Welcome back. I took August off, the first real break since I started this, and the world was decent enough not to end while I was away. Did you miss me? If so, there is a concrete way to say it: the paid tier runs $8 a month, about the price of one coffee, and it is what keeps these pieces coming. One note before we start, half good news and half not. This newsletter just crossed 2,100 subscribers, a number I still do not quite believe when I look at it. The less cheerful half: revenue is actually down even as the list has grown, which is the diplomatic way of saying most of that new growth is reading for free. That is what the free tier is for, and I am glad you are here. But if the work has earned a place in your week, a paid subscription is the difference between me spending weeks on an investigation like this one and spending an afternoon. If you can help, please do. Bloomberg: $35/month. Financial Times: $42/month. The Economist: $17/month. Original analysis by Tatsu with 18 footnotes: $8/month. Share this with anyone about to sign a legal AI contract. Every major legal AI vendor has, at one point or another, sold its product on a version of the same promise: it will not make things up. LexisNexis advertised "100% hallucination-free linked legal citations." Thomson Reuters said it avoids hallucinations "by relying on trusted content within Westlaw." Casetext, the company behind CoCounsel, said its tool "does not make up facts, or hallucinate."[1] In 2024, a team at Stanford's RegLab decided to test that promise against reality. They wrote 202 legal research questions, ran them through the leading tools, and had legal experts score every answer. The study was preregistered and later survived peer review in the Journal of Empirical Legal Studies.[1] It is the closest thing this industry has to an independent audit, and its finding is not ambiguous. The leading grounded legal research tools hallucinate between 17 and 33 percent of the time.[1][2] Broken out by product, the spread is worse than the headline. LexisNexis's Lexis+ AI was the best performer tested, and it was still fully accurate only about 65 percent of the time, hallucinating on more than one answer in six. Thomson Reuters's Westlaw AI-Assisted Research was accurate roughly 42 percent of the time and hallucinated on nearly one answer in three. Thomson Reuters's Ask Practical Law AI was accurate on fewer than one in five questions, and it managed that partly by refusing to answer 62 percent of the time. A tool that declines to answer two questions out of three is technically not hallucinating, in the same sense that a witness who takes the Fifth is technically not lying.[3] The vendors said hallucination-free. The study printed the error rate directly beneath the marketing copy. This matters because the downside is sanctions. Lawyers have been fined and referred to bar discipline for filing briefs with citations that turned out to be invented,[4][5] and "the software told me it was real" is not a defense that has worked for anyone. When a firm buys a legal AI tool, it is buying a promise about citation integrity, and that promise is the thing the leading study specifically measured and found wanting. To be fair to the category, grounding does help enormously. An earlier Stanford-affiliated study, "Large Legal Fictions," ran more than 800,000 queries through general-purpose models and found hallucination rates of 58 percent for GPT-4, 69 percent for GPT-3.5, and a spectacular 88 percent for Llama 2.[6] Against that baseline, a specialized tool that is right two-thirds of the time is a real improvement. The problem is the gap between "meaningfully better than raw ChatGPT" and "hallucination-free," because only one of those two phrases appeared in the sales deck, and it was not the accurate one. That is the scandal you can measure. The structure underneath it, the part that determines who you actually pay and what you actually get, is the one worth the subscription. Below the paywall: * Why the Harvey-versus-Claude frame is rigged, and the two incumbents it quietly deletes * Who really owns the grounding behind every "trusted" answer, and why Harvey subleases its credibility from a direct competitor * The DPA-versus-ZDR trap: how firms get their data retained after being told it was safe * The "enterprise-grade" model with the weakest data passport in Europe * How to actually procure one without overpaying on the parts nobody benchmarks $8/month, 18 footnotes, no vendor deck. 14-day free trial, cancel anytime. Harvey Versus Everyone Is the Wrong Question Walk into most legal AI procurement conversations in 2026 and you will be handed a binary: buy a managed legal platform, meaning Harvey, or deploy a frontier model, meaning Claude, and build the legal layer yourself. It is a clean frame. It is also a false one, and the two options it quietly deletes are the two that should worry the incumbents most. On the managed side, Harvey is a leader, not the category. Its most direct competitor is CoCounsel, owned by Thomson Reuters through its $650 million Casetext acquisition[7] and grounded in Westlaw.