What you will be able to do
- Explain what a hallucination is and why it is a limitation of current models rather than a setup mistake
- Connect hallucination to next-token prediction and to a knowledge base that is frozen and uneven
- Predict which parts of a response are most likely to be fabricated
- Keep a confident tone from being taken as evidence of accuracy
Key concept
Hallucination — A response that reads as correct but is false or unsupported. It comes from the way generative models produce text, so every output carries some risk of it, and no setting or instruction switches it off.
1.A known limitation, not a misconfiguration
Anthropic's help centre says it plainly: Claude can occasionally produce responses that are incorrect or misleading, and this is called hallucinating. The same article calls it a byproduct of the current limitations of frontier generative AI models. That framing is the first thing to take into the exam. A hallucination is not a sign that someone set up a project badly, forgot a setting, or wrote a weak system prompt. It is a property of the technology, so every output is exposed to it.
The developer documentation defines it the same way from the other side. Even the most advanced models can generate text that is factually incorrect, or that is inconsistent with the context they were given. That gives you two things to watch for. One is a claim about the world that is simply wrong. The other is a claim that contradicts the material you supplied. Both count, and a response can look polished while doing either.
It cannot remove hallucination, because hallucination is not a behaviour Claude chooses to do or not do. It comes from how the model generates text. Instructions that let Claude admit uncertainty are sound prompting practice, but they only reduce the risk. They are not a reason to stop reviewing outputs.
2.Why fluent text can be false
The AI Fluency course explains the mechanism. Generative AI is closer to a very sophisticated autocomplete than to a search engine. It writes answers word by word, based on what tends to follow what. The course draws the consequence: one property gives you both the fluency and the hallucination. The model is not looking a fact up and occasionally misreading it. It is producing the most plausible continuation, and plausible is not the same as true.
The second cause is knowledge. A model knows only what it saw in training, and that knowledge is frozen at a cutoff date. Coverage is also uneven. Topics that appeared often, recently and consistently in training data sit in the capability zone. Rare, post-cutoff, niche, local or contested topics sit in the limitation zone. The help centre gives a concrete case: Claude might not have the most up-to-date information and may get confused when asked about current events. A question about something that changed after the cutoff can produce a confident answer that describes how things used to be.
3.Where fabrication concentrates
Hallucination is not spread evenly across a response. The AI Fluency course names where it gathers: fabrication concentrates in specificity. That means names, dates, statistics, citations, URLs and quotes. Its rule of thumb is that the more precise a claim, the more it warrants verification. A general explanation of a common concept is the kind of well-worn path the model handles well. An exact figure, a named author or a page link is where an invented detail most easily hides.
| Situation | Zone | How much scrutiny |
|---|---|---|
| Summarising, reformatting or explaining a common concept | Capability zone | Spot-check the content |
| Mainstream, well-documented, stable topics | Capability zone | Lower, but still check anything specific |
| Rare, post-cutoff, niche, local or contested topics | Limitation zone | High: bring your own sources |
| Names, dates, statistics, citations, URLs, quotes | Where fabrication concentrates | Verify each one |
A support engineer asks Claude to summarize a technical paper and include supporting references. The response includes a citation to a journal article whose title and authors do not appear in the source document or in any database the engineer can find. What does this response demonstrate?
Correct answer: D — A hallucination, since Claude fabricated a citation with no basis in any verifiable source
- A. Incorrect. Bias involves skewed treatment of topics, groups, or viewpoints, not the invention of a nonexistent source.
- B. Incorrect. Inconsistency refers to conflicting statements within or across responses, not a fabricated fact that never existed.
- C. Incorrect. Nothing in the scenario indicates the document exceeded the context window; the issue is a fabricated reference, not truncation.
- D. Correct. A hallucination is content presented as fact that is not grounded in the source material or any verifiable reference; an invented citation is a classic example.
Sources3
4.Confidence is not evidence
The most dangerous thing about a hallucination is that it rarely looks like one. The help centre warns that Claude can display quotes that look authoritative or sound convincing but are not grounded in fact. The AI Capabilities and Limitations course describes the same thing: training leaves behavioural fingerprints, and one of them is loose confidence calibration. Put simply, how sure the model sounds is a weak guide to how likely it is to be right.
The course turns this into a probe you can run. Ask about one mainstream topic and one niche topic, then compare them. Watch whether the AI signals uncertainty differently for the two, or whether both come back in the same confident tone. When a well-covered answer and a thinly covered answer sound equally sure, tone has told you nothing, and the niche answer needs checking.
In a single multi-turn conversation, Claude tells a user early on that a product ships in 3 business days, then later in the same conversation states the same product ships in 7 business days without any new information being provided. How should a QA reviewer classify this behavior?
Correct answer: B — An inconsistency, since the two statements about the same fact conflict within one chat
- A. Incorrect. A shipping estimate is plausible and was actually stated twice; the problem is that the two values contradict each other, not that one is fabricated from nothing.
- B. Correct. An inconsistency is when a model gives conflicting answers to the same question without any change in input, which is exactly what happened here.
- C. Incorrect. Bias refers to skewed treatment of a topic or group, not a reviewer question about which of two conflicting numeric answers is favored.
- D. Incorrect. Response timing does not cause a model to contradict earlier output; the two turns are evaluated on content, not on elapsed time.
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.Telling Claude never to say anything untrue removes hallucination, so outputs no longer need checking.Why is that wrong?
Hallucination comes from the limits of current frontier models, not from missing instructions. Prompting can reduce it but cannot remove it, and outputs still need review.
Covered in A known limitation, not a misconfiguration
2.A response that is mostly accurate can be trusted in its specific details as well.Why is that wrong?
Fabrication clusters in precise details such as citations, URLs, statistics and quotes. A sound general explanation says nothing about whether those details are real.
Covered in Where fabrication concentrates
3.If Claude states something confidently and without hedging, it is probably correct.Why is that wrong?
Invented quotes and claims can sound just as convincing as true ones. How confident Claude sounds is not a reliable signal of accuracy.
Covered in Confidence is not evidence
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 occasionally produce responses that are incorrect or misleading.”
↩︎ A known limitation, not a misconfiguration“Claude might not have been trained on the most-up-to-date information and may get confused when prompted about current events.”
↩︎ Why fluent text can be false“a byproduct of some of the current limitations of frontier Generative AI models”
↩︎ Key concept“a byproduct of some of the current limitations of frontier Generative AI models”
↩︎ Exam trap 1“Claude can display quotes that may look authoritative or sound convincing, but are not grounded in fact.”
↩︎ Exam trap 3 - 2.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.”
↩︎ A known limitation, not a misconfiguration“Explicitly give Claude permission to admit uncertainty.”
↩︎ A known limitation, not a misconfiguration - 3.https://academy.claude.com/courses/ai-capabilities-and-limitations/next-token-predictionOfficial docs
“That single property gives you both the fluency and the hallucination.”
↩︎ Why fluent text can be false“The more precise a claim, the more it warrants verification.”
↩︎ Where fabrication concentrates“Fabrication concentrates in specificity: names, dates, statistics, citations, URLs, quotes.”
↩︎ Exam trap 2 - 4.
“Limitation zone: rare, post-cutoff, niche, local, or contested topics.”
↩︎ Why fluent text can be false“Pay attention to whether the AI signals uncertainty differently between the two, or whether both answers come with the same confident tone.”
↩︎ Confidence is not evidence - 5.
“recognize the behavioral fingerprints it leaves: sycophancy, verbosity, over-caution, and loose confidence calibration”
↩︎ Confidence is not evidence