English:Generative AI language models and hallucinations - AI Year 6
Introduction

Can a polished answer cite a book that its source does not contain? Investigate eight original fictional claims, repair five unsupported statements, keep justified information and test what changes when a later source appears. Learn why generation, retrieval and factual checking are different.
Separate generation from evidence
Follow the path from a request to a candidate answer
An instruction tells a system what you want it to do. Context is the information available for that response, such as a supplied passage or previous conversation. A generated answer is candidate text. External evidence is material you inspect to decide whether a factual statement is supported. Keep those four roles separate: an answer can repeat the words of your question without supplying evidence for them.
A language model uses patterns learned during training to generate sequences of tokens. Tokens can be whole words, parts of words or characters. The surrounding context affects which continuations are likely. This is a useful starting explanation, not a claim that every application works identically or that a likely continuation has been checked against a current source. A fluent sentence and an accurate statement are different achievements.
Suppose a request asks for a library's current opening time and a source. An answer may produce a convincing time followed by a plausible book title. Neither the confident tone nor the shape of that reference proves that the time or book was checked. Some applications can retrieve documents or use tools as well as generate language. You still need to open the returned source and check that it supports the exact claim, for the right place and date.
CodeAI: How Chatbots and Large Language Models Work. With teacher support, identify where generation differs from checking a document. The publisher's wider material is aimed at older grades; our concrete source-card investigation needs no learner chatbot account, private prompt or live model call.
Understand the term without giving the computer a human motive
Here hallucination refers to generated false or misleading material presented as factual. An outdated answer, invented citation and unsupported extension of a source all deserve attention, although their causes may differ. Do not diagnose a system's internal cause from one sentence alone. We do not need to say that a computer deliberately lied or believed a story in order to check the output.
A request to be accurate can help describe the task, but it is not an accuracy test. Asking the same system whether its answer is correct is not independent evidence either. A real check records the claim, relevant source, date, what the source actually says, and a verdict with an appropriate limit. When evidence is unavailable, leave the claim unverified and ask an appropriate adult or subject specialist rather than fill the gap with another confident sentence.
For this lesson all Westhaven records, authors, book titles and answer claims below are original fictional classroom materials. They are not real catalogue entries, not output measured from a particular model, and not the photographed library's records. We can practise a careful verification method without making public claims about a real person or organisation.
Investigate a complete claim and source ledger
Read the supplied evidence before the polished answer
The question is: what can we responsibly say about the fictional Westhaven club reading room and catalogue on 15 November 2026? These three cards are the complete evidence supplied for the first investigation. Dates and scope matter. N2 changes Thursday hours only and explicitly preserves the other N1 conditions. None of the cards reports live borrowing availability.
| Card | Issued | Original fictional record |
|---|---|---|
| C1 | 1 September 2026 | Westhaven Learning Library catalogue: River Windows, Ava Reed, 2024; Clockwork Orchard, Theo Moss, 2023. This is a selected classroom catalogue, not a world bibliography. |
| N1 | 1 October 2026 | Westhaven club reading room: Tuesday and Thursday 15:00 to 17:00; twelve seats; entry free for registered club members. No borrowing-availability record. |
| N2 | 1 November 2026 | Westhaven club update: from today Thursday is 16:00 to 18:00. Tuesday hours, seat count and member-entry condition in N1 are unchanged. |
Do not merge the selected catalogue with all books in the world. Do not extend one club's membership condition to every library. Record both N1 and N2 when explaining the Thursday change, so a reviewer can see why the earlier time no longer applies to the stated date.
Break an answer into separately checkable claims
Treat each statement below as part of a teacher-prepared candidate answer. Read K1 to K8 and try your own verdict column before the worked review. Hint: underline a date, a quantity or a word such as every. Partial step: compare the scope of K5 with the exact membership condition in N1.
| Claim | Prepared candidate statement |
|---|---|
| K1 | C1 lists River Windows by Ava Reed, published in 2024. |
| K2 | Moonlit Bridges by Ava Reed, 2025, is verified by C1. |
| K3 | The reading room opens on Thursday at 15:00 on 15 November 2026. |
| K4 | The room has thirty seats. |
| K5 | Every library is free for everyone. |
| K6 | This is the best reading room. |
| K7 | N2 takes effect on 1 November 2026. |
| K8 | River Windows is available to borrow right now. |
The claim identifiers let you discuss errors without rewriting the original answer or losing which version was checked. A claim-evidence-verdict ledger should also preserve the relevant source card and a proposed correction. An authoritative-looking reference is still only a reference-shaped string until the source is found and inspected.
