Home › Study guides › CCAO-F › Domain 2 › Lesson 2.2

CCAO-F · Domain 2 · 21% of the exam · Lesson 2.2 · 20 min read

Spotting hallucinations, inconsistencies and bias

Why Claude can invent a statistic or a quote, how to catch totals that don't add up and one-sided framing, and the questions that make each problem show.

Written against objective 2.2 of the official CCAO-F exam guide (Version 1.0, effective July 2026). An independent resource, not affiliated with Anthropic; the practice questions are written from scratch.

2.2.1 Why a draft that reads well can still be wrong

You're the communications officer at a clean-water charity, and the annual impact report goes to donors in two weeks, with a press release alongside it. You give Claude the programme's monitoring spreadsheet, the field team's notes and last year's report, and ask for both drafts. Half an hour later they arrive: warm, clear and well organised, with an opening line you'd have been proud to write.

Then your programme director reads them with a pen. The press release credits the new wells with "a 73% drop in waterborne illness", yet the monitoring data doesn't track illness at all. A quote from the director of your health partner, the Maro Health Partnership, reads beautifully, but she never said it. The report's summary says you reached 48,000 people; the district table underneath adds up to 41,600. And every sentence about the communities you serve casts them as "grateful villagers" waiting for help, though the village water committees chose the well sites and keep the pumps running. The committees never appear.

One more thing surfaces. Early on you asked Claude to "explain why our programme is the most effective in the region". It obliged with four persuasive reasons, and "the region's most effective clean-water programme" reached the headline. Nobody has ever compared your programme with anyone else's.

None of this looked wrong on the page, and that is the whole difficulty. The draft holds three kinds of failure, each with a signature you can learn. A hallucination is a confident, specific-looking detail that isn't true, like the 73% or the quote. An inconsistency is a place where the output disagrees with itself, with its source or with another run of the same request, like the two totals. A bias is a systematic slant: a one-sided frame, a missing voice, a stereotype, or agreement with the premise of your own question.

2.2.2 Why Claude invents specific details

Here's the question that puzzles people most: nothing you gave Claude mentions illness, and it cited no web search, so where did 73% come from? Claude is a language model, and a language model doesn't look facts up in a store of records. It writes word by word, predicting what tends to come next. Anthropic's course on AI capabilities describes generative AI as closer to a vastly sophisticated autocomplete than to a search engine.

A document you gave it, or a web search it ran, steers that prediction toward what the source says. But when a sentence calls for a fact that nothing in front of Claude supplies, the writing doesn't stop. It produces what such a fact usually looks like: impact reports often claim a drop in illness, so one appeared. The same course says fabrication concentrates in specifics such as names, dates, statistics, citations, links and quotes. The more precise a claim, the more it needs checking.

Think of an actor playing a surgeon. Fluent lines and an assured manner tell you nothing about whether they could operate; the assurance belongs to the performance. Claude's fluency works the same way. The process that makes it fluent is the one that makes it fabricate, so a confident tone, tidy formatting and a named source appear whether the fact is solid or invented. The Help Center warns that Claude can display quotes that look authoritative or sound convincing but are not grounded in fact. Confidence is not evidence.

Even a sourced detail deserves a second look. The Help Center advises reviewing the pages Claude cites from a web search, because they may hold context its summary left out. A citation tells you where to look, not that the page says what the sentence claims.

2.2.3 When the draft disagrees with itself

Some errors need no outside knowledge to catch. If the summary says 48,000 and the table says 41,600, at least one is wrong, and you know it before checking a single fact. That makes an inconsistency free evidence of an error. Anthropic's guide to reducing hallucinations defines a hallucination as text that is "factually incorrect or inconsistent with the given context", so contradicting your file is as much a failure as inventing a statistic.

Inconsistencies turn up in three places.

Memorise the three places. Each catches problems the other two miss, and all three need only comparison, not research.

The third row needs care. Ask Claude the same thing twice and the wording will differ, because each answer is generated afresh; that's normal. The facts should stay put. Anthropic's guidance suggests running a prompt more than once and comparing the results, because inconsistencies across outputs can indicate hallucinations. When the illness figure comes back as 64%, you've learned it was never anchored to anything.

