ChatGPT showed up in the legal profession the way most disruptive technologies do: a few early adopters using it well, a few public failures making the news, and most lawyers somewhere in the middle, quietly experimenting and trying to figure out where it actually belongs in their practice. Three years in, the picture is clearer. The American Bar Association has issued a formal ethics opinion on generative AI. Most state bars have followed with their own guidance. A handful of high-profile sanctions cases have made the limits painfully obvious. And firms that learned to use these tools well have started to pull ahead in research speed, document review throughput, and client intake response time. This guide walks through what ChatGPT actually does well in legal work, what it does poorly, the ethics rules every U.S. lawyer needs to know before they paste a single client fact into a chat window, and the practical workflow patterns that produce real productivity gains without putting your bar license at risk.

For related coverage on technology-driven practice changes, see our reporting on law firm culture and professional health and our guide to professional well-being for lawyers.

What ChatGPT actually is — and what it isn’t

ChatGPT is a large language model (LLM) built by OpenAI. At a technical level, it predicts the most likely next word given the words that came before, trained on a very large corpus of text. The version most lawyers are using day-to-day is GPT-4o or GPT-4.1 in a chat interface, sometimes inside a paid Team or Enterprise tenant where retention and training settings can be configured. It is not a database, not a search engine, and not a legal research platform. It does not have access to current case law, current statutes, or your client’s file unless you give it those materials inside the chat. When it answers a legal question without a source, it is generating plausible-sounding text — sometimes accurate, sometimes confidently wrong.

Treating ChatGPT as a brilliant, fast, fundamentally unreliable junior associate who needs every piece of work checked is the framing that produces good outcomes. Treating it as a research database produces sanctionable mistakes. The difference between those two framings is the difference between a productivity gain and a malpractice exposure.

LegalEagle’s widely-watched explainer on what happens when lawyers use ChatGPT carelessly — including the Mata v. Avianca case that became the canonical cautionary tale.

The ethics framework: ABA Formal Opinion 512 and what it requires

In July 2024 the ABA Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512, the first national ethics opinion specifically on generative AI in legal practice. It does not impose any new rule. It applies the existing Model Rules of Professional Conduct to GenAI use. The five duties it walks through:

  • Competence (Model Rule 1.1). A lawyer must understand the benefits and risks of any technology they use. With GenAI that means knowing how the tool works, what kinds of errors it produces, what data the tool retains, and whether the output is suitable for the use case.
  • Confidentiality (Rule 1.6). Inputs to a public consumer chatbot may be retained and used for training. A lawyer must not enter information that could allow a third party to identify or reconstruct a client’s confidential information unless the tool’s data handling, retention, and training-opt-out settings are appropriate — and unless the client has given informed consent where required.
  • Communication (Rule 1.4). Where the use of AI materially affects the matter, clients should be told. The opinion stops short of requiring per-task disclosure for every use, but flags that meaningful AI use in core legal work is the kind of thing clients are entitled to know about.
  • Supervision (Rules 5.1 and 5.3). Partners and supervising lawyers must train associates and staff on appropriate AI use, set firm policies, and supervise outputs. AI is treated like a non-lawyer assistant — useful, but the responsible lawyer owns every output.
  • Fees (Rule 1.5) and candor to the tribunal (Rule 3.3). Billing for AI-generated work that took 5 minutes as if it took 5 hours is a fee violation. Filing AI-generated content that includes fabricated citations is a candor violation, and courts have started imposing serious sanctions for it.

State bars have largely tracked Opinion 512. Florida’s Ethics Advisory Opinion 24-1, California’s GenAI guidance, and the New York State Bar Association’s report all reach materially similar conclusions on the duties of competence, confidentiality, supervision, and disclosure. The variation is mostly at the margins — for instance whether oral disclosure to a client is sufficient, or whether written disclosure is required for material AI use.

