture: when not to use AI

AI should be rejected or postponed when:

• the target skill is exactly the operation AI would perform;
• privacy cannot be protected;
• the task rules prohibit it;
• the user cannot verify the answer;
• the source material cannot legally or ethically be uploaded;
• the decision requires professional responsibility outside the user’s competence;
• independent retrieval practice is more valuable than convenience;
• human conversation is the learning objective.

When AI is the better tool

AI can be especially useful when:

• many varied practice examples are needed;
• the learner needs immediate low-stakes questioning;
• several wording alternatives should be compared;
• a long explanation needs a first orientation;
• the learner needs simulated dialogue;
• repetitive formatting or transformation creates unnecessary friction;
• the human can verify the result cheaply.

When search is better than AI

Use direct search when the primary goal is to locate:

• a current official page;
• an exact source;
• a live timetable;
• a policy;
• a published paper;
• a dictionary entry;
• a specific quotation.

AI can assist discovery, but direct source access reduces one transformation layer.

When a dictionary/corpus is better than AI

For exact language evidence—sense, collocation, regional label, historical use, register distribution—specialist reference tools give a clearer evidence trail.

When a human is better than AI

Human support is especially important when the task requires:

• relationship and trust;
• high-stakes safeguarding;
• deep learner diagnosis over time;
• ethical accountability;
• nuanced oral interaction;
• knowledge of local context not present in the system;
• judgement under genuinely novel conditions.

The human–AI handoff

A strong workflow makes handoffs explicit.

Human defines → AI assists → human verifies → AI may revise → human decides → independent human retest.

The handoff prevents the interface from silently becoming the owner of the task.

AI-English maturity levels

Level 1: consumer

Accepts answers at face value.

Level 2: operator

Can prompt for desired output.

Level 3: critic

Can identify errors, bias and missing evidence.

Level 4: verifier

Can choose authoritative sources and confirm claims.

Level 5: collaborator

Can use AI while preserving human decision-making and authorship.

Level 6: designer

Can create workflows where AI supports learning or production without destroying the target capability.

Level 7: independent transfer

Can perform the core intellectual work when AI is unavailable, inappropriate or wrong.

SETC’s goal is not to keep learners at Level 1 with a more sophisticated interface. It is to build toward Level 7.

The tutor’s AI observation sheet

During a three-student tutorial, the tutor can record:

• what the learner attempts before AI;
• what they ask AI to do;
• whether they read the answer critically;
• which suggestions they accept;
• whether they verify new claims;
• whether they can explain the final version;
• whether they transfer after support is removed.

This makes AI use visible without turning the lesson into surveillance.

The parent’s AI observation sheet

A parent does not need to inspect every prompt. Look for capability signals:

• Can the child start work without AI?
• Can they explain what AI contributed?
• Can they identify one mistake?
• Can they do a similar short task independently?
• Are they reading original sources?
• Are they protecting personal information?

The learner’s AI self-check

Before submitting work, ask:

1. What part is mine?
2. What part came from AI?
3. What facts did AI introduce?
4. Have I verified them?
5. Does the final wording still mean what I intend?
6. Have I followed the rules?
7. Could I explain this without the screen?

The future-proof principle

Interfaces will change. Models will improve. Tool names will disappear. Some current weaknesses will shrink and new failure modes will appear.

The durable English capacities remain:

• understand;
• question;
• search;
• verify;
• reason;
• write;
• speak;
• adapt;
• take responsibility.

Digital & AI English should therefore be revised around evidence, not fashion.

The everyday Digital & AI English operating protocol

Most learners do not need to remember every section of this owner while working. They need a compact routine that preserves the deeper system.

Before using AI

1. Name the task. Reading, writing, search, feedback, translation, research, practice or final production?
2. Name the learning goal. What capability are you trying to build?
3. Name the boundary. What should AI not do because that operation belongs to you?
4. Check the rules. Is AI permitted here?
5. Protect data. Remove unnecessary personal or confidential information.

While using AI

6. Ask specifically. Give purpose, audience and constraints.
7. Keep source boundaries visible. Separate supplied evidence from generated additions.
8. Notice new claims. Every new factual claim creates a verification burden.
9. Read critically. Do not accept fluency as proof.
10. Make decisions yourself. Accept, reject or revise with reasons.

Before using the output

11. Verify the load-bearing facts.
12. Check current information. Policies, products, exams and rules can change.
13. Check language evidence. Use dictionaries/corpora where usage matters.
14. Check meaning. Did the rewrite alter certainty, scope or attribution?
15. Check audience. Is the register right?
16. Check authorship. Does the use comply with the task?

After the task

17. Explain what AI contributed.
18. Name what you learned.
19. Retest later without AI.
20. Reduce support if performance survives.

Twenty-second verification triage

When time is limited, ask four questions:

• Could this fact have changed?
• Would it matter if this claim were wrong?
• Is there an obvious primary owner?
• Did AI introduce the claim rather than my source?