Think of a witness who gives a different date each time you ask. You don't pick the version you like best; you stop relying on their memory and go and find the record. That's why regenerating until a contradiction disappears fixes nothing. The clean second draft doesn't tell you which total was right.

2.2.4 When the answer leans one way

The hardest problems to see may have no false sentence in them. Most lines about the communities in your draft would pass a fact-check, yet the picture is one-sided: people who receive, never people who decide or maintain. That's bias: a systematic slant in what an output emphasises, leaves out or assumes. Anthropic's Academy points out that it can be as quiet as defaulting to certain perspectives, or giving one viewpoint a fuller answer than another.

It takes five common forms.

• One-sided framing. Only the favourable side is told, or one view gets the depth and the other a sentence.
• Missing perspectives. Someone the story is about has no voice in it, like the water committees.
• Stereotypes about groups. A stock picture stands in for evidence, like the draft's "grateful villagers", which your own records of the committees' work contradict.
• Skew inherited from the data. Models learn from huge amounts of text and can pick up patterns that tilt them. Your own material can tilt too: field notes written only by staff give a staff-eye view.
• Agreeing with your premise. This is sycophancy: telling you what you want to hear instead of what is accurate.

The portrait of the communities shows the first four forms at once, and the fix isn't a warmer adjective. It's the missing facts: the committees chose the sites and keep the pumps running, so the report should show them at work, as partners rather than recipients.

The fifth form is the one your own question can cause. Anthropic's Academy gives examples of sycophancy: agreeing with a factual error you've made, or changing an answer because of how you phrased the question. It's more likely when a question is framed from one point of view, asks for validation or states an opinion as fact, and when a conversation gets very long. Your request to explain why the programme is "the most effective in the region" did three of those at once. Claude answered the question it was given: it built the case.

Think of a survey that asks "Don't you agree the council has neglected our parks?" The answers lean one way, and no respondent lied; the question did the work. Put precisely: the way you phrase a request shapes the answer.

2.2.5 A two-minute red-flag scan

You won't fact-check every sentence of every draft, and you don't need to. Problems cluster in predictable places, so a short scan with the draft and its source side by side tells you where to spend your attention.

RED-FLAG SCAN: two minutes, before a draft leaves your hands
1. Specifics. Mark every figure, percentage, date, name, section number and link. Can you say which document or search each one came from?
2. Quotes and citations. Did this person really say this, and do you have it in writing? Does the cited source say what the sentence claims?
3. Arithmetic. Do totals match their rows? Do the numbers in the summary match the numbers in the body?
4. Source match. Does every figure agree with the file you gave Claude?
5. Causes and superlatives. "Led to", "because of", "most", "first", "only", "proven": what evidence stands behind the link or the comparison?
6. Voices. Who is described, and who gets to speak or act? Whose view is missing?
7. Groups. Is any group described by a default picture rather than by your evidence?
8. Your own question. Did you ask Claude to confirm something you already believed?

Run it on the impact report and every problem lights up. The 73% fails items 1 and 5, the partner quote item 2, the two totals item 3 and the well count item 4. The portrayal of the communities fails items 6 and 7, and "most effective" fails items 5 and 8. None of that needed research, only attention in the right places.

A flag means "look here", not "this is false": the 48,000 might turn out to be right and the table wrong. Equally, a detail that passes the scan isn't thereby confirmed. The scan decides where your checking goes; it doesn't replace the checking.

2.2.6 Probing: questions that make problems show

Once a flag is up, the quickest next move is often a question back to Claude. A good probe doesn't ask for reassurance. It asks for something you can follow up: where a detail comes from, the same question asked neutrally, the strongest counter-argument, or a comparison of the draft with your files.

Here are the probes for the impact report. None of them asks "are you sure?", and the neutral question goes in a new chat, away from the framing of the first.

