CCAO-F 2.2 — Spotting hallucinations, inconsistencies and bias

Naming the problem

What is wrong with this part of the answer?

A specific with no source
Possible hallucinationverify it or remove it
Two statements disagree
Inconsistencygo back to the data
One-sided or flattering
Possible biasre-ask neutrally, both sides
None of these
Carry on checkingaccuracy and completeness still apply

Each kind of problem has its own tell. Naming it tells you what to do next: find a source, find the contradiction’s root, or re-ask neutrally.

Hallucinations: fluent, specific and unsupported

A hallucination is content that is presented as fact but is not grounded in anything real. Anthropic’s help centre describes two common forms: statements and quotes that sound authoritative but have no factual basis, and confusion about recent events, because each model is trained on data up to a cutoff date and may not know what happened after it. The dangerous thing about a hallucination is that it is written in the same confident voice as everything else.

Hallucinations are not spread evenly. They cluster where the model is filling a gap with something plausible — and those places are predictable.

Red flag	Example	Why it is risky
Precise numbers with an attribution	“A 2025 industry survey found 47% of buyers…”	Plausible statistics are easy to generate and hard to spot
Exact quotes from real people	A CEO “said” a crisp line in an interview	Invented quotes are one of the failure types Anthropic names
Legal and regulatory references	“Under Section 14(3)(b) of the Act…”	Section numbers look authoritative and are rarely checked
Citations, titles, URLs	A report title or link that sounds real	A made-up reference looks exactly like a real one
Recent events	Last month’s price change or leadership change	May be after the model’s training cutoff
Details “from” your documents	“Clause 9 allows early termination”	The document may say something different, or nothing

Notice that the red flags are about the kind of claim, not the tone. A good first scan runs through the answer and marks every name, number, date, quote and reference. Those are the claims that must trace to a source; the connecting prose around them rarely matters as much. How you then verify each one is covered in 2.3.

Worked example · The returns policy with a precise-sounding law

Dana, a support team lead at an online retailer, asks Claude (with no web search and no documents) to draft a customer-facing note explaining the legal right to return goods bought online in her country.

The draft reads well. It states that customers have “a 30-day statutory right to return under Section 12(4) of the Consumer Protection Act”, and adds that “the regulator issued new guidance in March”.
She marks the specifics: a period (30 days), a section number, a named act, and a recent event. Every one of them is exactly the kind of claim that hallucinates.
She checks the recent event first. Claude had no web search, so anything about “March” is either from before its training cutoff or invented. She cannot tell which from the text alone.
She checks the law on the regulator’s website. The act exists, but the return period and the section number are different from the draft.
She rewrites the prompt to use the regulator’s page as the source, and has the final wording checked by the legal team before it goes to customers.

Takeaway: Nothing in the draft looked wrong. The only reliable signal was the type of claim — a precise legal reference and a recent event — which is where she concentrated her checking.

Inconsistencies: when the answer disagrees with itself

An inconsistency is two parts of the output — or two outputs — that cannot both be true. It is the easiest problem to detect because you do not need an outside source: the contradiction is on the page. It is also a strong clue, because at least one of the two statements is wrong, and often the error behind it is a hallucination or a misread figure.

Where to look	What it looks like
Summary vs detail	The executive summary says revenue grew 12%; the table shows 9%
Numbers vs arithmetic	Line items that do not add up to the stated total
Across the conversation	An assumption stated early (“prices exclude VAT”) quietly changes later
Across runs	The same question asked twice gives two different figures or conclusions
Output vs source	The response describes a document in a way the document does not support

The cross-run check is one Anthropic recommends explicitly: its guidance on reducing hallucinations suggests running the same prompt more than once and comparing, because inconsistencies between runs can indicate a hallucination. Some variation in wording is normal and harmless. Variation in facts, figures or conclusions is not.

