What you will be able to do
- Recognise a hallucination and trace it to a likely cause such as stale knowledge, thin coverage or ungrounded quoting
- Choose a detection technique (chain-of-thought, best-of-N, iterative refinement) that exposes a suspected hallucination
- Apply the documented fixes: permission to say "I don't know", quote-first grounding, cite-and-retract, and external knowledge restriction
Key concept
Hallucination — A hallucination is output that reads as plausible and confident but is factually wrong or contradicts the context Claude was given. To diagnose one, look at whether the claim is anchored in anything the model can actually see.
1.What a hallucination is and where it comes from
When an LLM application gives a bad answer, hallucination is usually the first suspect. Anthropic defines it as text that is "factually incorrect or inconsistent with the given context." Both halves of that definition matter when you diagnose. An answer can be wrong about the world, or it can be wrong about the document you supplied. In both cases the output tends to look correct: the support documentation warns that "Claude can write things that might look correct but are very mistaken." You can't judge reliability by how fluent or confident an answer sounds.
The causes are fairly predictable. A model knows only what it saw in training, and that knowledge stops at a cutoff, so questions about current events fall outside what it has learned. Coverage is also uneven. Topics that appeared often and consistently in training are well supported. Rare, niche, local, contested or post-cutoff topics sit at the edge, and that is where errors cluster. The model also can't say where a piece of knowledge came from. Anthropic's course calls this source amnesia. A related problem is fabricated attribution, where Claude can display quotes that look authoritative but are not grounded in fact.
Training adds another factor: loose calibration between how confident the model sounds and how reliable it actually is. In practice, a confident answer on a niche topic is no more trustworthy than a hedged one.
| Failure | What you observe |
|---|---|
| Staleness | Outdated information presented as current, e.g. on a recently changed regulation |
| Uneven coverage | Depth and accuracy drop on rare, niche or local topics, often with the same confident tone |
| Inherited bias | The model treats an outsider's view as the 'default' or 'normal' case |
| Source amnesia | The model cannot attribute where a piece of knowledge came from |
Checkpoint 1 of 4· Check yourself
A support bot answers questions about a regulation that changed after the model's training. It confidently describes the old rules. Which knowledge failure is this?
Knowledge is frozen at the training cutoff, so a rule that changed afterwards shows up as stale information presented as current.
“Claude might not have been trained on the most-up-to-date information and may get confused when prompted about current events.”Source: support.claude.com
2.Detecting a suspected hallucination
Since hallucinations look correct, you need ways to make them show themselves. Anthropic's guidance gives three, and each one exposes a different problem.
Chain-of-thought verification. Ask Claude to explain its reasoning step by step before it gives a final answer. When the reasoning is visible, you can see where a conclusion rests on an unsupported assumption.
Best-of-N verification. Run the same prompt several times and compare the outputs. A grounded fact should come back the same each time. A claim that changes between runs probably isn't anchored in anything.
Iterative refinement. Feed Claude's output back in a follow-up prompt and ask it to verify or expand on its earlier statements. This can catch inconsistencies and correct them.
There is a human step too. When Claude works from web search results, review the sources it cites. The original pages may contain context that the synthesis left out, and an answer can only be as good as the sources behind it. Anthropic's guidance is direct: users should not treat Claude as their only source of truth, and they should scrutinise high-stakes advice.
Checkpoint 2 of 4· Match them up
Match each detection technique to what it reveals
Tap a term, then the definition that fits it.
Each technique makes the problem visible in a different way: through the reasoning, through variation across runs, through self-verification, or through the original sources.
“Ask Claude to explain its reasoning step-by-step before giving a final answer. This can reveal faulty logic or assumptions.”Source: platform.claude.com
3.Fixing the cause: uncertainty, grounding and retraction
Once you have confirmed a hallucination, choose the fix that matches its cause.
The model guessed when it lacked information. Explicitly give Claude permission to admit uncertainty. Anthropic calls this a simple technique that can drastically reduce false information. The prompt should name an exact fallback phrase:
Focus on financial projections, integration risks, and regulatory hurdles. If you're unsure about any aspect or if the report lacks necessary information, say "I don't have enough information to confidently assess this."The answer drifted away from a long document. For long documents (over 20k tokens), have Claude extract word-for-word quotes before it does the task, and base its analysis only on those quotes. That ties the answer to text that actually exists.
Claims can't be audited. Have Claude cite a supporting quote for each claim. You can also add a pass after generation where Claude looks for a supporting quote for every claim and retracts any claim it can't support.
General knowledge leaked into a document-bound task. Tell Claude to use only the information in the provided documents and not its general knowledge.
