What you will be able to do
- Recognise a response that conflicts with the supplied context or with itself
- Read different answers to the same question across fresh chats as a warning sign
- Identify inherited default-assumption bias, sycophancy and uneven hedging
- Apply the three responses the help centre actually recommends
1.When a response contradicts its own context
Anthropic's developer documentation defines hallucination broadly enough to include inconsistency. Claude can generate text that is factually incorrect or inconsistent with the given context. That second half matters when you review work. A response can be fluent and still disagree with the brief, with a document you supplied, or with something it said earlier in the same answer.
The AI Capabilities and Limitations course describes a related pattern in instruction-following. A model can follow a detailed instruction perfectly and then ignore a simple one in the very next message. So a reviewer checks two things. Does each claim fit the source material? Do the parts of the response agree with each other? A summary that gives one total in the opening line and a different one in a later table has an internal inconsistency, whichever figure turns out to be right.
These sources do not explain how Claude picks between two conflicting documents in a project's knowledge. For example, they do not say why it might repeat last year's policy when this year's is also present. What they do support is spotting the result: an answer that is inconsistent with the context it was given. Treat that as a finding to act on, not as something to explain away.
Step-by-step reasoning exposes the logic behind the final answer. If a step rests on an assumption the context does not support, you can see it. This is developer prompting guidance, so use it as good practice and not as something the help centre recommends.
2.Different answers to the same question
A second kind of inconsistency only shows up when you ask again. Because the model generates text by sampling, the same request in a fresh conversation can come back different. The AI Fluency course has you test this directly: run the exact same request for specific facts in a new conversation and compare the outputs. Some details stay the same and some change.
The developer documentation turns this into a signal. Its Best-of-N technique runs Claude through the same prompt several times and compares the outputs, and it states that inconsistencies across outputs could indicate hallucinations. So if two chats give two different dates for a deadline that appears in no document you supplied, you should not pick the one that sounds better. The disagreement tells you that neither answer rests on a grounded fact, and the value has to come from a source outside the chat.
A user asks Claude for a detailed, confident summary of a major news event that occurred after the model's reliable knowledge cutoff. Claude produces specific dates, names, and outcomes that cannot be verified in any source. What is the most likely underlying cause of this response?
Correct answer: C — Claude hallucinated plausible-sounding details to fill a gap past its knowledge cutoff
- A. Incorrect. The scenario states the summary addressed the event in question, so a topic mismatch does not explain the fabricated details.
- B. Incorrect. Systematic favoring of certain groups or viewpoints is bias; inventing unverifiable specifics about a recent event is a knowledge-gap hallucination.
- C. Correct. When a query falls beyond the model's reliable knowledge cutoff and no external grounding is provided, the model can generate confident but fabricated specifics instead of acknowledging the gap.
- D. Incorrect. No tool use or search was described in the scenario, so there are no cached results to attribute the fabricated details to.
3.Bias: defaults, agreement and uneven hedging
Bias in a response is usually quieter than offensive content. The Knowledge module lists inherited bias among the model's characteristic failures, meaning bias in what counts as the default or normal case. It suggests a probe: without naming your assumption, ask a question that would reveal whether the AI takes the outsider's view of your field, for example by asking it to describe a typical customer. Then note what it treats as normal.
Training leaves other fingerprints as well. The course lists sycophancy, verbosity, over-caution and loose confidence calibration. Sycophancy leans a response toward what the asker seems to want. Loose calibration means certainty and hedging don't track the evidence. Together they explain a pattern that is easy to miss when you ask for balanced content: one side is stated plainly while the other is wrapped in qualifiers. No single sentence is false, but the response as a whole is not balanced. Classify it as bias and correct it, rather than approving it because each fact checks out.
4.What the help centre tells you to do
Spotting a problem only helps if you respond to it. The help centre article on incorrect or misleading responses recommends three things, and only three. Keep them apart from the prompting techniques in the developer guides, such as permission to say you don't know, quoting documents first, or showing reasoning. Those are sound practice, but this article does not state them.
| Action | When it applies | What it achieves |
|---|---|---|
| Do not rely on Claude as a singular source of truth | Always, and especially for high-stakes advice | Keeps independent checking in place |
| Review Claude's cited sources | When working with web search results | Shows context or details the synthesis left out or misread |
| Use the thumbs down button | When a response was unhelpful | Tells Anthropic about the problem response |
Reviewing sources works because, as the article notes, the quality of Claude's answer depends on the sources it draws on, and checking the original content reveals anything misread out of context. The AI Fluency course suggests the same test with citations: re-run a request for specific facts in a tool with citations turned on, such as Research mode in Claude, and see whether having sources to check changes your score. The course links all of this to Discernment. Knowing an output was generated tells you what kind of scrutiny to apply.
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.When two chats give different answers to the same factual question, go with the more detailed or more confident one.Why is that wrong?
If repeated runs disagree, the disagreement itself suggests hallucination. Neither answer should be trusted until it is checked against an outside source.
Covered in Different answers to the same question
2.A response is only biased if it contains offensive or openly one-sided statements.Why is that wrong?
Bias also shows up in which case the model treats as default or normal, and in uneven hedging, even when every individual statement is accurate.
Covered in Bias: defaults, agreement and uneven hedging
3.The help centre's fix for hallucination is to give Claude permission to say it does not know.Why is that wrong?
That technique comes from developer prompting guidance. The help centre article recommends the thumbs-down button, reviewing cited sources, and not relying on Claude as a singular source of truth.
Covered in What the help centre tells you to do
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinationsOfficial docs
“can sometimes generate text that is factually incorrect or inconsistent with the given context.”
↩︎ When a response contradicts its own context“Ask Claude to explain its reasoning step-by-step before giving a final answer. This can reveal faulty logic or assumptions.”
↩︎ When a response contradicts its own context“Best-of-N verification: Run Claude through the same prompt multiple times and compare the outputs.”
↩︎ Different answers to the same question“Inconsistencies across outputs could indicate hallucinations.”
↩︎ Exam trap 1 - 2.
“It follows a detailed instruction perfectly, then ignores a simple one in the very next message.”
↩︎ When a response contradicts its own context“recognize the behavioral fingerprints it leaves: sycophancy, verbosity, over-caution, and loose confidence calibration”
↩︎ Bias: defaults, agreement and uneven hedging - 3.https://academy.claude.com/courses/ai-capabilities-and-limitations/next-token-predictionOfficial docs
“Run the exact same specific-facts request in a fresh conversation. Compare the two outputs.”
↩︎ Different answers to the same question“Re-run Probe 2 in a tool with citations enabled (like Research mode in Claude).”
↩︎ What the help centre tells you to do“Knowing the output was generated tells you exactly what kind of scrutiny to apply.”
↩︎ What the help centre tells you to do - 4.
“Note what it treats as normal.”
↩︎ Bias: defaults, agreement and uneven hedging“Characteristic failures: staleness, uneven coverage, inherited bias in what counts as”
↩︎ Exam trap 2 - 5.https://support.claude.com/en/articles/8525154-claude-is-providing-incorrect-or-misleading-responses-what-s-going-onOfficial docs
“checking original content helps you identify any information that might be misinterpreted without the full context.”
↩︎ What the help centre tells you to do“You can use the thumbs down button to let us know if a particular response was unhelpful”
↩︎ What the help centre tells you to do“Users should not rely on Claude as a singular source of truth and should carefully scrutinize any high-stakes advice given by Claude.”
↩︎ Exam trap 3