A consistency pass on a board memo

Fails: Summary figures match the detailed tablesummary 12% growth; table 9%
Passes: Totals equal the sum of their parts
Check: Assumptions stay the same throughoutVAT excluded on p.1, included on p.3
Passes: A second run gives the same conclusionwording differs; figures and recommendation match
Passes: Dates and names spelled the same everywhere

Every item here can be checked without leaving the document — which makes a consistency pass the fastest check you can run.

Bias: a skew you might not notice

Bias is harder to spot than a wrong number, because each sentence can be true while the whole is lopsided. For business users, four forms matter most.

Agreeing with you (sycophancy). Anthropic’s research found that five leading AI assistants consistently tended to favour responses that match the user’s stated views over accurate ones — a side effect of training on human preferences, since people tend to rate agreeable answers highly. A question that assumes its own answer tends to get that answer back.
One-sided treatment of contested questions. Claude is trained to be even-handed on political and contested topics and to present multiple perspectives, and Anthropic measures this. That is a design goal, not a guarantee; check whether each side got comparable depth.
Unfair treatment of people. In an Anthropic study of hypothetical decisions such as loans and hiring, an earlier Claude model showed patterns of both positive and negative discrimination when demographic details varied. Prompting changes reduced it significantly, and Anthropic states it does not endorse using language models to make automated decisions in these high-risk areas.
Skewed inputs. An answer can only be as balanced as its sources. If the documents you uploaded all come from one team, or the web results all come from one kind of site, the output inherits that slant.

Leading versus neutral framing

Invites agreementtext

Explain why switching to Vendor A
is the right decision for us.

Invites an assessmenttext

Compare Vendor A and Vendor B for our
support desk, using only the attached
proposals.

Criteria: total 3-year cost, contract
exit terms, integration effort, SLAs.

Make the strongest case for each vendor.
Then say where the evidence is thin, and
what would change your recommendation.

The left prompt assumes the conclusion, so an agreeable answer is the likely result. The right prompt asks for both sides against stated criteria and for the weak spots in the evidence.

Pushback reveals sycophancy too. If you reply “I don’t think that’s right” and Claude reverses a well-supported answer without new evidence, the reversal tells you nothing about which answer is true. Claude is designed to be truthful and calibrated rather than to tell people what they want to hear, but you evaluate the output in front of you, not the design goal. A good answer should change when you give it new facts, not when you give it disapproval.

Worked example · Interview summaries that describe people differently

An operations manager uses Claude to turn interview notes on six candidates into one-paragraph summaries for a hiring panel. The notes include names and some personal details.

Reading the summaries side by side, she notices that two candidates who did similar things are described differently: one is “confident and decisive”, the other “somewhat forceful”.
She checks the notes. The underlying behaviour described is nearly identical. The difference is in the summary, not the evidence — a skew she would have missed reading one summary at a time.
She changes the process: removes names and personal details before summarising, gives Claude the same scoring criteria for every candidate, and asks for evidence quoted from the notes under each criterion.
She keeps the decision human. The panel uses the summaries as an aid and makes the hiring decision itself — the kind of use where review is required, covered in 2.4.

Takeaway: Bias often shows up only in comparison. Reading outputs side by side, with the same criteria for each, is how you see it.

Traps the wrong answers are built from

Tempting but wrong	Do this instead
Expecting errors to look odd or hesitant	Target the claim types that hallucinate: specifics, quotes, references, recent events.
Asking Claude “are you sure?” to settle it	Check the source; a reversal under pushback is not evidence either way.
Letting Claude choose between two contradictory figures	Go back to the data, find which is right, and fix both places.
Asking a question that assumes its answer	Frame neutrally, give criteria, and ask for the case on each side.
Reviewing outputs about people one at a time	Compare them side by side against the same criteria, without demographic details.

You should now be able to

Distinguish a hallucination, an inconsistency and a bias in a given response.
Name the claim types most likely to be hallucinated and scan for them first.
Run a consistency pass: summary against detail, totals against parts, run against run.
Recognise sycophancy and leading prompts, and rewrite a request neutrally.
Spot unfair or one-sided treatment by comparing outputs against shared criteria.

Spotting hallucinations, inconsistencies and bias