7AAVCD42 Chap.3 Language Models, Prediction and Hallucination
Language Models, Prediction and Hallucination
Fluency is one performance criterion
Transformer-based large language models calculate relations in context and generate plausible continuations. This chapter compares that prediction mechanism with factual grounding and human understanding. It gives you a claim audit that separates readable structure, verifiable assertions, framing and intended use.
The evidence section then shows how to discuss hallucination and bias without inferring frequency or cause from a single output. The aim is precise evaluation rather than celebration or dismissal.
Separate generation from verification
A generated answer can be grammatical, coherent and responsive because those properties follow from learned language patterns and context. Factual trust requires a separate route.
Extract each checkable claim, including names, dates, quotations, citations and legal duties. Match it to an appropriate source. Do not ask the same model to certify an unsupported reference because a second fluent answer does not create independent evidence.
Use depends on the task. Brainstorming may tolerate suggestions that are later checked. A summary must remain faithful to the supplied document.
A current policy claim needs an authoritative and dated source. A submitted assignment also raises authorship and module rules. State the intended use before judging whether remaining uncertainty is acceptable.
Diagnose hallucination precisely
Reserve hallucination for generated content that presents an unsupported assertion as credible.
A shallow response, a disputed interpretation and a biased frame are different weaknesses. Record the exact sentence, the evidence needed and the consequence of relying on it. This produces an auditable finding and avoids treating all model limitations as the same problem.
A claim ledger can make the work visible. Mark claims as confirmed, corrected or unresolved. Add the source type and date.
The ledger reveals whether generation saved effort or shifted effort into checking. It also gives the report a transparent basis for deciding which output can be used as orientation and which should be discarded.
Evaluate bias through comparison
One stereotyped answer demonstrates possibility under the recorded prompt and model conditions. It does not establish frequency or cause.
A stronger investigation uses repeatable prompts, comparison conditions and a coding rule. Qualitative analysis can still examine one case closely, but its conclusion should remain at case scale and should distinguish description from causal explanation.
Look for defaults, omissions and unequal authority. Which identity is treated as normal? Which sources are presented as expert? Which affected group is absent?
Then connect the pattern to a plausible mechanism in data, model design, prompting or deployment. Match the recommendation to that mechanism.
Missing representation may require new evidence, while an organisational incentive will not be repaired by dataset changes alone.
Write within the module boundary
The module permits selective use for learning technical concepts while requiring submitted wording to be your own. Use tools within the current instructions, preserve your source trail and verify factual claims independently.
A good coursework paragraph explains the prediction mechanism, tests the evidence and shows why the social consequence depends on a particular use rather than on language models in general.
What this chapter covers
- 01
Transformer Architecture
- 02
Next-Token Prediction
- 03
Factual Grounding
- 04
Hallucination
- 05
Model Bias
Worked example · free
Audit a confident policy summary
- 2Extract verifiable claims.
- 3Check authoritative sources.
- 2Match trust to intended use.
Key terms
- Transformer Architecture
- Transformer Architecture processes relationships across a context and underpins contemporary large language models.
- Factual Grounding
- Factual Grounding connects a generated assertion to reliable evidence about the world or supplied material.
- Model Hallucination
- Model Hallucination is fluent generated content that presents an unsupported fact or source as credible.
Language Models, Prediction and Hallucination FAQ
Why can fluent text still be wrong?
Language models optimise plausible continuation from learned patterns and context. Grammar and coherence can succeed even when a date, quotation, source or causal claim lacks grounding.
How should I test a generated claim?
Extract the exact assertion, identify the evidence needed and check an appropriate authoritative source. Then decide whether any remaining uncertainty is acceptable for the intended use.
Can one biased output prove a general pattern?
One output shows that the behaviour occurred under recorded conditions. Frequency and cause require repeated comparisons, a coding rule and evidence about training or deployment mechanisms.
Assessment move
Take one generated paragraph, colour-code claims, sources, framing and qualifications, then rewrite its conclusion at the scale the verified evidence supports. Build a claim ledger with confirmed, corrected and unresolved columns. Check every citation outside the model and record the source date. Repeat the prompt with one controlled change, then code whether the output shifts in fact, framing or tone.
For bias practice, define the comparison before collecting examples and preserve outputs that do not fit your initial expectation. Rehearse the difference between possibility, frequency and cause aloud. Before using any generated material for study, state the intended use and the harm of error. Keep submitted wording your own and follow the current module-specific boundary on KEATS.
Add one source that challenges the dominant framing and explain whether it changes a fact, interpretation or recommendation. This guards against treating verification as a narrow search for confirmation. Finish by deleting every unresolved generated assertion from the draft evidence table.