Tracking hallucination marketing claims from legal tech vendors
Some people have not heard of the WayBack Machine
When parties in legal proceedings explain what led them to file briefs riddled with hallucinations, they often recount having been too trusting of the claims by legal tech vendors, and especially efforts to soothe concerns about this pitfall.
Here is a typical example, taken from a recent case, of an attorney confessing:
Earlier this year, I began using an AI program called EVE to help me draft pleadings. I was hesitant to use AI because of the horror stories I heard about hallucinated case law. EVE made several presentations to my law firm and assured me in training sessions there were safeguards in place to minimize hallucinations.
[…]
Here is the marketing language from Eve’s website:
Can you trust your Al to do legal work? It is not enough to minimize the risk of poor answers or hallucination. At Eve, we’ve worked up an extra layer that we add to the [retrieval-augmented generation, or RAG] methodology and increase the trust (and ease of verification) of AI’s answers. We call it “Trust but Verify”.
Reading this triggered my interest in checking whether an intuition I had was true: that vendors have historically touted their products as hallucination-free (or in some other way guaranteed no hallucinations), but that they later backtracked from this position in light of the evidence - leading to the new “trust but verify” byword many of them have now adopted.1
And so, in this piece, I want to track the claims made by some LegalTech vendors2 in the past and today with respect to how they handle hallucinations from their offerings. I am relying on internet-based written marketing material, trying to highlight the changes in how these products are and were presented.3 (Links below commonly lead to the WebArchive, which takes snapshot of websites at various times.)
We’ll see that my intuition was not entirely true: the main vendors were a bit more cautious than I had first thought, though most still overclaimed in this respect and eventually backtracked, at least implicitly.
Prelude: the Stanford study
But first, a bit of recent history so as to get a rough timeline.
ChatGPT is introduced in November 2022 and puts generative AI in everyone’s mental map. The first legal-only chatbot/AI tools are introduced over the course of 2023, with several iterations up to the present day. This includes historical legal editors tooling up, legal tech pure players trying to disrupt them and, as we saw recently, model providers themselves adding legal toolkits to their offerings.
In 2024, a study by a team from Stanford had a look at the claims of reliability by the main legal tech vendors, and found them wanting (Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools).
This is a very nice study. But this is also, by now, an old and outdated study, that was already beyond its time when it went out: the authors recognised that they did not test newer models and approaches.4
As such, people (they are many) citing this study in 2026 to make a point against legal tech vendors are foolish and cringe. The headline results are far from what the state of the art can now provide in terms of accuracy and “grounded” results. I have my cavils as to whether a generated text can ever be “accurate”,5 but I cannot and would not deny that tools, be they legal-tech specific or off-the-shelf models, are much better now than they used to be, including with respect to hallucinated legal material.
At the same time, the opposite attitude - models never hallucinate, thanks to thinking/RAG/agents/whatever-other-silver bullet - is also foolish, and constantly belied by practice. Which is why most actors in this field have moved away from that claim, knowing they can’t guarantee it.
But the real value of the 2024 Stanford study for this piece is that it offers an easy time threshold to compare claims by legal tech vendors: before that, a legal executive (or marketing team) could have relied on their engineers’ claims that hallucinations are solved with RAG; after that, this position became much harder to maintain.
And with that out of the way, let’s move to the examples, starting with.
LexisNexis (Lexis+ AI, Protégé)
This is the most striking example, and needs to be unpacked carefully.
In April-May 2024, Lexis Nexis’s VP of Product Management posted a blog post entitled “How Lexis+ AI Delivers Hallucination-Free Linked Legal Citations”, pointing out that:
We raised some eyebrows last fall when we announced the launch of Lexis+ AI, our new generative artificial intelligence (Gen AI) solution designed to transform legal work, “with linked hallucination-free legal citations” that are grounded in the world’s largest repository of accurate and exclusive legal content from LexisNexis.
This, at least, was in the original post, now accessible through the WebArchive. If you visit the same URL now, the post’s title has changed to “How Lexis+ AI Delivers Trustworthy Linked Legal Citations”, and the opening paragraph just quoted has … disappeared.
The rest of the blog post, then and now, did qualify and elaborate on the claim of zero hallucinations (“our promise is not perfection, but that all linked legal citations are hallucination-free”). However, even this is not a claim they are still making, apparently.
(For completion, I noted that (i) the URL has not changed, despite the title switch, and it still records the “hallucination-free” claim; and (ii) you can still find a version of the original post, including the original title and the “hallucination-free” initial paragraph, in the version of this post for the Australian market.6)
Anyhow, Lexis’s original claim has not gone unnoticed: at least one party in a US case cited the exact paragraph quoted above, further stating that “Plaintiff’s reliance on Lexis+ AI was grounded in the platform’s reputation and its explicit assurance of citation accuracy”. That attorney later got fined 3,500 USD, was ordered to inform the Bar about it and, presumably, cancelled his subscription.
Lexis eventually changed tack, and its subsequent offerings were instead set, sensibly, in terms of minimising or reducing hallucinations, for instance when it introduced Protégé in January 2025.
