CertSafari
    Anthropic Claude Certified Associate – Foundations (CCAO-F) - December 7 onwards· Lessons

    Domain 2 · Lesson 5/21

    Hallucination detection and verification for Claude outputs

    Hallucination Detection and Verification

    17 min read
    5% of exam
    6 sources
    Published 18 Nov 2026
    Docs as of 28 Oct 2026

    What you will be able to do

    • Explain why fluent, confident Claude output can still contain fabricated content, and where that fabrication concentrates
    • Recognise four signals of unsupported content: fabricated citations and links, confident invented specifics, inconsistencies, and bias
    • Explain why asking Claude to double-check its own answer does not confirm a fact
    • Verify a draft's statistics and named facts against sources, in the right order, before the draft is used
    • Handle a citation that does not resolve: confirm it is fabricated, remove the claim, and source a checked replacement

    Key concept

    Hallucination — Text that Claude generates that reads as correct but is factually wrong or unsupported by the context it was given. It comes from the same word-by-word generation that makes Claude fluent, so a confident tone tells you nothing about whether a claim is true.

    1.Why fluent text can be fabricated

    Claude writes an answer one word at a time, each word chosen from what tends to follow the words before it. It is not looking anything up. That one mechanism explains both why the prose reads so smoothly and why some of it can be invented. Summarising, reformatting and explaining common concepts follow patterns the model has seen many times, so they usually come out right. The risk lies in tasks that require telling what is true apart from what merely sounds true, because the generation process has no built-in way to make that distinction.

    Fabrication concentrates in specifics: names, dates, statistics, citations, URLs and direct quotes. The working rule is that the more precise a claim, the more it needs verifying. A sentence saying a market is growing is low risk. A sentence giving that market's growth as an exact percentage, credited to a named report, is exactly where an invented figure will appear.

    How confident Claude sounds is not evidence either way. Training on human judgments of good responses leaves a loose calibration between how confident a model sounds and how reliable it actually is. An invented statistic is written in the same assured tone as a correct one. A useful test in your own field is to compare a mainstream topic with a niche one and see whether the answers signal uncertainty differently or sound equally confident.

    Two further gaps produce confident errors. The first is knowledge cutoff. Without tools such as web search, Claude has no information from after its training cutoff. If you ask for a competitor's current pricing with search turned off, you can get a tidy table of figures that were stale, or invented, when it was generated. The second is invented capabilities. Claude can claim to have sent an email or produced an external document even though it has no access to tools that have not been explicitly connected.

    Checkpoint 1 of 9· Check yourself

    Web search is turned off. You ask Claude for a competitor's current pricing, and it returns a neat table of exact prices with no hedging at all. How should you treat the table?

    Sources1234

    2.Four signals of unsupported content

    Unsupported content shows up in four recognisable ways. Each one needs a different kind of check, so name the signal before deciding how to check it.

    Signals of unsupported content and what each one looks like in practice
    SignalWhat it looks likeWhy it happens
    Fabricated citations and linksA reference to a real-sounding report, or a full URL with a detailed path on a real domain, that leads nowhereCitations and URLs are precise specifics, the zone where fabrication concentrates
    Confident invented specificsExact figures, dates, job titles or quotes stated without hedgingStated confidence is only loosely tied to actual reliability
    InconsistenciesClaims that contradict the provided document, each other, or a fresh run of the same promptEach run samples its words afresh, so unsupported details can change between runs
    BiasOutsider default assumptions treated as normal, or agreement with the user's framingInherited bias from training data, plus a trained pull toward sycophancy

    Inconsistency is the signal you can check most cheaply. A hallucination can contradict the context you supplied, such as a contract summary that names a different notice period from the contract. It can also show up between runs. Run the same request for specifics in a fresh conversation and compare the two. Details that change between runs are the ones the model is generating rather than recalling, and inconsistencies across outputs can point to hallucinations.

    Bias is harder to see because it looks like reasonable defaults. Claude's knowledge carries inherited bias in what counts as "default" or "normal". If you ask it to describe a typical customer in your field, it may give the outsider's picture. Sycophancy works in a similar way. If a prompt opens with "I think this strategy is bulletproof," the answer may endorse that framing instead of pushing back. To check for bias, compare the output with what people inside the field know, and invite disagreement explicitly.

    Checkpoint 2 of 9· Match them up

    A reviewer marks four problems in a Claude-drafted report. Match each problem to the signal of unsupported content it shows.

    Tap a term, then the definition that fits it.

    Sources532

    3.Why asking Claude to double-check is not verification

    Typing "please double-check your answer" in the same chat asks the same model, working from the same text, to judge itself. That has some value. Feeding Claude's output back to it and asking it to verify or expand on earlier statements can catch and correct inconsistencies within the text. But a model that generated a wrong date with confidence can confirm the same wrong date with confidence. Nothing in that loop brings in information from outside, which is why Claude should not be treated as the only source of truth.

    Three tempting explanations for why self-review falls short are wrong. The first is that Claude cannot read earlier messages in the conversation; it can, and iterative refinement depends on it. The second is that a second request just returns the same text; Claude does revise answers when asked. The third is that self-review only rechecks numbers; names, dates and quotes can all be re-examined. The real limit is that confirmation needs a source outside the generated text: the original document, an authoritative publication, or a search result you open and read yourself.