The knowledge simply isn't in the model. If the topic is after the cutoff or too niche, prompting won't fix it. Web search, retrieval (RAG/MCPs) and tool use exist to give the model information it was never trained on.
Checkpoint 3 of 4· Put it in order
Put the cite-and-retract workflow for a document-bound press release in order
- 1.For each claim, find a direct quote from the documents that supports it
- 2.Draft the press release using only information from the supplied documents
- 3.Review each claim in the draft
- 4.Remove any claim without a supporting quote and mark where it was removed with []
Verification runs after drafting. Every claim needs a supporting quote, and a claim without one is removed with a visible marker.
“For each claim, find a direct quote from the documents that supports it.”Source: platform.claude.com
Checkpoint 4 of 4· Exam question
A retrieval-augmented HR assistant answers employee questions using uploaded policy PDFs. Reviewers find it confidently states a specific vacation-accrual percentage that does not appear anywhere in the source documents. Which change to the assistant's instructions best addresses the root cause of this failure?
Correct answer: A — Require the assistant to quote exact policy text before answering, admitting uncertainty when none exists.
- A. Correct. Forcing the assistant to ground claims in extracted quotes and explicitly permitting an admission of uncertainty directly targets the mechanism behind the fabrication: the model filled a gap with a plausible-sounding invented figure instead of acknowledging the documents didn't cover it.
- B. Incorrect. Explicitly encouraging the model to fall back on general training knowledge is the opposite of the fix — it legitimizes exactly the behavior that produced the fabricated percentage.
- C. Incorrect. Forcing a numeric answer in every case removes the model's ability to say it doesn't know, making fabrication more likely, not less.
- D. Incorrect. Restating the full document doesn't stop the model from inventing details for questions the document never addresses, and it wastes context on unrelated sections.
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.A response that includes quotes or citations must be grounded, so it doesn't need verifying.Why is that wrong?
Claude can produce quotes that look authoritative but are not grounded in fact. Citations only help when each one is checked against the source and unsupported claims are retracted.
2.Extracting quotes first is the standard fix for every hallucination, whatever the task.Why is that wrong?
The guidance targets quote-first grounding at long documents over 20k tokens. A model guessing without information needs permission to say "I don't know", and missing knowledge needs retrieval or tools.
Covered in Fixing the cause: uncertainty, grounding and retraction
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.https://support.claude.com/en/articles/8525154-claude-is-providing-incorrect-or-misleading-responses-what-s-going-onOfficial docs
“Claude can write things that might look correct but are very mistaken.”
↩︎ What a hallucination is and where it comes from“Users should not rely on Claude as a singular source of truth”
↩︎ Detecting a suspected hallucination“Claude can display quotes that may look authoritative or sound convincing, but are not grounded in fact.”
↩︎ Exam trap 1“Claude might not have been trained on the most-up-to-date information and may get confused when prompted about current events.”
↩︎ Checkpoint - 2.
“Identify staleness, uneven coverage, inherited bias, and source amnesia as characteristic knowledge failures”
↩︎ What a hallucination is and where it comes from“Web search, retrieval (RAG/MCPs), and tool use exist specifically to patch these gaps”
↩︎ Fixing the cause: uncertainty, grounding and retraction - 3.https://academy.claude.com/courses/ai-capabilities-and-limitations/how-ai-gets-its-characterOfficial docs
“loose calibration between stated confidence and actual reliability”
↩︎ What a hallucination is and where it comes from - 4.https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinationsOfficial docs
“Run Claude through the same prompt multiple times and compare the outputs.”
↩︎ Detecting a suspected hallucination“Use Claude's outputs as inputs for follow-up prompts, asking it to verify or expand on previous statements.”
↩︎ Detecting a suspected hallucination“Explicitly give Claude permission to admit uncertainty. This simple technique can drastically reduce false information.”
↩︎ Fixing the cause: uncertainty, grounding and retraction“If it can't find a quote, it must retract the claim.”
↩︎ Fixing the cause: uncertainty, grounding and retraction“Explicitly instruct Claude to only use information from provided documents and not its general knowledge.”
↩︎ Fixing the cause: uncertainty, grounding and retraction“Even the most advanced language models, like Claude, can sometimes generate text that is factually incorrect or inconsistent with the given context.”
↩︎ Key concept“For tasks involving long documents (>20k tokens), ask Claude to extract word-for-word quotes first before performing its task.”
↩︎ Exam trap 2“Inconsistencies across outputs could indicate hallucinations.”
↩︎ Prediction“Ask Claude to explain its reasoning step-by-step before giving a final answer. This can reveal faulty logic or assumptions.”
↩︎ Checkpoint“For each claim, find a direct quote from the documents that supports it.”
↩︎ Checkpoint