AI Competency for Lawyers

Tools such as ChatGPT, Claude, Gemini and Copilot are built on large language models (LLMs). An LLM is trained on vast amounts of text to predict which words are most likely to come next. ABA Formal Opinion 512 describes these tools as producing a statistically probable output when prompted.

That design explains both their strength and their main weakness. They are excellent at producing fluent, well-structured language: first drafts, summaries, plain-English explanations. But a model does not consult a database of cases when it answers. A citation has a very predictable shape (party names, a reporter, a volume, a page, a year), so a model can produce something that looks exactly like a real authority with nothing behind it. This is called a hallucination.

Two distinctions matter in practice:

• General assistants answer from patterns learned in training, sometimes supplemented by web search.
• Grounded legal tools retrieve documents from a defined legal database and show you the sources. Grounding reduces errors but does not remove them.

Finally, every model has a context window: the maximum text it can consider at once. Very long documents may be truncated or summarised without a clear warning, which is why you should ask for page or paragraph references and check them.

The quality of AI output depends heavily on the instruction. A useful structure for legal prompts is CRAFT, taught in AI for Lawyers 2025: give the AI context and a role, a precise request, the audience, the format you need and the tone.

In practice, a strong legal prompt includes:

1. Role and jurisdiction. “Act as an experienced employment lawyer advising under New York law.”
2. Task. “Review the non-compete clause below and identify enforceability risks for the employer.”
3. Context. The relevant facts or text, with names and identifying details removed where appropriate.
4. Format. “A table: issue, why it matters, proposed wording.”
5. Guardrails. “Do not invent citations. Flag uncertainty. Do not assume facts.”

Treat the first answer as a draft. Ask follow-up questions, request alternatives, and tell the tool what was wrong. Try our free Legal Prompt Builder to practise.

In Mata v. Avianca (S.D.N.Y. 2023), lawyers filed a brief citing cases that did not exist and were sanctioned $5,000. Since then, courts have handled AI-fabricated citations in large numbers, and penalties have escalated to six-figure awards and, in April 2026, a licence suspension in Nebraska. In March 2026 the Sixth Circuit stated in Whiting v. City of Athens that no filing should contain citations the lawyer has not personally read and verified.

A practical routine before anything AI-assisted is filed or sent:

1. List every authority and quotation.
2. Find each one in a trusted database.
3. Read the passage relied on and confirm it supports the proposition.
4. Match quotations word for word.
5. Check subsequent history.
6. Record how AI was used and comply with any court rule.
7. Sign off personally.

Asking the same chatbot “is this case real?” is not verification; it may confirm its own invention. Use our AI Citation Checker to build your list.

Your duty of confidentiality (Model Rule 1.6 and its state equivalents) applies fully to AI tools. Opinion 512 explains that the risk depends on the client, the matter, the task and the tool. It warns that self-learning tools, which may use inputs to improve, create a risk that confidential information could surface in later outputs, and indicates that informed client consent is needed before inputting confidential information into such tools. Boilerplate in an engagement letter is not enough for that consent.

Practical safeguards:

• Use business or enterprise plans whose terms prohibit training on your data, not personal or free accounts.
• Read the tool’s terms of use, privacy policy and data-retention settings, or have someone qualified review them.
• Input only what the task requires, and remove names and identifying details where possible (try the Confidentiality Redactor).
• Follow your firm’s AI policy and approved-tools list.

From January 1, 2027, California’s SB 574 adds a statutory limit for California attorneys on entering confidential information into AI systems that are not restricted to the attorney and others bound to protect it.

On July 29, 2024, the ABA issued Formal Opinion 512, its first formal opinion on generative AI. It applies existing Model Rules across six areas: competence (understand the tool and verify output to an appropriate degree), confidentiality, communication with clients, candor toward the tribunal and meritorious claims, supervision of lawyers, staff and vendors, and reasonable fees.

States have followed. Many bars have issued formal opinions or guidance. In January 2026, Colorado became the first state to amend its Rules of Professional Conduct for AI. New York’s court system adopted Part 161, effective June 1, 2026, which permits AI in court papers but requires attorneys to ensure there is no fabricated material. California enacted SB 574, effective January 1, 2027, the first statute governing lawyers’ use of generative AI.

Individual judges also issue standing orders on AI use and disclosure. Check them for every court you appear in. See what applies in your state with our State Bar AI Guidance Navigator.

Opinion 512 applies Model Rule 1.5 to AI. Lawyers billing hourly must bill for their actual time, which includes time spent prompting, reviewing and verifying, but not time saved. A flat fee may become unreasonable if AI makes the work much faster than the fee assumed. Lawyers generally should not charge clients for learning to use a tool they will use across their practice, and the cost of general-purpose tools built into firm software is overhead. Costs of a third-party tool used for a particular matter may be charged where reasonable and agreed.

On communication, Opinion 512 indicates that lawyers should tell clients about AI use where the client asks, where engagement terms or client instructions require it, where informed consent is needed (for example for some confidential inputs), and where AI output will influence a significant decision in the representation. Many firms now describe their AI use in engagement letters; our Disclosure Clause Generator helps you draft that language.

Individual care matters, but firms need systems. Opinion 512 indicates that lawyers with managerial authority should establish clear policies on permissible AI use and ensure people are trained. A sensible programme:

1. Policy. A short written AI use policy covering approved tools, confidentiality, verification, filings, billing and incidents. Start with our AI Use Policy Generator.
2. Approved tools. Chosen by task and risk, with vendor terms reviewed. Use the Vendor Due Diligence Builder.
3. Pilot. Test tools on recently completed matters where you already know the right answer.
4. Train. On real firm tasks, including prompting and verification.
5. Supervise and measure. Track time saved, errors caught and client feedback, and review the policy regularly.

Take the Firm AI Readiness Assessment to see where to start.