Thomson Reuters (Westlaw Precision, CoCounsel, Practical Law)
The story is a bit more complicated for legal editor giant Thomson Reuters (“TR”), parent of Westlaw and Practical Law.
CoCounsel, TR’s chief offering in this field, had been developed through CaseText - acquired by TR in August 2023. Before its website was shuttered, CaseText had been explicit that:
Unlike even the most advanced LLMs, CoCounsel does not make up facts, or “hallucinate,” because we’ve implemented controls to limit CoCounsel to answering from known, reliable data sources—such as our comprehensive, up-to-date database of case law, statutes, regulations, and codes—or not to answer at all.
TR itself, in introducing Westlaw Precision, sought to soothe concerns about hallucinations, stating:
We avoid [hallucinations] by relying on the trusted content within Westlaw and building in checks and balances that ensure our answers are grounded in good law.
At the same time, the same piece by TR had also advised attorneys to “be sure you’re checking the sources the results come from and using it as a starting point for your research”.
This is probably why, in its later marketing of CoCounsel, TR eschewed the earlier claims that the tool does not hallucinate, warning that it sought to reduce or minimise hallucinations, and calling on users to verify everything. Indeed, I also found an early report from one of its principals couched in terms of caution and professional responsibility that remains a good read today.
V-Lex / Vincent / FastCase
Up until mid-2024, V-Lex’s main marketing page for its Vincent AI tool had this to say about hallucinations:
Safeguarding against hallucination
Authorities are sourced directly from the largest collection of legal and regulatory information, housed by vLex, to ensure only real cases and materials are referenced. By using trusted data, Vincent AI can safeguard against the dangers of unconstrained LLMs.
Needless to say, this did not last, and the “Safeguarding against hallucination” part of the marketing material later changed tack to insist on the “Trust but verify” principle - and highlight how Vincent AI can help with this, a much more defensible approach.
Still, a certain ambiguity remained; for instance, in a blog post from February 2025 announcing the new version of their tool, V-Lex boasted that “Unlike general-purpose AI tools prone to hallucination, Vincent AI is powered by vLex’s vast library of structured legal data, providing added confidence with citations to verified authoritative sources.”
Harvey
Despite being one of the earliest players in that field, I could find little evidence that Harvey ever slipped and promised hallucination-free outputs
On the contrary, Harvey has been remarkably transparent about its hallucination rate, merely promising that they have reduced it in their new version published in August 2024, and eventually publishing a blog post and benchmark on this subject in October 2024.
Legora
Nothing to say about them ! Archival data to work with is scarce, and the few snapshots on the WayBack Machine are not easy to exploit.
Still, just like Harvey, it appears they played it smartly in this respect.
CourtAid
This is one minor vendor used by a lawyer in Australia (together with ChatGPT) that found himself in the hallucination database.
And no wonder: CourtAid has a December 2023 blog post, still online, in which the company claims that “CourtAid.ai sets itself apart by eliminating the potential for misinformation and hallucination”.
A claim that does not seem to have been repeated in its marketing material going forward, matching the pattern of the other vendors.
UnlikelyAI
This startup initially marketed itself chiefly around claims of being hallucination-free; this is the very first thing mentioned on their landing page in mid-2025. Referring to the launch, and drumming up the hype, this LinkedIn post asked:
Hallucination-free AI: could this be the trust breakthrough legal has been waiting for?
Reader, it was not: in the most likely scenario ever to occur, UnlikelyAI eventually dropped “hallucination-free” from its marketing material, although they still imply that their solution does not hallucinate, thanks to neurosymbolic AI or something.
****
I am not here to point fingers, especially since the tally here turned out to be more complicated than I imagined, and there were as many overclaimers and ambiguous hedgers as careful operators. But it’s good that fewer actors now boast “zero hallucination”,7 and there is hope even those will eventually relent.
And all in all, this review goes to my frequent point that AI is an immature technology, and that we are all learning its powers and limits as we move along: likewise, the legal-tech industry is growing up, one quietly edited webpage at a time.
Writing this piece taught me that “Trust, but verify” is a Russian proverb, and has its own Wikipedia page.
Readers should feel free to let me know if I missed important ones, and I will feel free to edit this page as often as needed.
This is a limitation, and it’s likely that human vendors and the personnel providing training and onboarding for these products have been rather more optimistic - or reckless - in addressing the concern over hallucinations.
Additional pushback came from the vendors, although they could not dispute the main finding that their solutions did hallucinate at times: see here for a timeline.
Namely, to the extent you are relying on a probabilistic model, you are still at the end of it getting lucky that the results match “accurate” or “grounded” results, even if that’s 99.9% of the time. But I am conscious that this is a galaxy-brained take, and does not detract from the fact that, in practice, 99.9% might work out enough for most use cases.
I should also mention that, at some point mid-2025, Lexis retroactively added a disclaimer to all posts on their “Legal Insights” pages, stating that the views from “externally authored materials” published on “this site” (i.e., lexisnexis.com) do not necessarily reflect their views. This looks very strange as applied to a blog post by their VP of Product Management.