Worked review: correct precisely and retain the limits
| Claim | Verdict | Evidence inspected | Reason or appropriate action |
|---|---|---|---|
| K1 | Supported | C1 | Keep the precise catalogue statement. |
| K2 | Unverified citation | C1 | C1 does not list it; remove the unsupported citation, without claiming the book cannot exist elsewhere. |
| K3 | Outdated | N1 and N2 | For this date N2 changes Thursday opening to 16:00. |
| K4 | Contradicted | N1 and N2 | N1 says twelve; N2 explicitly leaves the seat count unchanged. |
| K5 | Unsupported scope | N1 | N1 supports only this club room and registered members. |
| K6 | Needs clarification | No comparison criteria supplied | Ask best for whom and for which purpose; do not present an undefined comparison as established fact. |
| K7 | Supported | N2 | Keep the dated update statement. |
| K8 | Not established | C1 and N1 | A catalogue entry is not a current borrowing-availability record. |
K2, K3, K4, K5 and K8 are five distinct claims that must not pass unchanged as supported facts. K1 and K7 are supported within this fictional file. K6 needs a stated comparison purpose and criteria. These eight prepared cases are not a measured hallucination rate for any model or a promise about future outputs.
A suitable corrected answer is: the selected C1 catalogue lists River Windows by Ava Reed, 2024. According to N1 and the N2 update, on the stated November date the reading room opens Tuesday 15:00 to 17:00 and Thursday 16:00 to 18:00, has twelve seats and offers free entry to registered club members. C1 does not verify Moonlit Bridges, and these cards do not establish current borrowing availability. An undefined best claim should be removed or clearly reframed as a question about needs.
Notice what was not corrected by guessing. We did not substitute a different invented book as evidence, assert that no such book exists anywhere, invent a live stock check or turn a club condition into a national rule. A smaller accurate statement is more useful than a longer unsupported one.
Test source boundaries and transfer the check
Change one source and revisit only affected conclusions
Consider a separate boundary case. A new fictional catalogue card C2, dated 20 November 2026, now lists Moonlit Bridges by Ava Reed, 2025. If C2 is an inspected authentic card in that later fictional investigation, it supports that entry then. It does not show that C1 contained the entry, or retroactively verify the original answer's citation to C1 on 15 November. Retain both versions and the date of each decision.
This boundary demonstrates why not found in the supplied catalogue is a limited result, not proof of nonexistence everywhere. If a real source cannot be opened, label it inaccessible or unverified. If two sources disagree, compare authority, date, purpose and scope; state the unresolved issue if you cannot settle it. Do not silently choose the sentence that matches the generated answer.
Another boundary is a genuine source attached to the wrong claim. N2 really supports the date of an opening-time change in our file; it does not verify the author of River Windows. Finding a real document is necessary for a citation check but insufficient for claim support. Ask where in that source the exact statement is established.
CodeAI: AI: Training Data & Bias. Consider which facts or perspectives may be missing and why polished wording does not repair missing evidence. This is conceptual context, not the source of the Westhaven cards or an endorsement of our fictional candidate answer.
Transfer the method to a different subject
Imagine an original fictional museum notice M1: an object is a replica made in 1998, after a vessel catalogued as medieval; no specific medieval year is supplied. A candidate answer calls the displayed object an original made in 1298. Split the claim: original conflicts with replica, while the exact year 1298 is not supplied. An appropriate correction describes a replica made in 1998 after a medieval vessel, and leaves the precise medieval date unresolved. Do not cite the library cards for museum facts.
The method transfers, but the relevant evidence changes. A catalogue may establish an edition, while a dated institution notice may establish current hours. For historical or scientific questions, ask which primary record or appropriate expert source can support the particular claim. A generated explanation can help you formulate the question; it cannot replace the evidence you need.
Produce a useful evidence log and a safe school workflow
Retain the original request, candidate statements, dated cards, separate verdicts, revised answer and remaining questions. Record which checks were actually performed. A proposed future librarian interview or model comparison remains unperformed until it happens. Do not label a worksheet, an inaccessible citation or a prompt instruction as an independent expert check.
