AI for lawyers: the 9 tools vetted on client data, what Opinion 512 allows, and the 864 fake citation cases

Which AI tools can a lawyer put client files into?

A business tier is the floor, and every major assistant sells one. ChatGPT Business, Claude Team, Gemini in Google Workspace and Microsoft Copilot on a work account all keep your prompts out of model training by default, as of October 2026. The consumer apps are where Opinion 512's consent rule bites. We read each vendor's own data page on 12 October 2026:

The table is where a lawyer's reading starts, not where it ends. ABA Formal Opinion 512, issued 29 July 2024, says "all lawyers should read and understand the Terms of Use, privacy policy, and related contractual terms and policies of any GAI tool they use to learn who has access to the information that the lawyer inputs." California's State Bar puts the same duty more bluntly in its 2026 Practical Guidance: "Reasonable efforts require more than reliance on generalized marketing assurances." So a partner who signs up for ChatGPT Business still reads the DPA, and checks whether web search is on before a client name goes into a prompt.

The opinion also draws the line that saves most of a lawyer's AI use from the consent problem. When you use a tool "for idea generation in a manner that does not require inputting information relating to the representation", Opinion 512 says no client consent is needed. An outline of a CLE talk, a plain English rewrite of a public statute, or a draft engagement letter template needs no client facts at all. What never to paste into AI lists the data that does.

How many lawyers have been caught filing AI-invented citations?

Lawyers filed 864 of the 2,184 decisions in Damien Charlotin's AI Hallucination Cases database at its 10 October 2026 update, and 1,492 of all the decisions come from US courts. Self-represented litigants filed 1,259. We downloaded the database's CSV on 12 October 2026 and counted the lawyer filings by the year of the decision:

The legal research platforms aren't a way out. Most entries say the tool is implied or unidentified, but 155 of the lawyer decisions name it. ChatGPT or another OpenAI product appears in 82 of those. Westlaw, CoCounsel or Lexis appears in 26: Westlaw or CoCounsel in 16, Lexis in 13, and three name both. At least 19 more name another product sold to lawyers, such as Fastcase, vLex, Paxton, Irys, Spellbook or Eve. That matches the Stanford test published in the Journal of Empirical Legal Studies in 2025. It found that Lexis+ AI and Westlaw's AI research tools "each hallucinate between 17% and 33% of the time".

The bill is uneven. Of the 586 US decisions where a lawyer filed the error, 171 list a money penalty or a costs award. Thirteen of those carry the database's $1 placeholder for an amount it couldn't find, and across the other 158 our count puts the median at $3,000. The largest is $110,204 in Couvrette v. Wisnovsky (D. Or., 23 March 2026), a $15,500 sanction plus about $94,700 in the other side's costs. The best known, Mata v. Avianca (S.D.N.Y., 22 June 2023), cost two lawyers and their firm $5,000 for six decisions ChatGPT invented. How to fact check AI answers has the quick test for a fake cite, and AI for paralegals walks through Mata and two later sanctions in detail.

What do California, Florida and New York require of lawyers using AI?

California asks the most, and from 1 January 2027 part of it is statute. Governor Newsom signed SB 574 on 30 September 2026, as Chapter 858. It adds section 6068.1 to the Business and Professions Code: an attorney "shall not delegate the practice of law" to generative AI, must take reasonable steps to verify AI output, citations included, and must disclose AI use to the court for all documents submitted to it. It bars confidential, nonpublic or personally identifying information from any generative AI system unless access is limited to the attorney and authorized people bound by confidentiality, which Holland & Knight says may rule out some consumer tools. An amended Code of Civil Procedure section 128.7 says no filing may carry a citation the responsible attorney hasn't personally verified. The State Bar's Practical Guidance, first issued in November 2023, was replaced by a 2026 version that adds agentic AI.

