AI Hallucination Examples

Multiple documented cases exist where attorneys used AI-generated legal research containing fabricated case citations that were later filed in court. Courts in several jurisdictions have since issued guidance on AI use in legal filings.

When asked to summarize the research landscape on a topic, AI tools will sometimes populate reference lists with papers whose titles, authorship, and journal placement are all generated rather than retrieved from actual publication databases.

Specific chapter titles and page ranges are generated to match what 'should' exist given a book's stated scope — even when the book itself is real but the specific sections don't exist.

AI tools frequently state current executives, department heads, or organizational structures from their training data period — without flagging that the information may have changed. This is particularly common for government positions and technology company leadership.

Scientific findings that were widely cited at the time of training but have since been challenged, retracted, or superseded may be presented as current consensus without caveat.

Software pricing, API limits, hardware specifications, and product features change frequently. AI tools trained before a major product revision will state outdated specifications with no indication of the time gap.

Statutory requirements, regulatory thresholds, and procedural rules that were in force during the training period may be presented as current law — even when subsequent amendments have changed the applicable rules.

An AI may state a statistic correctly, then apply a formula to it correctly, then arrive at a wrong conclusion because the initial statistic was subtly wrong. The final answer inherits and amplifies the original error.

When explaining how something works, AI tools often get the conceptual process correct while filling in specific data points (dates, names, numbers) with plausible but inaccurate values.

A term or concept from one field may be applied using the rules of another field — particularly where the same word has different technical meanings in different disciplines (e.g., 'hazard' in epidemiology vs. engineering vs. insurance).

AI tools will generate quotes in a person's documented 'style' and attribute them to that person. The person is real; the quote is not. This is particularly common for academics, historical figures, and public intellectuals.

Professional credentials, institutional affiliations, and academic titles are frequently wrong for individuals who are real but not extensively covered by online sources. The model 'fills in' plausible-sounding credentials.

Where two or more people share a name or are closely associated in training data, biographical details from one may be attributed to another — creating a composite profile that accurately describes neither individual.

An AI will describe a function that logically should exist in a library, complete with a plausible name, parameter list, and return type — when the function doesn't exist. The pattern fits the library's naming conventions, which makes the error harder to spot.

Code patterns that were standard in an older framework version continue to appear in AI responses after the library has shipped a breaking change. The code runs without error in old versions and silently fails in current ones.

Behavior that changed across versions — default values, configuration option names, module paths — is often stated without version qualification, leading to version-specific bugs that can be difficult to trace.

Legal rules described accurately for one jurisdiction are applied to questions about another — particularly for areas where rules are commonly confused (employment law, privacy law, tax treatment) across US states, EU member states, or other national systems.

Specific numerical thresholds in tax law, employment law, and environmental regulation change regularly. AI tools will state the values from their training period without indicating that these figures change on statutory schedules.

Section numbers and subsection references that plausibly match the structure of real legislation but don't correspond to actual provisions are generated with the same confidence as accurate citations.