AI Hallucinations: What They Are, Why They Happen, and How to Spot Them

An AI hallucination is not limited to an obviously absurd answer. The harder cases are plausible. A model can invent a paper that sounds real, attach a real link to a claim the source never made, use an outdated price, or produce a calculation whose final number looks reasonable enough to escape notice.

That is why hallucination detection is less about catching strange wording and more about verifying the underlying claims. The most useful question is not “Does this sound like something an expert would say?” but “What evidence would prove or disprove this?”

Why AI hallucinations happen

Language models generate likely continuations from the context they receive. They are not automatically performing a source check every time they produce a sentence. Google's guidance for web publishers makes the same practical point: generative models predict likely sequences of words, so their output can contain inaccuracies and should be fact-checked before publication. You can read that guidance in the Google Search documentation.

Hallucinations can have several causes at once. The model may lack the needed information, the question may be ambiguous, the context may be incomplete, or the system may infer a pattern that does not hold in this specific case. A retrieval tool can also find the wrong source or the right source can be interpreted too broadly.

There is also an incentive problem. OpenAI's 2025 research on why language models hallucinate argues that common training and evaluation setups can reward guessing instead of acknowledging uncertainty. In other words, a system can be pushed toward producing an answer even when the more reliable behavior would be to say it is unsure. The research is available from OpenAI.

Six common types of AI hallucination

1. Fabricated facts

The model states a date, event, feature, person, price, law, statistic, or technical detail that is simply not true. These are often easy to miss when the surrounding explanation is accurate.

2. Invented or distorted citations

A title, author, DOI, journal, quotation, or URL may be invented. Another version is more subtle: the source is real, but it does not support the claim attributed to it. See our separate guide on checking AI citations and sources.

3. Wrong numbers with plausible formatting

Percentages, totals, averages, unit conversions, growth rates, or financial calculations can be wrong even when the answer shows a neat formula. A visible calculation is not the same as an independently reproduced one.

4. Stale information presented as current

Product features, pricing, company roles, software versions, regulations, schedules, and market data can change quickly. An answer may have been true at one point and still be wrong for today's decision.

5. Scope errors

A source about one country, age group, product tier, or time period is generalized to everyone. This can be especially misleading because both the source and the quoted number may be real while the conclusion is not.

6. Unsupported conclusions

Every individual fact may be correct, but the final recommendation or causal claim does not follow from them. This is a reasoning failure rather than a simple factual error.

Warning signs worth checking first

There is no single wording pattern that proves an answer is hallucinated. Still, some situations deserve extra scrutiny because the cost of a hidden error is higher.

• Very specific numbers with no source or methodology.
• Quotations that are unusually perfect for the point being made.
• Academic references that cannot be found by title, author, or DOI.
• Claims about current pricing, policies, availability, or regulations without a current source.
• A confident answer to a question with missing or ambiguous premises.
• A conclusion that is much stronger than the evidence summarized immediately before it.

A practical workflow for detecting hallucinations

Step 1: Split the response into claims

Do not verify the paragraph as a whole. Mark the pieces that can be independently true or false: dates, names, statistics, quoted statements, product features, calculations, and causal claims.

Step 2: Mark what needs fresh information

Ask whether the fact could have changed recently. If it could, use a current source and check the source date. This matters for anything operational: prices, software, laws, schedules, executives, inventory, market figures, and service availability.

Step 3: Verify the source exists and supports the claim

A citation has two jobs. It must point to a real source, and that source must support the exact claim. Checking only the first job is not enough.

Step 4: Recalculate numbers

Extract the input values, confirm their units, write the formula separately, and reproduce the result. If the inputs came from external sources, verify those before trusting the arithmetic.

Step 5: Test the scope

Compare the claim with the population, geography, date, product version, and conditions in the supporting evidence. Many apparently factual mistakes are really scope mismatches.

Step 6: Inspect the reasoning separately

Once the facts are checked, ask whether the conclusion actually follows. Look for missing alternatives, unstated assumptions, causal claims based on correlation, and recommendations that depend on a preference the user never supplied.

Step 7: Preserve uncertainty

Sometimes the evidence is incomplete or reliable sources disagree. The correct result is not always a forced yes or no. A trustworthy workflow should be able to say that a claim remains uncertain.

Why asking the AI “Are you sure?” is not enough

Self-correction can be useful, but it is not independent verification. The same model can repeat the original error, invent a new justification, or become more confident after being challenged. The check becomes stronger when it introduces external evidence, deterministic calculation, or a separate evaluation of the reasoning.

Confidence language should therefore be treated as presentation, not proof. The evidence behind the answer matters more than whether the answer says “I'm highly confident.”

Match the verification effort to the stakes

You do not need a forensic investigation for every AI response. A low-stakes brainstorming idea can be used with light review. A client report, published article, financial model, legal interpretation, medical decision, or compliance statement deserves much stronger checking.

NIST's Generative AI Profile treats confabulation as one of the risks organizations should manage in generative AI systems. The practical lesson for individual users is similar: decide in advance how much error you can afford, then apply a verification process that matches that risk. The profile is available from NIST.

AI hallucination checklist

• Separate the answer into checkable claims.
• Flag claims that require current information.
• Verify that every important citation exists.
• Confirm that each source supports the exact claim.
• Recalculate important numbers independently.
• Check geography, date, population, and scope.
• Review the conclusion separately from the facts.
• Keep uncertainty when the evidence is incomplete.
• Escalate high-stakes decisions to qualified human review where appropriate.

How Verdictr helps with hallucination checks

Verdictr is built around the idea that different failure modes need different checks. A workflow can examine factual claims, current web evidence, user-supplied evidence, calculations, and reasoning rather than treating the whole response as one block of text.

If you already have an AI answer, VERIFY is the most direct path. If you are starting with a question, ANSWER can perform review work before presenting the result. For a more complete process, see How to Verify AI Answers.

Related guides