Florida's rule is softer. Florida Bar Ethics Opinion 24-1, adopted 19 January 2024, says "it is recommended" that a lawyer get the client's informed consent before using a third party generative AI program with confidential information, and that no consent is needed when nothing confidential is disclosed. It adds a rule most firms miss: a website chatbot "must include a disclaimer indicating that the chatbot is an AI program and not a lawyer or employee of the law firm."

New York has guidance, not a statute. The New York State Bar Association's House of Delegates approved its Task Force on Artificial Intelligence report on 6 April 2024. It says that even when a client consents, the lawyer "should obtain assurance that the Tool provider will protect your client's confidential information." The New York City Bar's Formal Opinion 2024-5, dated 7 August 2024, adds in a footnote that when a client supplies citations, "a lawyer must review the decisions to make sure that they are genuine and properly cited." California and Florida agree with Opinion 512 on billing. You can't pad an hourly bill with time the tool saved, and the ABA says a lawyer "may not charge a client to learn about how to use a GAI tool" used for clients regularly.

Which legal AI tools are built for lawyers, and what do they publish?

Harvey, CoCounsel and Lexis+ with Protégé are the three names on almost every list, and none of them publishes a price. Each says it doesn't train on customer data. Here's what each vendor's own site states, read 12 October 2026:

Page one's "best AI for lawyers" lists run to 10 or 15 names. Most of the extras do other jobs: Clio runs the practice, Relativity and Everlaw handle discovery review, and Luminance competes with Spellbook on contracts. These five cover the jobs a lawyer asks AI to do: research that reads real authority, review of a firm's own documents, and contract drafting. The research column is the one to pay for. CoCounsel and Protégé answer from the reporters you'll cite, so a wrong answer is easier to catch. A general assistant answers from its training and the open web. That's no guarantee: Westlaw, CoCounsel or Lexis is named in 26 of the lawyer cases above.

Which AI prompts work for lawyers, and how do you check the output?

The prompts that work give the model a document and forbid it from citing law, so every claim it makes points back to something you can open. Here are three, each with its check. The first needs a tier that's cleared for client files. The other two need no client facts.

A first pass on a contract, on a business tier or a legal tool:

```

You are reviewing a [contract type] for [our client's role, no names]. Use only the attached document. List every clause that shifts risk to our side: indemnity, limitation of liability, termination, assignment, governing law, renewal. For each one, quote the clause word for word with its section number, say in one sentence why it matters, and mark it standard, unusual or missing. If a clause type is absent, write ABSENT. Do not cite any case, statute or rule.

```

The check: open every section number and read the quote against the page. A quote you can't find is the model writing, not reading.

A research plan with no authority in it:

```

I am researching [issue] in [jurisdiction] for [motion or memo]. Do not cite any case, statute or rule. Give me the elements to prove, the question a court asks on each element, where jurisdictions split, and 10 search strings I can run in Westlaw or Lexis.

```

The check: every authority comes from the research platform, and each one gets KeyCited or Shepardized before it goes in a draft.

A cite list for the final check, with no client facts:

```

Below is my draft brief. Put every citation in a table: the citation as written, the proposition the brief cites it for, any quoted words, and the pin cite. Do not say whether a case exists or is good law. Do not fix or add anything.

```

The check is yours alone: pull each case, find the quoted words at the pin cite, confirm the holding, and read later history. In California that last pass is the law from 2027, since SB 574 wants the responsible attorney to have personally verified every cite. Does ChatGPT make up sources covers why models invent the citations this check catches.

These prompts run in any assistant. In Whizi, which publishes this page, the Pro plan at $29.99 a month runs Claude Opus, Claude Sonnet and ChatGPT Sol in side by side model chats, so a second model can attack the first one's research plan. Whizi doesn't use prompts or files to train Whizi-owned models. But each message goes to OpenRouter and the model's maker, Whizi's privacy policy won't promise zero retention across providers, and neither the policy nor the docs offer a data processing addendum. By Opinion 512's baseline that makes it a tool for the second and third prompts, not the first.