    The same applies to best-of-N comparisons. If three runs agree, that makes a fabrication less likely, but agreement is still a signal, not a source. Use disagreement between runs to decide which claims to check first, then check them against something Claude did not write. How much of this checking an output needs depends on its stakes, which is covered under review depth and escalation.

    Checkpoint 3 of 9· Check yourself

    Your Claude-drafted market note gives a competitor's product launch date. You type "please double-check that date" in the same chat, and Claude replies that the date is correct. What would actually confirm the date before the note goes out?

    Checkpoint 4 of 9· Exam question

    Why is typing "Please double-check your answer" in the same conversation not enough to verify a factual claim Claude has made?

    Sources56

    4.Verifying a draft before it is used

    Take a Claude-drafted market brief with eight statistics that is going to a client and has not been checked yet. Verification follows a fixed order. Each step needs the one before it, and the last step is there because the fixes themselves can add new errors.

    Work through the brief in this order:

    1. List every statistic and named fact the client could rely on or challenge. This is the inventory you will check against.
    2. Find a primary or authoritative source for each item, either the source the draft cites or one from your own independent search.
    3. Compare each figure and its wording against its source. Mark mismatches, and mark any item for which you found no source.
    4. Correct mismatched items using the source. Delete or flag any item you could not support.
    5. Re-read the revised brief for new inconsistencies introduced by the edits, then release it.

    Finding one error changes how you treat the rest. Suppose a spot-check finds a wrong launch date in a ten-claim briefing. Fixing that date and sending the other nine is the wrong move. The other nine sit in the same high-risk zone of precise specifics and were written in the same confident tone, so the error is evidence that every specific in the briefing needs checking.

    Some checking can go into the prompt itself. You can have Claude cite a quote and source for each claim so the response can be audited. You can also have it go back over its own draft and find a supporting quote for every claim, removing any claim it cannot support. This makes your external check faster, because each claim arrives tied to a passage you can open. It does not replace that check.

    A prompt that makes Claude find a supporting quote for each claim and visibly remove any claim it cannot supporttext
    Draft a press release for our new cybersecurity product, AcmeSecurity Pro, using only information from these product briefs and market reports.
    <documents>
    {{DOCUMENTS}}
    </documents>
    
    After drafting, review each claim in your press release. For each claim, find a direct quote from the documents that supports it. If you can't find a supporting quote for a claim, remove that claim from the press release and mark where it was removed with empty [] brackets.

    Two more prompt instructions reduce how much there is to catch. Explicitly allowing Claude to say "I don't have enough information" can drastically reduce false information, because the model no longer feels pushed to fill the gap with something plausible. Telling it to use only the documents provided, and not its general knowledge, means every claim should trace back to text you can check. For long documents, asking Claude to pull out word-for-word quotes before it does the task grounds the analysis in the actual text.

    Checkpoint 5 of 9· Fill the gap

    This prompt makes every claim in the press release traceable to text you can check. Which word completes it?

    Draft a press release for our new cybersecurity product, AcmeSecurity Pro, using  ?  information from these product briefs and market reports.
    <documents>
    {{DOCUMENTS}}
    </documents>
    
    After drafting, review each claim in your press release. For each claim, find a direct quote from the documents that supports it. If you can't find a supporting quote for a claim, remove that claim from the press release and mark where it was removed with empty [] brackets.

    Checkpoint 6 of 9· Check yourself

    Halfway through reviewing a ten-claim client briefing, you find that one product launch date is wrong. The briefing is due out today. What is the right next step?

    Checkpoint 7 of 9· Exam question

    A strategy analyst has a Claude-drafted market brief containing eight statistics. The brief will go to a client, and nothing in it has been checked yet. Which sequence should the analyst follow to verify the brief?

    Sources5

    5.When a citation or link does not resolve

    A link that returns "page not found" is not proof that the work is fake. Pages move. But a link that does not resolve is the typical form of a fabricated citation: a plausible title, a real domain, and a detailed path that nobody ever published. Claude can produce links that do not work, and a realistic domain shows only that the model knows what that vendor's URLs tend to look like. Before deciding, check whether the work itself exists.

    For an unshared report with one reference you cannot reach at its cited link, use this order:

    1. Search the publisher's site or a library database for the exact title and authors, to find out whether the work exists anywhere.
    2. If there is no record of it, mark the citation as fabricated and remove the claim it was supporting. That claim has nothing behind it.
    3. Ask Claude to propose a replacement for the claim, and give it source documents you have already confirmed, with an instruction to use only those.
    4. Open the replacement source and confirm that it exists and actually states the claim before you add it to the report.

    The replacement has to meet the same standard as the original. If you ask Claude for a new source from its general knowledge, it can fabricate a second citation as easily as the first. Restricting it to confirmed documents and then opening the result yourself closes that gap. Even with real sources, read the original. Original pages often contain context that the synthesis leaves out, and checking them helps you spot information that was misread without that context.

