Task statement 2.2 · Output evaluation and validation

2.2 · Hallucinations, Inconsistencies and Bias

In Wrenfield Housing, the course example, the communications team asks Claude to turn the tenant handbook into a draft Q&A of ten common requests from tenants. A fluent draft can still cite a clause that does not exist, give two values for one fact, or work for only some tenants. A person must catch all three before tenants see it.

Why it's tested

The tempting check is to read the draft once and ask whether it sounds right, because a fluent answer feels like a correct one. Claude writes an invented detail in the same sure voice as a true one, so the tone gives no warning. That read also misses a contradiction between two answers far apart, and it never asks who the draft leaves out. Doubting every sentence equally fails too, because the check gets so slow that people skip it.

KnowledgeThree problems a fluent response can hide

Anthropic's docs define a hallucination as text that is factually incorrect or inconsistent with the context Claude was given. The Claude Help Center explains why it happens. In trying to be helpful, Claude can occasionally produce incorrect or misleading responses, including quotes that seem authoritative but have no basis in fact. The invented part is often the most specific part: a clause number, a quotation, a figure or a name. The sign is therefore a precise detail that nothing you gave Claude contains. Anthropic's docs add that good prompting reduces hallucinations but does not remove them, which rules out treating an exact-looking reference as proof.

An inconsistency is two statements that cannot both be true, such as two deadlines for the same task. They may sit in one response or in two responses to the same request. Anthropic's docs treat it as a warning: when the same prompt gives different outputs, the difference could indicate a hallucination. A value taken from your material comes out the same each time, while a guessed value can change each time it is needed. The sign is one fact stated twice with two values, often far apart, so you find it only by setting the two places side by side.

Bias is a slant in a response that favours, disfavours or leaves out a group of people or a point of view. An Anthropic study found that in some test settings, a Claude model's decisions changed when only the demographic details in the prompt changed. Careful prompting reduced this effect. In everyday drafting, bias more often shows up as an assumed reader. Claude has no context about you, your task or your organisation, so it writes for a typical reader it imagines, and anyone unlike that reader can drop out of the text. The signs are an instruction only some readers can follow, a group described by a stereotype, or one side of a question argued at length while the other gets a line.

A reviewer finds all three in four passes:

  1. Mark the specifics: every name, number, date, quotation and reference. Ask whether the material you gave Claude contains each one.
  2. Pair the repeats: find each fact stated more than once, and compare the values.
  3. Picture the readers: ask who the draft assumes is reading, and which readers it would fail or describe unfairly.
  4. List what you found: each flagged item is then checked against a source (2.3), never sent on unchecked.

Interactive · where each problem shows in the draft Q&A

Four answers from the draft tenant handbook Q&A on the left, in question order from top to bottom, each with an arrow to the sign a reviewer looks for on the right. Q5 cites clause 14.3 of the tenancy agreement, a reference not in the handbook Claude was given: a hallucination. Q2 says four weeks' notice and Q7 says two weeks for the same act, linked by a line labelled same fact: an inconsistency. Q9 offers only the online form and email replies, assuming every tenant is online: bias. A flagged answer's sign is dashed and labelled flag; the other answers fade. THE DRAFT Q&A THE SIGN A REVIEWER LOOKS FOR Q2 · Ending your tenancy Give four weeks' notice in writing. SIGN One fact stated twice notice also given in Q7 FLAG: INCONSISTENCY Q5 · Decorating your home Clause 14.3 of your tenancy agreement lets you decorate. SIGN A reference not in the source Claude saw only the handbook FLAG: HALLUCINATION Q7 · Moving to another home Give two weeks' notice on your current tenancy. SIGN One fact stated twice differs from Q2's four weeks FLAG: INCONSISTENCY Q9 · Reporting a repair Use our online form; we will reply by email. SIGN One reader assumed no phone or letter route FLAG: BIAS SAME FACT 4 ≠ 2 Each problem has its own sign, so the reviewer looks for signs, not for a feeling.

Choose a problem to see which answer shows it, and its sign.

Hallucination: Q5 cites clause 14.3 of a tenancy agreement Claude never saw, and the agreement has no such clause, however exact it looks.

Inconsistency: Q2 says four weeks' notice and Q7 says two weeks for the same act, ending a tenancy, so at least one is a guess.

Bias: Q9 offers only the online form and email, so tenants who would phone or write are left out.

Dashed and labelled FLAG: the answer that shows the chosen problem. Faded: answers that do not.

KnowledgeTelling the three apart

Each problem leaves a wrong or unfair sentence, but the fault sits in a different place. A hallucination is one claim with no support. An inconsistency is a pair of claims that conflict, so at first you may not know which one is wrong. Bias can sit in sentences that are each true, because the problem is who the text serves. A read for facts alone misses it. Naming the problem tells you what to compare the sentence with:

Hallucination

A precise detail, such as a clause, quotation or figure, that nothing you gave Claude contains.

