Guide · Legal

Verifiable AI for Legal Research: How to Stop Fabricated Citations

The defining risk of AI in legal work is the fabricated citation: a case, a quote, or a holding that reads as real and is not. Lawyers have been sanctioned for filing them, and studies of AI legal-research tools marketed as low-hallucination still measured error rates of 17% to 33%. The fix is not a more confident model. It is a system that checks whether a cited authority was actually retrieved, flags anything the model only claimed to do, and keeps an auditable record. ThinkerWave does this with a patent-pending verification step, and it runs on your own machine so privileged matter never leaves it.

The fabricated-citation problem

The failure that has actually cost lawyers is not a vague or hedged answer. It is a confident, well-formatted citation to authority that does not exist. It has led to sanctions, to withdrawn filings, and to the kind of headline no firm wants. And it is not limited to consumer chatbots: a study of AI legal-research tools sold as low-hallucination still found them producing incorrect or misgrounded answers between 17% and 33% of the time.

The reason is structural. A language model generates the most plausible continuation, and a plausible-looking citation is exactly what it is good at producing, whether or not the case is real. The output arrives in the same authoritative voice a correct answer would. For a filing or a memo, that confidence is the trap.

Why “it cites sources” is not a safeguard

Retrieval helps, and grounding an answer in a real corpus removes a class of pure invention. But it does not close the gap. A cited case can be real and still not stand for the proposition it is offered for. A quote can be stitched together from fragments. A string of citations can look thorough while one of them was never actually retrieved. The only reliable check is to compare each claim against the record of what the system truly did, not against how confident the answer reads.

How ThinkerWave catches it: verification against the record

ThinkerWave keeps a complete log of every search and computation it performs. An independent evaluator, separate from the part that wrote the answer, then cross-checks the answer’s claims against that log. If the answer asserts a case, a quote, or a calculation that has no matching entry in the record, that claim is identified as unverified, what our patent calls “phantom verification,” and penalized rather than passed through.

This matters because it targets the exact failure mode legal work cannot tolerate: an answer that claims to have checked something it never checked. The mechanism, cross- verifying an agent’s claims against its actual tool-execution record, is the subject of ThinkerWave’s patent application (Indian Patent Application #202611044024). Every quantitative claim in the output also carries an explicit flag for whether it was produced by a real computation or only asserted, so a reviewer can see at a glance what is grounded and what is not.

What we saw it do

On a hard research question, an early attempt did exactly what the safeguard is built to catch: it produced an answer supported by “evidence” it had not actually gathered. The independent evaluator flagged it plainly, noting that no real searches had been conducted and that the supporting snippets were fabricated rather than real retrieved quotes. In a separate case it caught an attempt claiming a result was “computation confirmed” when no computation had run. The system does not assume its own work is honest; it checks, and it catches the bluff, including its own.

What reliable AI for legal work looks like

Four practical properties, in order of how much they protect you:

• It verifies claims against the record, so a fabricated or unretrieved citation is flagged before it reaches a filing.
• It refuses what it cannot support, instead of manufacturing a confident answer to a question the authorities do not settle.
• It keeps an auditable reasoning trail, so the work can be checked by a supervising lawyer, not taken on trust.
• It runs locally on your own machine, so privileged documents and client matter never leave your control.

How a legal team uses it

Point it at the research or review task where a wrong answer is expensive: does this line of authority hold, what does this contract actually oblige, where is the exposure in this agreement. Describe the question and attach the materials. It works the problem from several angles, verifies what it asserts against what it actually found, flags the weak authority and the open questions, and hands back a record a lawyer can check line by line.

The lawyer stays accountable. The system does the exhausting cross-checking and shows the trail; the human makes the call with the fabrications already caught and the uncertainty made visible.

Frequently asked questions