The hallucination problem and the cases lawyers should know

The single biggest source of professional discipline arising from GenAI in 2023-2025 has been “hallucinations” — confident, fluent, completely fabricated case citations. The canonical example is Mata v. Avianca, Inc. (S.D.N.Y. 2023), in which a New York lawyer submitted a brief containing six fictitious cases that ChatGPT had invented. The court imposed $5,000 in sanctions, mandatory disclosure to the named judges, and a public order that put the lawyers’ names in every legal-tech ethics CLE for the next decade. The lawyers were not sanctioned for using ChatGPT. They were sanctioned for failing to verify its output before filing.

Mata was the first widely-publicized incident. It is not the last. Federal and state courts have since sanctioned lawyers in Park v. Kim (2nd Cir. 2024), Mid Cent. Operating Eng’rs Health & Welfare Fund v. HoosierVac, and a steady drip of state-court matters where filings cited cases that did not exist. The pattern is consistent: a busy lawyer asked ChatGPT for case law on point, did not check the citations against an authoritative database, and filed the brief. The fix is also consistent: every citation in any document that will be filed, sent to opposing counsel, or relied on for a substantive legal conclusion must be verified in Westlaw, Lexis, Bloomberg, Fastcase, or another authoritative source before the document leaves the lawyer’s hands.

Where ChatGPT actually earns its keep in a law practice

Used inside its competence envelope, ChatGPT is genuinely useful. The categories where lawyers consistently report meaningful productivity gains:

  • Drafting first cuts of routine documents. Demand letters, engagement letters, deposition outlines, summary judgment background sections, deposition prep questions, settlement letters, intake summaries. The model produces a structured first draft in seconds; the lawyer edits, customizes, and signs off.
  • Summarizing long documents. Discovery productions, deposition transcripts, regulatory filings, expert reports. Paste in (or upload, in Enterprise tenants) the document and ask for a structured summary, key admissions, inconsistencies, or witness chronology.
  • Translating between registers. Turning a technical legal memo into a plain-English client email. Turning client narrative into a chronological fact pattern. Turning a discovery response into a witness outline.
  • Brainstorming and issue spotting. “Here are the facts of a wrongful termination matter — what causes of action and defenses should I be thinking about?” The model generates a useful checklist that a lawyer then validates.
  • Document review triage. First-pass relevance review on a small to mid-sized production, with a lawyer checking the model’s calls. For large productions, purpose-built legal tools (Everlaw, Relativity AI, Reveal) outperform a general chatbot.
  • Marketing, intake, and client-communication content. Blog posts, FAQ content, intake forms, automated client emails, social media drafts. Low ethics risk, high time savings.
  • Coding and spreadsheet work. Writing a short Excel formula, parsing a CSV, building a quick automation. Lawyers without technical training get genuine leverage here.

Where ChatGPT should not be doing the work

The categories where general-purpose ChatGPT consistently underperforms or creates outsize ethics risk:

  • Citation-driven legal research. Use Westlaw AI, Lexis+ AI, Vincent AI, vLex, or another tool that is grounded in a verified case database. General ChatGPT does not have reliable access to current case law and will fabricate citations confidently.
  • Tax, immigration, securities, and other rapidly-changing regulatory practice. The training data is a snapshot in time. Rules change. The model will not know about a regulation issued after its training cutoff and may not know about updates inside its training window.
  • Anything you cannot verify. If a lawyer cannot independently confirm an output, that output should not be in a filed document, an opinion letter, or a client deliverable.
  • Direct client advice. ChatGPT does not have the context of the client’s full file, jurisdiction, and risk tolerance. It is fine for a lawyer to use it as a thinking aid; it is not fine to forward its output to a client as legal advice.
  • Anything involving non-anonymized confidential client data in a free or consumer tier. The free tier of ChatGPT may use inputs for training. The Plus, Team, and Enterprise tiers have different retention and training settings. Lawyers must understand which tier they are on and configure data settings accordingly. The Enterprise tier — with a Business Associate or Data Processing Agreement — is the cleanest fit for confidential matter work.

