AI
Google's own documentation describes AI Overviews as built on top of its normal ranking systems and generally resistant to the hallucinations that affect standalone chatbots. That claim is worth testing against independent data before you rely on an AI Overview — or any AI-generated summary — while reviewing a contract, a filing, or a compliance requirement.
What the research actually found
A 2026 academic study tracked 55,393 trending queries across 19 topic categories over 40 days, extracting 98,020 individual factual claims from the resulting AI Overviews and checking each one against the pages the Overview itself cited as its source. The headline number: about 11% of claims were unsupported by the cited pages, and the dominant failure mode was omission — the Overview stating something the source didn't actually say, rather than directly contradicting it. That's a subtler and more dangerous failure than an obvious factual error, because the citation link is present and looks legitimate; it just doesn't back up the specific sentence attached to it.
The same research flagged specific categories where inaccuracy concentrates: claims about named individuals or businesses, precise statistics and percentages, events after a model's training cutoff, niche/specialized topics, and — explicitly — legal information. Legal document review sits at the intersection of several of these risk categories at once: it typically involves specific named parties, precise figures (dates, dollar amounts, statutory citations), and often niche or jurisdiction-specific rules.
Courts are starting to treat this as a real liability question
This isn't a hypothetical concern. In May 2026, a regional court in Munich ruled that Google can be held directly liable for a false statement made in an AI Overview about an identifiable person, on the reasoning that the Overview generates a new, independent statement rather than simply linking to a third party's claim (Google is appealing; the ruling is specific to German law and one type of defamation claim, not a general finding that all AI Overview content is legally actionable everywhere). Separately, attorney Damien Charlotin maintains a public, continuously updated database of court cases in which a party's use of AI produced hallucinated citations or facts that reached a judge — it has documented over 1,600 cases, with sanctions ranging from stern warnings to five- and six-figure fines. Several of those sanctions were issued to lawyers who cited case law an AI tool invented outright.
The practical lesson for anyone using AI to accelerate legal document review: an AI-generated summary, overview, or citation is a starting point for verification, not a substitute for reading the underlying statute, contract clause, or case text yourself. If a chatbot names a case, cites a code section, or quotes a figure, open the primary source before it goes in front of a client, a court, or a compliance file.
| Use Case | AI Overview / Chatbot Answer Reliability | Recommended Practice |
|---|---|---|
| General topic orientation ("what is a force majeure clause") | Reasonably reliable for well-established, widely-documented concepts | Fine to use as a starting explanation |
| Named case law, statute citations, specific figures | Highest documented hallucination risk category | Always verify against the primary source (court record, statute text, official filing) before relying on it |
| Client- or matter-specific document review | Not evaluated by the cited research — no accuracy guarantee exists | Use AI only to flag sections for human review, not as the review itself |
Practical Challenge
Take an AI Overview or chatbot answer to a legal question you're curious about. Open every source it cites and check whether each specific claim is actually stated in that source, or just adjacent to something the source says. Note how often the two diverge.
Concept Check
Sources & Further Reading
- Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact (arXiv, 2026) — the study behind the 98,020-claim, 11%-unsupported-claims figure cited above, including the category breakdown for where hallucination concentrates.
- AI Features and Your Website — Google Search Central — Google's own documentation on how AI Overviews are generated and what it claims about their reliability.
- AI Hallucination Cases Database — Damien Charlotin — a continuously updated, searchable database of real court decisions involving AI-hallucinated legal citations and the sanctions courts have imposed.