Use the supplied fictional text on paper or in a teacher-approved document. Any separate live generative-AI activity requires the school's approval, age-appropriate conditions, supervision and suitable safeguards. Do not upload pupil identities, private work or sensitive details to test whether a chatbot will repeat them. The source-checking learning goal is achievable without doing so.
Before sharing a factual answer, ask: is the source accessible, does it support this precise statement, is it current for the task, and what remains unknown? Keep justified information, correct supported errors and remove or clearly flag unresolved claims. This editorial Year 6 enrichment develops critical judgement; passing its quiz does not make future generated text automatically reliable.
Interactive Tasks
Quiz: Test Your Knowledge
What is a generated answer before its factual claims are checked? (Candidate text requiring appropriate evidence) (!An automatically verified current source) (!A guarantee that every citation was opened) (!A complete record of every relevant fact)
What can a token represent? (A word part or character as well as a whole word) (!Only a complete book) (!Always exactly one sentence) (!A verified reference by definition)
Which check is independent of the generated answer? (Inspect a relevant external source for the exact claim) (!Ask the same answer to praise its own accuracy) (!Trust confident wording) (!Count how many references look formal)
Which source supports the November Thursday opening time? (N2 read together with the unchanged N1 conditions) (!The selected book catalogue alone) (!The generated time without evidence) (!The library photograph alone)
What follows from C1 not listing Moonlit Bridges? (That C1 does not verify the claimed entry) (!That the book cannot exist anywhere) (!That a different invented citation is acceptable) (!That all books in C1 are false)
Which seat count is supported by the supplied notices? (Twelve) (!Thirty) (!Twenty) (!No room exists because one claim was wrong)
Why is every library is free for everyone unsupported? (N1 concerns one club room and registered members) (!All broad claims are automatically accurate) (!A formal reference permits any conclusion) (!The seat count proves national policy)
What does C2 dated later change? (It can support its own later entry without rewriting C1) (!It proves C1 always contained that entry) (!It makes all original claims correct) (!It removes the need to inspect dates)
Does a current catalogue entry establish borrowing availability now? (No current availability needs its own evidence) (!Yes every listed book is always available) (!Yes if the answer sounds fluent) (!Yes if the title has two words)
How should the museum candidate answer be revised? (Describe the 1998 replica and leave the exact medieval year unresolved) (!Keep the unsupported year because it is precise) (!Call the replica an original) (!Use library opening hours as proof)
Memory Game
| Instruction | The task requested of the system |
| Context | Information available for the response |
| Token | A unit used in modelling and generating language |
| Candidate answer | Generated text awaiting appropriate checking |
| Evidence ledger | A record connecting claims sources verdicts and limits |
| Citation | A reference whose existence and exact support still need checking |
Drag and Drop
| A convincing reference absent from the supplied catalogue | Mark that citation unverified without proving global nonexistence |
| An old opening time changed by a later notice | Use the applicable dated update |
| A title entry without a live stock record | Leave current borrowing availability unknown |
| A real notice attached to an unrelated author claim | Check exact claim relevance rather than mere source existence |
| An undefined best comparison | Ask for purpose and comparison criteria |
Crossword Puzzle
| TOKEN | What unit can be a word part or character? |
| CONTEXT | What surrounds the task and informs a response? |
| SOURCE | What external material must be inspected for support? |
| CITATION | What reference can look plausible yet remain unverified? |
| VERDICT | What judgement records support contradiction or uncertainty? |
| REPLICA | What is the fictional museum object rather than an original? |
LearningApps
Optional extra practice. Choose an appropriate activity with your teacher. If the external collection is unavailable, use the native activities and independent tasks in this course.
Cloze Text
Open Tasks
Easy
- Separate four roles: Mark instruction, supplied context, candidate answer and independent evidence in a short original example.
- Explain a token: Explain why a token need not equal a whole word without claiming to inspect a real model tokenizer.
- Check a quantity: Use N1 and N2 to review K4 and give a precise correction.
- Limit an absent citation: Explain K2 without turning absence from C1 into proof of nonexistence everywhere.
Standard
- Review eight claims: Build the complete K1 to K8 ledger with source, date, verdict and reason.