If the answer to any is yes, verify before relying on it.

The one-minute learning check

Close the AI output and answer:

• What was the main idea?
• What decision did I make?
• What evidence supports it?
• What can I now reproduce?

If the learner cannot answer, the interaction may have produced exposure rather than learning.

The five-minute source check

For an important generated claim:

1. find the source owner;
2. open the source;
3. locate the relevant passage/data;
4. compare the claim strength;
5. record the real citation.

The ten-minute transfer check

Change the topic and repeat the target skill without AI for ten minutes. This quickly reveals whether the learner owns the method or only remembers the output.

The pause rule

Pause AI use when:

• you are prompting repeatedly without knowing what you want;
• you cannot tell whether the answer is true;
• the tool is writing more than you are thinking;
• you are sharing information you should not share;
• you are avoiding the exact practice you need;
• the institution’s rules are unclear.

Return to the human task model first.

The escalation rule

Move from AI to another source when the question requires a stronger authority:

• dictionary/corpus for language usage;
• official specification for examination rules;
• primary research for evidence;
• institution for policy;
• qualified professional for high-stakes professional advice;
• human teacher/tutor for persistent learner diagnosis.

The independence rule

For any skill SETC claims to teach, AI-supported success is provisional until independent evidence exists.

This protects the educational mission from a seductive mistake: confusing a stronger human–machine system with a stronger learner.

The final Digital & AI English principle

The arrival of generative AI does not reduce the importance of English. It increases the importance of the parts of English that machines cannot responsibly own for us: understanding what a question means, judging evidence, choosing what to say, knowing what another person needs, recognising uncertainty, preserving source truth, taking responsibility and learning enough to act when the tool is absent or wrong.

Digital fluency is valuable. AI fluency is valuable. Human judgement remains the owner.

Institutional quick audit before adopting an AI-English workflow

Before a school, tutor or education website adopts a new AI workflow, answer these questions in writing.

1. Learning target: Which human capability is the workflow supposed to strengthen?
2. AI role: What exact operations will the system perform?
3. Human role: Which decisions remain with the learner or teacher?
4. Evidence: How will we know whether learning improved rather than only the immediate output?
5. Transfer: When will the learner perform a fresh task with reduced or no AI support?
6. Verification: Which generated claims require source-level checking?
7. Privacy: What data enters the system, and is all of it necessary?
8. Equity: Does the workflow depend on paid access, device quality or support that some learners do not have?
9. Bias: Which language varieties, identities or assumptions might be treated as default?
10. Assessment validity: Does AI replace the capability the assessment claims to measure?
11. Authorship: What use is permitted, recorded or disclosed?
12. Recovery: What happens if the AI output is wrong or the system is unavailable?
13. Review trigger: Which policy, capability or evidence change would require the workflow to be reviewed?

If these questions do not have clear answers, the workflow is not yet mature enough to become part of the curriculum merely because the technology is impressive.

A premier education system should adopt tools by educational evidence, not novelty. The tool can change next year; the learner’s ability to read, reason, verify, communicate and act independently must remain the durable asset.

SETC review protocol for a fast-moving English territory

Digital & AI English should be reviewed by evidence class rather than by calendar alone. Foundational principles such as authorship, source verification, privacy, audience and independent transfer remain unless better evidence changes them. Tool-specific statements should be rechecked when interfaces, model capabilities or provider policies materially change. Institutional statements should be rechecked when schools, examination boards, universities or regulators update their rules.

Every review should ask three questions: What changed in the technology? What changed in the evidence about learning? What changed in the rules governing use?

If none changed materially, the page should not be rewritten merely to appear fresh. If one changed, update the smallest affected layer and preserve the stable owner architecture. This prevents a fast-moving topic from forcing the entire English knowledge graph into permanent churn.

The page therefore carries two clocks: a fast clock for tools and policy, and a slow clock for language, learning and human judgement. SETC should maintain both without confusing them.

Existing Secondary 2 AI-English applications

These earlier SETC pages remain stage-specific applications under the Secondary 2 owner. They now sit beneath this Digital & AI English system without being rewritten:

• What Jobs Should AI Be Allowed to Do in English Tuition?
• When Should a Student Attempt English Work Before Using AI?
• How Should Students Check AI Feedback Before Acting on It?
• How Should Tutors Preserve Student Voice and Authorship When AI Is Used?
• How Should Tuition Test What a Student Can Do Without AI?

Together they apply this owner’s rules for attempt-first work, tool boundaries, feedback verification, authorship and unaided transfer to a specific school-stage context.

Cross-stage practice

When Algorithms Choose — Read, Verify, Debate and Write is the practical learning-pack application: learners read claims about recommendation systems, verify evidence, debate trade-offs and produce independent writing rather than treating an algorithmic output as an answer.