You give the AI a document and ask for a summary. You expect the summary to contain only information from the document. Instead, the AI adds facts, statistics, and claims that were never in the source. The summary becomes a hybrid of your document and the AI's training data - and you cannot tell which is which.
Intrinsic hallucination. The model distorts information that is in the source. A number gets changed. A name gets swapped. A causal relationship gets reversed. The source says "revenue declined 3%" and the summary says "revenue declined 5%." The information is from the source, but wrong.
Extrinsic hallucination. The model adds information that is not in the source at all. The source document discusses a company's Q3 results. The summary adds a detail about the CEO's background that appears nowhere in the document. The model injected something from its training data into your summary.
Research shows that up to 30% of AI-generated summaries contain at least one hallucinated fact. For longer documents, the rate is higher - the model's attention degrades over long contexts, and it fills gaps with plausible-sounding completions rather than returning to the source.
The whole point of a summary is that you don't have to read the source. That's exactly what makes summarization hallucination so effective - the user has delegated reading to the AI and has no baseline to catch fabrications.
The user does not check. Summarization is a delegation task. You ask the AI to summarize because you do not want to read the full document. This means you are unlikely to verify the summary against the source. The entire point is to skip the source.
Summaries mix compression with generation. A good summary does not just extract sentences. It synthesises, rewords, and combines information. This means the model is generating new text, not copying source text - and generation is where hallucination happens.
Long documents exceed attention. For documents over a few thousand tokens, the model's attention to specific details weakens. It captures the general themes but may lose specific numbers, names, and details. When it generates the summary, it fills these gaps with probable values rather than accurate ones.
The training incentive. RLHF rewards complete, fluent summaries. A summary that says "the document mentions several financial metrics that I cannot accurately reproduce without re-reading" scores lower than one that confidently (incorrectly) states those metrics. The model learns to fill gaps rather than flag them.
Legal document review. Law firms use AI to summarise contracts, depositions, and case law. A hallucinated clause in a contract summary - or a fabricated precedent in a case summary - can misdirect legal strategy. Legal hallucination in summarization carries professional liability.
Medical record summarization. Healthcare providers use AI to summarise patient records, lab results, and clinical notes. A summary that adds a medication the patient is not taking, or changes a dosage, or fabricates a test result could inform treatment decisions based on false data.
Business intelligence. Executives rely on AI summaries of reports, earnings calls, and market research. A hallucinated statistic in a summary - one that was never in the source report - can drive incorrect business decisions. The executive trusts the summary. The summary is wrong.
Research synthesis. Researchers use AI to summarise papers. A summary that adds a finding not present in the paper, or attributes a conclusion to the wrong study, corrupts the research process. The hallucination enters the researcher's notes as a fact.
The fix is the same principle applied to the specific task. Every claim in the summary must be traceable back to the source document. Not "does this claim sound right?" but "does this specific statement appear in or follow from the source text?"
This requires post-generation validation - checking the summary output against the input document, claim by claim. It is the same grounding principle: the source document is the reality that the summary must be verified against.
Check applies this principle across all AI output: validate against the source of truth before the output reaches the user. For summarization, the source document is the truth. For system operations, the live environment is the truth. The pattern is always the same: verify against reality.
120 verifications a day free. No card, no signup.