Setting up ChatGPT for a law firm: practical configuration

The minimum configuration any firm using ChatGPT for client work should put in place:

  • Account tier. Use ChatGPT Team or Enterprise for any matter work. These tiers default to no training on inputs and provide tighter retention controls. Solo and small-firm practitioners should at minimum use ChatGPT Plus and turn off “Improve the model for everyone” in Settings → Data Controls.
  • Written firm AI policy. Cover allowed tools, prohibited inputs (raw client identifiers, privileged communications, etc.), required output verification, billing rules, and disclosure rules. The policy is what the supervision rules require — and what carrier underwriters increasingly ask for.
  • Training and onboarding. Every lawyer and staff member who uses the tool gets at least an hour of structured training: what to enter, what not to enter, how to verify outputs, how to log AI-assisted work for billing.
  • Verification protocol. Every citation, factual claim, and substantive legal proposition produced by the tool must be checked against an authoritative source before it leaves the firm. Document the check.
  • Billing rules. If the tool reduced the time required for a task, the bill reflects the reduced time. The lawyer’s value-add — judgment, supervision, finalization — is still billable, but billing 5 hours for 5 minutes of work because the tool produced output is a fee-rule violation.
  • Client disclosure rules. Decide where on the spectrum the firm sits — silent use (low-impact tasks), case-by-case disclosure, or always-disclose-in-engagement-letter. Document the choice and apply it consistently.

Specific prompts that work for legal work

Generic ChatGPT prompts produce generic output. Lawyers who get the most out of these tools tend to use structured, role-specific prompts. A few patterns that consistently outperform the bare “draft me a letter” approach:

  • Role + context + task + constraints. “You are an experienced employment lawyer drafting a demand letter for a client claiming wrongful termination under California FEHA. Facts: [paste anonymized facts]. Draft a demand letter that is firm, factually specific, asks for [demand], and runs about 2 pages. Use plain English and avoid legalese.”
  • Output format specification. “Return the result as a numbered chronology with each entry containing: date, event, source document. Then list any apparent inconsistencies.”
  • Iterative refinement. Draft, critique, redraft. “What’s weak about this argument? What would the strongest counter-argument be? Now redraft to anticipate that counter.”
  • Verification framing. “List five cases that might support [proposition]. Mark each as ‘I’m confident this exists’ or ‘I’m uncertain — check before relying.’ Provide the citation only if you’re confident.” This will not eliminate hallucinations but it improves your odds of the model flagging its own uncertainty.
  • Document upload + targeted question. Inside Plus/Team/Enterprise, upload the document and ask a specific question. “What does Section 4 of this lease say about the landlord’s repair obligations?” Much more reliable than asking the model to recall the same lease from memory.

The legal-AI tools beyond ChatGPT

For citation-grounded research and document work, purpose-built legal AI tools are now the right answer. The current landscape:

  • Harvey. Aimed at large firms; integrates with internal databases and document management. Strong at drafting, document review, and matter-specific research. Significant enterprise pricing.
  • Westlaw AI / CoCounsel. Thomson Reuters’ integration of GenAI with the Westlaw case database. Citations are real because the underlying corpus is real.
  • Lexis+ AI. The LexisNexis equivalent. Same general value proposition.
  • vLex Vincent. Strong international and federal coverage.
  • Eve. Plaintiff-side litigation workflow, particularly employment and personal injury.
  • Spellbook, Robin AI, Ironclad. Contract drafting and review.
  • Everlaw AI Assistant, Relativity aiR, Reveal AI. Document review at scale, integrated with the e-discovery platform you are likely already using.
  • Clio Duo, MyCase IQ, PracticePanther IQ. Practice-management-integrated AI features for billing narratives, intake summaries, and case work.

The choice between general ChatGPT and a purpose-built legal AI tool is not either-or. Most firms doing this well use ChatGPT Team or Enterprise for the long tail of drafting, summarization, brainstorming, and internal-content tasks, and a citation-grounded legal AI tool — Westlaw AI, Lexis+ AI, Harvey, or vLex — for any work product that depends on real, verifiable case law.