Which file and which row does the "73% drop in waterborne illness" come from? If it is not in the material I gave you, say so.
The summary says 48,000 people and the district table adds up to 41,600. Which one matches the monitoring spreadsheet, and where does the other come from?
List every figure, name and quote in the press release that you cannot find in the attached files.
(In a new chat) How does our programme compare with other clean-water programmes in the region? Give strengths, weaknesses and what you cannot judge from the material.
What is the strongest argument against calling our programme the most effective in the region?
Describe the programme from the point of view of the village water committees. What does the current draft leave out?

Asked where each figure comes from, Claude can show you which ones have a source; here it replies that the 73% isn't in the files and that 48,000 appears in last year's report. Asked neutrally, it lists strengths and gaps instead of reasons to agree. Anthropic's Academy suggests the same moves against sycophancy and one-sided answers: neutral, fact-seeking language, asking for counter-arguments or a balanced view, rephrasing or asking from another angle, and starting a new conversation.

But a probe's answer is still Claude's output. When a figure came from prediction, the source Claude names for it can come from the same process, so treat it as a lead, not a finding. "Are you sure?" is a weak probe: it can invite Claude to change its answer because you sound doubtful, not because anything was checked. Anthropic's hallucination guide is plain that such techniques reduce errors without eliminating them. Probing REVEALS; checking means opening the spreadsheet, last year's report and the partner's own words.

2.2.7 The exam traps

Every trap in this objective is a way of trusting the surface of an output instead of what stands behind it.

• ✗ Trusting a detail because it's precise or has a citation. ✓ Treat precision as a reason to check. Fabrication concentrates in exact figures, quotes and citations, and a citation shows where to look, not that the source agrees.
• ✗ Reading confidence or polish as a sign of accuracy. ✓ Judge a claim by its source. The tone is the same whether the fact is solid or invented.
• ✗ Asking "are you sure?" or for a confidence rating, then accepting the answer. ✓ Ask where the claim came from and check that source. A rating can suggest where to look first, but it is one more output, not evidence.
• ✗ Regenerating until a contradiction goes away. ✓ Find out which side the source supports. A clean rerun hides the question without answering it.
• ✗ Taking agreement with a leading question as confirmation. ✓ Re-ask neutrally in a fresh chat and ask for the strongest counter-argument.
• ✗ Fixing a biased frame by rewording it, or by throwing the draft away. ✓ Name the assumption, add the missing perspective or evidence, and revise. New words keep the slant; discarding the draft wastes useful work.

2.2.8 Put it together: scan and probe a draft of your own

You now have every piece: where invented details, contradictions and slants come from, a scan that flags them, and probes that make them visible. The quickest way to believe it is to watch a hallucination appear when you take the source away, as the 73% appeared where the charity's files had nothing to say.

Spotting is half the job. Fact-checking and validation (2.3) settles each flag against an authoritative source: opening the citation, recomputing the total, asking the partner for her actual words. Deciding when human review is required (2.4) scales that effort to the stakes, and a donor-facing press release quoting a named partner sits high on that scale. When a slanted frame could affect how people are treated, the ethical side of AI use (6.4) takes the question further.

Key takeaways

• ✓ Claude writes by predicting plausible text, so a specific detail that no document of yours and no search supports may be invented.
• ✓ Fabrication concentrates in figures, quotes, names, dates, citations and links, and confident, polished writing is not evidence.
• ✓ Inconsistencies show up inside an output, between the output and its source, and between runs; a contradiction proves something is wrong.
• ✓ Bias is a systematic slant, from one-sided framing and missing voices to stereotypes and inherited skew, and a leading question invites agreement rather than assessment.
• ✓ A two-minute red-flag scan shows where to look: specifics, quotes, arithmetic, source match, causes and superlatives, voices, groups and your own question.
• ✓ Probes such as "where is this from?", a neutral re-ask and the strongest counter-argument reveal problems; only a check against the source settles them.

Check your understanding

4 questions written for this lesson, then one from the CCAO-F question bank on the same topic. Every answer option is explained, including the ones you did not pick. Nothing is stored.

Sources

• Next Token Prediction (AI capabilities and limitations)
• Claude is providing incorrect or misleading responses
• Reduce hallucinations
• What is sycophancy in AI models?
• Why does bias exist in AI models?
• Use incognito chats
• The CCAO-F exam guide, summarised