When AI searches thousands of legal sources

You must now record which lawyer verifies which sources before legal advice leaves your firm, because AI systems that search hundreds or thousands of authorities do not guarantee correctness — they shift the volume of checking you must do, and with it your liability for what you missed.

An analysis of 20 October 2026 of a research mode searching hundreds to thousands of legal sources argues that breadth of search does not produce guaranteed accuracy, and that the signing lawyer must verify every cited passage, jurisdiction and date. The vendor describes a system that typically assesses 800 to 1,500 sources, sometimes exceeding 2,000, and flags each as highly relevant, relevant, flagged or rejected. In our assessment, the development matters not because it produces a smarter answer, but because it moves the economic boundary of research — work that once stopped when time or budget ran out can now run further, including unfavourable and older rulings. That shift places the burden of systematic checking squarely on you.

Does searching more sources mean the answer is more reliable?

No. A larger number of searched or cited sources says nothing about whether a proposition is actually supported by the cited passage, whether the ruling remains good law, or whether it applies to the correct jurisdiction. Research in the Journal of Empirical Legal Studies found substantial reliability problems among commercial legal-AI tools: performance varied by system, with some producing inaccurate, hallucinated or incomplete answers in a substantial share of test questions. Earlier work in the Journal of Legal Analysis described how general language models can fabricate or misrepresent legal authorities and are poorly calibrated about their own reliability. A randomised experiment at the University of Minnesota Law School found that AI assistance improved performance in some structured tasks, but participants who began with stronger analyses sometimes produced lower-quality work after incorporating AI-generated revisions. The consistent lesson is that AI can broaden the candidate list of authorities, but does not shift the judgement itself. Those who leave the reasoning to the system rather than using it as a tool run the risk that sound analyses are diluted.

What concrete duties does the wider search impose on your firm?

Three practices follow directly from the research findings and from the vendor's own account of how the system works:

1. Record which lawyer verifies which sources before the advice is signed — document in advance which member of your team reads and approves which categories of sources, and do not leave verification to whoever has time available.
2. Sample and document the rejected sources — because the vendor states that sources can be filtered as rejected or flagged, a relevant counter-ruling may remain out of view; open a sample of excluded sources and record why exclusion was justified before the memo is sent.
3. Read back every load-bearing proposition in the full ruling, not the summary — because independent research shows these tools can attach incorrect propositions to real rulings, a citation list does not suffice as evidence; verify each key claim against the full text and sign off that check per proposition.
4. Establish a named senior decision point on which authorities govern the client's facts — because the final advice now rests on more authorities, any error reaches further; revise your internal liability and review protocol with a recorded determination by a named partner or director on which rulings control the client's circumstances.

How should you deploy this capacity in practice?

Use the wider search as a first sift and as a counterweight, not as a finished product. The system can surface authorities you would have missed under time pressure, and can show you rulings that conflict with your initial direction. That is valuable. It does not, however, replace your own reading. The difference between more available research and better-founded advice lies in who reads, checks and signs — and now, in recording that someone did.

What does the regulatory framework require?

Under the EU AI Act, human oversight in professional use remains a requirement. Searching more widely changes nothing about that obligation. The framework does not prescribe how you verify; it requires that you do, and that you can demonstrate it. The shift from dozens of sources to hundreds or thousands makes that demonstration harder to improvise after the fact.

Tooling can surface authorities and flag conflicts. It cannot read a ruling for you, cannot judge whether it applies to your client's facts, and cannot carry the professional responsibility for the advice you give. That remains yours to exercise, and now, to record.