The AI "Hallucination" Problem — and Why Contract Review Is a Different Case Than Legal Research

The stories are not exaggerated: lawyers have been sanctioned for submitting briefs with AI-invented case citations that don't exist. A public tracking project, the AI Hallucination Cases Database, had documented more than 1,300 such cases globally by April 2026 — up from under 100 in mid-2025. Federal courts have issued sanctions ranging from a few thousand dollars to tens of thousands, and at least one attorney has been suspended over it.

That's a genuine, serious, and rapidly growing problem. Anyone using AI in a legal context should take it seriously — which is exactly why it's worth understanding precisely what kind of task produces this failure, and what kind doesn't.

The 13x growth in documented cases in under a year is itself worth sitting with — it suggests this isn't a problem that's being solved as the underlying models improve, at least not yet, in the specific context of unverified legal research use.

Look closely at the documented cases and a pattern emerges: every one involves AI legal research or brief drafting — asking a model to recall or generate case law, then citing what it produced without checking that the cases are real. That's a fundamentally different task from contract review. Legal research asks AI to pull facts from its training memory, where it has no way to verify against an authoritative source in real time. It's answering from recall, and recall can be confidently wrong.

Contract review is a different shape of problem entirely: the AI isn't being asked to remember a court case from months of training data — it's being asked to analyze a document that's sitting right in front of it, provided in the same request. That's a task built for verification, not recall.

Put another way: legal research asks "what does the law say," a question the model can only answer from what it happened to absorb during training. Contract review asks "what does this specific document say," a question the model can answer by actually reading the text you gave it — a categorically easier and more checkable task.

Because the contract is provided as input rather than pulled from memory, a well-built system can do something legal research tools structurally cannot: check its own answer against the source. Before presenting a finding — "this clause says X" — the system can confirm that the exact quoted language actually appears in the document you uploaded. If it doesn't, that's a red flag the system can catch and surface, rather than a confident error that slips through.

This is exactly the failure mode a contract-review tool should be engineered to close off. It's not a guarantee that every finding is legally correct — a real attorney's judgment call is still a judgment call — but it does eliminate the specific, well-documented "invented the source material" failure that's driving the hallucination headlines.

This distinction is also why comparing a contract-review tool's reliability to a legal-research tool's documented hallucination rate is genuinely apples-to-oranges — they're different tasks with different failure modes, and treating them as equivalent risk categories obscures more than it clarifies.

This is the part that gets lost in general "AI hallucinates" coverage: whether a tool verifies its findings against source text is a design decision, not an unavoidable property of AI itself. A general chatbot with no verification step and a purpose-built contract tool with a document-anchoring and second-pass verification layer are running the same underlying model technology but producing very different reliability profiles.

That's the practical question worth asking about any contract-review tool you're evaluating: does it show you where in your document a finding came from, and has it checked that the quote is real? If the answer is no, or if you can't tell, treat its output the same way you'd treat an unverified web search result — a starting point, not a finding.

Vendors who've built this verification layer tend to be happy to explain it in specific, technical terms, since it's a genuine differentiator. Vague reassurance ("our AI is very accurate") without a specific explanation of how it checks its own work is itself a useful signal.

The headline risk with AI right now is real, but it's concentrated in legal research and drafting, not in reviewing a document you already have. That doesn't mean contract-review AI is risk-free — it means the risk is different and more manageable, provided the tool is built to verify rather than just generate.

If you're evaluating any AI tool for contract work, ask specifically how it handles this. See how RCS anchors and verifies every finding against your actual contract before you decide how much to trust any tool's output.

The broader lesson generalizes well beyond contract review, too: whenever you're evaluating an AI tool for any task, ask specifically whether that task involves the model recalling facts from memory, or analyzing material you've handed it directly — because that single distinction predicts a great deal about how much you should actually trust its output.

As the documented hallucination-case count keeps climbing in the legal-research context, expect the industry conversation to keep conflating the two failure modes — which is exactly why it's worth being the person in the room who can explain, specifically, why contract review isn't the same risk category.