Frequently asked questions

Can I lose my law license for using ChatGPT?

Not for using it. For misusing it, yes — bar discipline, malpractice claims, and court sanctions are all on the table. Every reported case of professional discipline tied to GenAI has involved one of three things: filing fabricated citations without checking them, exposing client confidential information through a consumer-tier chatbot, or billing for AI-generated work as if a lawyer had done it manually. None of those are inherent to the technology. All of them are avoidable with a basic firm policy and a verification habit.

Do I have to tell my client I’m using AI?

It depends on the use and the jurisdiction. ABA Formal Opinion 512 says disclosure is required where the use of AI materially affects the matter — which most ethics commentators read to include any use that goes into core work product. State bars vary on whether disclosure must be in writing, in the engagement letter, or can be oral. The conservative answer is to add a short paragraph to the engagement letter describing the firm’s use of AI tools and to document the client’s acknowledgement. That posture survives any plausible reading of the rules across jurisdictions.

What’s the difference between ChatGPT, ChatGPT Plus, Team, and Enterprise for a law firm?

The free tier and Plus tier may use inputs to improve the model unless the user opts out. Team and Enterprise default to no training on inputs and offer SOC 2 reporting, longer chat history controls, and stricter data handling. For any work that touches client confidential information, Team or Enterprise is the right tier. Enterprise additionally supports Data Processing Agreements that satisfy most state bar guidance on confidentiality.

Can ChatGPT replace Westlaw or LexisNexis for legal research?

No. General ChatGPT does not have current, citation-grounded access to case law and routinely fabricates citations that look real. Westlaw AI / CoCounsel, Lexis+ AI, vLex, and similar tools are built on top of authoritative databases and are appropriate research tools. ChatGPT can be useful for brainstorming, structuring research questions, or drafting around research that has already been verified — but it is not a substitute for a real research database.

How do I bill for ChatGPT-assisted work?

Bill the actual time you spent — drafting, reviewing, supervising, finalizing — at your normal rate. Do not bill for hours the AI did the work. Some firms bill the AI tool itself as a cost passed through to the client (similar to LexisNexis charges); others absorb it as overhead. Both are defensible if disclosed in the engagement letter. What is not defensible is invoicing 5 hours for a 5-minute prompt.

Sources

  • American Bar Association Standing Committee on Ethics and Professional Responsibility, Formal Opinion 512: Generative Artificial Intelligence Tools (July 2024). Available at ABA Ethics Opinions.
  • The Florida Bar, Ethics Advisory Opinion 24-1 (Generative AI). Available at Florida Bar Ethics Opinions.
  • State Bar of California, Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law (Nov. 2023). Available at California State Bar COPRAC.
  • New York State Bar Association, Report and Recommendations of the New York State Bar Association Task Force on Artificial Intelligence (April 2024). Available at NYSBA publications.
  • Mata v. Avianca, Inc., 22-cv-1461 (S.D.N.Y. 2023) (sanctions order on fabricated AI-generated citations).
  • U.S. Courts, Standing Orders on Generative AI — most federal districts now have a standing order addressing GenAI use in filings. Check the local rules of any court before filing AI-assisted work.
  • Stanford Center for Legal Informatics (CodeX), Hallucination-Free Legal Research (2024) — empirical study of hallucination rates across legal AI products.
  • Thomson Reuters Institute, Future of Professionals Report (annual) — survey data on AI adoption rates and use cases in U.S. law firms.
  • Model Rules of Professional Conduct, Rules 1.1, 1.4, 1.5, 1.6, 3.3, 5.1, 5.3 — the underlying ethical framework Opinion 512 applies.

This article is general professional-practice information for lawyers. It is not legal advice for your clients and is not ethics advice for your specific bar admission. Consult your jurisdiction’s rules of professional conduct and your firm’s malpractice carrier for binding guidance.