Compare withThe material Claude was given, then the real source.

Inconsistency

One fact with two values, in one response or across two responses to the same request.

Compare withThe other place the fact appears, then the source to see which is right.

Bias

An assumed reader, a stereotype, or one side of a question given most of the space.

Compare withThe full range of people the text is for.

SkillWorked example: the tenant handbook Q&A

Wrenfield's communications team attached the current tenant handbook and asked Claude for a Q&A answering ten common requests from tenants. The draft read well. The handbook says tenants give four weeks' notice in writing to end a tenancy, ask Wrenfield before decorating, and can report a repair by phone on the repairs line, on the online form or by letter:

# before: four answers from the draft
Q2 To end your tenancy, give four weeks' notice in writing.
Q5 Clause 14.3 of your tenancy agreement lets you decorate freely.
Q7 Moving to another Wrenfield home? Give two weeks' notice.
Q9 Report repairs on our online form; we will reply by email.
# after: each flagged answer checked against the handbook
Q5 Please ask us before you decorate.
Q7 Moving to another Wrenfield home? Give four weeks' notice.
Q9 Report a repair on the repairs line, the online form or by letter.

Claude was shown only the handbook, and the tenancy agreement has no clause 14.3. Q5 is therefore a hallucination, given away by a precise reference to a document nobody attached. Q2 and Q7 are an inconsistency, because ending a tenancy needs the same notice whether the tenant leaves Wrenfield or moves to another of its homes. Q2 matched the handbook, so Q7 was the guess. Q9 is bias by omission: it is true, but it was written for a tenant who is online, so the phone and letter routes dropped out. The communications lead marked all three and confirmed each against the handbook before approving the Q&A, as Wrenfield's AI policy requires.

A fix can carry its own bias. The quick repair for Q9 was a line saying "older tenants can phone instead", which swaps one assumption for a stereotype about age. The repair that works lists every route for every tenant and lets each reader choose.

Ask one question: does every specific detail appear in the material you gave Claude, say the same thing everywhere it appears, and work for every reader the text is for? Yes → the draft goes on to its checks and review. No → name the problem as a hallucination, an inconsistency or bias, and check that answer against the source before anyone sends it.
Takeaway
  • A hallucination is a fluent, specific claim with no basis in what Claude was given, so exact-looking references are where to look first.
  • An inconsistency is one fact with two values, and it often means at least one of the two is a guess.
  • Bias can hide in true sentences: ask who the draft assumes is reading, and fix it without adding a stereotype.

check yourself

0 / 3 answered

A charity fundraiser attaches the charity's annual report and asks Claude for a paragraph for a grant application. The draft quotes the chair: "Every child deserves a reading mentor." Web search was off, and the annual report contains no quotation from the chair at all.What is the most likely root cause?

Select one.

An operations manager asks Claude to turn a new hybrid-working policy into a one-page staff guide. The guide gives the required office days in two sections, and every instruction starts "On your laptop". The policy sets three office days and lets staff without a laptop use the office desktops.Which TWO checks would most directly surface the problems this evidence points to?

Select two.

A library outreach worker asks Claude for a flyer about a free course on email and online forms. The draft says the course is "perfect for seniors who struggle with technology". The worker wants the flyer to reach everyone new to these tools, without suggesting that any age group lacks skills.Which is the most proportionate fix?

Select one.

sources

  1. Claude Platform Docs, "Reduce hallucinations": hallucination as text "factually incorrect or inconsistent with the given context"; inconsistencies across outputs of the same prompt could indicate hallucinations; the techniques reduce hallucinations but do not eliminate them. Checked 2026-10-15.
  2. Claude Help Center, "Claude is providing incorrect or misleading responses. What's going on?": in trying to be helpful, Claude can occasionally produce incorrect or misleading responses; training data may not be up to date; quotes that seem authoritative but lack a factual basis. Checked 2026-10-15.
  3. Anthropic, "Evaluating and mitigating discrimination in language model decisions": answers that varied with demographic details in select settings, both positive and negative discrimination; careful prompting reduced both; Anthropic does not endorse or permit automated decisions for the high-risk uses studied. Checked 2026-10-15.
  4. Anthropic, "Measuring political bias in Claude": the aim that Claude treats opposing viewpoints with equal depth, engagement and quality of analysis. Checked 2026-10-15.
  5. Claude Help Center, "Introduction to prompt design": Claude has no context about you, your task or your organisation. Checked 2026-10-15.
  6. Wrenfield's handbook Q&A, clause 14.3 and the notice periods are this course's example.