- Repair five unsupported claims: Correct or flag K2 K3 K4 K5 and K8 without inventing replacement facts.
- Write a bounded answer: Write a concise November answer preserving membership conditions and unknown availability.
- Test the later catalogue: Compare what C1 on 15 November and C2 on 20 November actually support.
Advanced
- Challenge a real-looking source: Explain why N2 can be genuine yet irrelevant to a book-author claim.
- Transfer to the museum: Separate replica status and dates, correct the supported error and preserve what remains unknown.
- Design a disagreement check: Propose source-authority, scope and date checks when two future records conflict; label the proposed test unperformed.
- Build a reusable review protocol: Produce an accessible evidence-log template with an uncertain verdict, human escalation and a school-safe non-digital route.
Learning Check
- Generation and evidence: Explain why likely continuation is not factual verification.
- Five claims needing intervention: Review the unsupported factual claims separately.
- Supported and ambiguous claims: Keep what is justified without accepting everything.
- Version and citation boundary: Interpret a later matching catalogue entry.
- Museum transfer: Correct the different fictional output.
- Useful review evidence: Design a proportionate check and handoff.
Worked checks and assessment guidance
- A model generates token sequences using learned patterns and available context. Retrieval may provide documents in some applications, but a fluent answer or citation-shaped string does not establish that the exact claim was checked. Inspect relevant independent evidence and record its limits.
- K2 is not verified by C1, without a global nonexistence conclusion. K3 is outdated for the stated November date because N2 changes Thursday to16:00. K4 conflicts with twelve unchanged seats. K5 extends one club membership condition to every library. K8 treats a title entry as live availability, which the cards do not record.
- K1 is supported by C1 and K7 by N2. K6 needs a defined audience, purpose and comparison criteria; do not certify an undefined best claim. Two supported statements do not make the whole answer reliable, and one false statement does not automatically falsify all other statements.
- An inspected C2 may support its dated entry in the later fictional investigation. It does not retroactively add the title to C1 or validate the old citation. Preserve dates, exact source identity and the original verdict rather than rewriting history.
- M1 says replica made1998 after a medieval vessel. Original is contradicted and1298 is not supplied. Describe the1998 replica and leave the exact medieval date unresolved. Use museum evidence rather than unrelated library records; a real source still needs claim-level relevance.
- Retain request, original statements, accessible dated sources, exact support or contradiction, revised wording and remaining questions. Flag inaccessible sources and seek an appropriate adult or expert where needed. Future interviews/model trials are not performed evidence. Use fictional classroom materials rather than uploading private pupil data.
Learning evidence
Retain the eight-claim ledger with exact source cards and dates, five corrected or flagged factual claims, two precisely supported claims and an ambiguous comparison. Include a bounded revised answer, later-catalogue boundary and museum transfer with remaining unknowns. Distinguish actually inspected evidence from proposed checks. No real model trial, institution record, live availability check or independent expert certification is established.
OER and Wikipedia
Google: introduction to language models informs the brief conceptual explanation of tokens, context and generation. DfE: generative AI in education informs critical judgement and school safeguards. These sources do not supply or verify the invented library and museum records. Read 15 November 2026. All source cards, candidate claims, review ledgers and worked corrections are original classroom simulations, not measured responses from a named model or real institution records.
International books at Kent Library.jpg, Roc0ast3r, CC0. Unchanged contextual library photograph; not Westhaven, not evidence for fictional opening times or invented books.
https://www.youtube.com/watch?v=X-AWdfSFCHQ — CodeAI: How Chatbots and Large Language Models Work. Original recording embedded unchanged.
https://www.youtube.com/watch?v=x2mRoFNm22g — CodeAI: AI: Training Data & Bias. Original recording embedded unchanged. Teacher-supported optional conceptual context; no live learner account required.
Linked learning areas
aiMOOC Projects
Schulfach+


aiMOOCs


aiMOOC Projekte


MOOCwiki · Deutsch
Nach dem Lernen ist vor dem Lernen
Entdecke direkt den nächsten Lernkurs. Weitere Inhalte erscheinen, wenn Du weiter nach unten scrollst.
Zur MOOCwiki-HauptseiteMediathek
Mediathek wird aus dem Wiki geladen ...
Keine passenden Inhalte gefunden. Bitte ändere Suche oder Filter.
NEWSLernweltNOAH fragen