[8] Sitting right next to it is Lexis+ AI, the first-party product from LexisNexis, built on the very research corpus that Harvey pays LexisNexis to license. Then there is Legora, a fast-growing entrant that is particularly strong in Europe,[9] plus a specialist layer, Robin AI and Spellbook and Luminance for contracts, Paxton AI and vLex Vincent for research. On the frontier side, "native model" has been treated as a synonym for Anthropic, which is strange, because the single most-deployed enterprise model is OpenAI, and it is one of the models running inside Harvey.[10] The honest native column is Claude, OpenAI, Google Gemini, and Microsoft Copilot, evaluated on where they differ rather than on the governance baseline they all now share. Any evaluation that names only Harvey and Claude has skipped the incumbents with the deepest moats and the vendor with the largest install base. That omission is the shape of the market's own marketing. Rent Duopoly Nobody Puts in the Deck Here is the fact that reorganizes everything else. Citation integrity is primarily a property of the research corpus the model is grounded in, not the model itself, and that corpus is a duopoly: Thomson Reuters's Westlaw on one side, LexisNexis on the other. KeyCite and Shepard's, the two citators that tell you whether a case is still good law,[11] are each locked to one of those two houses. CoCounsel grounds on Westlaw. Lexis+ AI grounds on LexisNexis. And Harvey, the independent, grounds through a commercial partnership with LexisNexis,[12] which happens to sell a competing product built on the same data. Harvey pays LexisNexis for the grounding that makes it trustworthy. LexisNexis also sells its own tool that does the same thing. Somewhere inside that sentence is a renewal negotiation. For a buyer, this has two consequences the vendor will not volunteer. A tool grounded on a single corpus inherits that corpus's coverage gaps and its citator's judgment calls about which authority still counts. And a tool that grounds through a partnership rather than owning the data is exposed to the commercial terms of that partnership in a way a first-party product is not. You are not just choosing an interface. You are choosing a landlord, and in Harvey's case, a landlord who subleases from a competitor. "Hallucination-Free" Is the Tell, Not the Feature Return to the marketing claim, because its persistence is diagnostic. No vendor has published a benchmark showing zero hallucinations. They have published the sentence. The Stanford team went looking for the evidence behind "hallucination-free" and found the phrase unsupported, which is why the study quoted it by name. The deeper hole is this: the two products most relevant to an actual Harvey purchase decision, CoCounsel and Harvey itself, have no independent hallucination benchmark at all. The Stanford study tested Lexis+ AI, Westlaw AI-Assisted Research, Ask Practical Law AI, and GPT-4. It did not test CoCounsel, and it did not test Harvey. So when a competing deck cites "the Stanford numbers" against CoCounsel, it is misattributing figures that belong to a different Thomson Reuters product.[3] The two most-sold platforms in the category are, empirically, unmeasured. A newer benchmark from Vals AI in October 2025 showed real improvement across the tools it tested, but the market leaders, Westlaw CoCounsel and LexisNexis and vLex, opted out of the legal research portion, so those flattering numbers describe smaller tools, not the ones a firm would shortlist.[13] The takeaway is that "hallucination-free" is a claim the leading study disproved, and any procurement process that accepts it at face value has skipped its one job. Contract You Signed Is Not the Contract You Think Now to the governance fine print, where a subtle conflation costs firms real exposure. Two contractual objects get treated as one, and they are not the same thing. A Data Processing Addendum limits how a vendor uses your data and secures how it is processed. It typically still permits the vendor to retain that data for a window, for abuse monitoring and debugging. Zero Data Retention is the different promise: that prompts and outputs are not persisted at all. ZDR is a separate, sales-approved, endpoint-by-endpoint amendment. It is not a default. OpenAI's own documentation says as much in plain language, and enterprise chat and API products default to controlled retention measured in weeks, not zero.[14] The contract everyone signed says the vendor will not train on your data. It does not say the vendor will not keep it. Those are different promises, and firms tend to learn the difference in that order. This is where Harvey has a genuine, defensible edge that survives scrutiny: it contractually requires both no-training and Zero Data Retention across all of its underlying