    Checkpoint 8 of 9· Check yourself

    One reference in an unshared report gives a detailed URL on a real publisher's domain, and the link returns "page not found". You search the publisher's site and a library database for the exact title and authors and find no record of the work. What should you do next?

    Checkpoint 9 of 9· Exam question

    An analyst finds that one reference in a Claude-drafted report cannot be found at its cited link. The report has not been shared. Any replacement source must pass the same checks as the original. Which sequence should the analyst follow?

    Sources465

    Practise it for real

    Measure how often Claude fabricates specifics in a field you know well, and see how run-to-run variation exposes them

    1. 1.Pick a topic where you are the expert. In a Claude chat, ask for five checkable specifics: three cited sources, an author's name and an exact figure.

      Why: Specifics are where fabrication concentrates, and your own expertise lets you check them.

      You should see: A fluent answer with five precise-looking items.

    2. 2.Check every item against a source you open yourself, and score how many were fully accurate out of five.

      Why: Confirmation has to come from outside the generated text.

      You should see: A score, plus a note of how confident any wrong items sounded.

    3. 3.Run the exact same request in a fresh conversation and compare it with the first output.

      Why: Details that change between runs are being generated, not recalled.

      You should see: Some items stay the same and others change. The ones that changed are your strongest suspects.

    4. 4.Re-run the request in a tool with citations enabled, such as Research mode, and score it again.

      Why: Cited sources give you something concrete to check each claim against.

      You should see: Claims tied to sources you can open, and possibly a different score.

    Stuck? Get a nudge

    Ask yourself whether you would have caught these fabrications in a field you did not know well.

    Sources

    (6)

    Every claim above is drawn from one of these pages, quoted as it was written on the date shown.

    1. 1.
      “It writes answers word by word based on what tends to follow what. That single property gives you both the fluency and the hallucination.”
      ↩︎ Why fluent text can be fabricated
      “The more precise a claim, the more it warrants verification.”
      ↩︎ Why fluent text can be fabricated
      “The more precise a claim, the more it warrants verification.”
      ↩︎ Exam trap 3
      “Fabrication concentrates in specificity: names, dates, statistics, citations, URLs, quotes.”
      ↩︎ Prediction
      “Fabrication concentrates in specificity: names, dates, statistics, citations, URLs, quotes. The more precise a claim, the more it warrants verification.”
      ↩︎ Checkpoint
    2. 2.
      “loose calibration between stated confidence and actual reliability”
      ↩︎ Why fluent text can be fabricated
      “See whether the AI validates your framing or pushes back.”
      ↩︎ Four signals of unsupported content
      “loose calibration between stated confidence and actual reliability”
      ↩︎ Exam trap 1
      “Fine-tuning uses human judgments about good responses, and those judgments leave fingerprints: a pull toward sycophancy”
      ↩︎ Checkpoint
    3. 3.
      “Pay attention to whether the AI signals uncertainty differently between the two, or whether both answers come with the same confident tone.”
      ↩︎ Why fluent text can be fabricated
      “Without tools, it has no access to any information after that date.”
      ↩︎ Why fluent text can be fabricated
      “inherited bias in what counts as”
      ↩︎ Four signals of unsupported content
    4. 4.
      “Claude does not have access to other tools or software that are not explicitly integrated, including email, word processors, or file transfers.”
      ↩︎ Why fluent text can be fabricated
      “In an attempt to be a helpful assistant, Claude can sometimes hallucinate its capabilities.”
      ↩︎ When a citation or link does not resolve
      “In an attempt to be a helpful assistant, Claude can sometimes hallucinate its capabilities.”
      ↩︎ Exam trap 4
    5. 5.
      “generate text that is factually incorrect or inconsistent with the given context”
      ↩︎ Four signals of unsupported content
      “Inconsistencies across outputs could indicate hallucinations.”
      ↩︎ Four signals of unsupported content
      “asking it to verify or expand on previous statements. This can catch and correct inconsistencies.”
      ↩︎ Why asking Claude to double-check is not verification
      “having it cite quotes and sources for each of its claims”
      ↩︎ Verifying a draft before it is used
      “Explicitly give Claude permission to admit uncertainty. This technique can drastically reduce false information.”
      ↩︎ Verifying a draft before it is used
      “Explicitly instruct Claude to only use information from provided documents and not its general knowledge.”
      ↩︎ Verifying a draft before it is used
      “Explicitly instruct Claude to only use information from provided documents and not its general knowledge.”
      ↩︎ When a citation or link does not resolve
    6. 6.
      “Users should not rely on Claude as a singular source of truth and should carefully scrutinize any high-stakes advice given by Claude.”
      ↩︎ Why asking Claude to double-check is not verification
      “checking original content helps you identify any information that might be misinterpreted without the full context.”
      ↩︎ When a citation or link does not resolve
      “Claude can display quotes that may look authoritative or sound convincing, but are not grounded in fact.”
      ↩︎ Key concept
      “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 2

    Ready to test yourself?

    Practise Anthropic Claude Certified Associate – Foundations (CCAO-F) - December 7 onwards in quiz mode.

    Spotted a mistake, or was something unclear? Tell us.