ChatGPT for Lawyers, Tested: What Works, Where It Fails

Our tests support specific drafting workflows, not reliance on ChatGPT as legal authority. We tested GPT-6 Astra through an API harness, not the consumer app. For broader tool selection, see our guide for lawyers.

Can ChatGPT build a research plan without inventing citations?

It can produce an issue outline and research plan when instructed to omit citations, as our tested prompt demonstrated. Treat the result as a search agenda, with every legal proposition still to verify.

The table reports average criteria passed out of 5, followed by runs meeting every criterion out of 3. These are task results, not an overall model ranking.

Test date: 2026-10-19. Models called through the API, not consumer apps.

Claude Fable 5.1 and Gemini 4 Argon, added on 2026-10-20 with the same prompts and rubrics and judged by GPT-6 Astra, scored 4.33 of 5 (1 of 3 runs) and 5 of 5 (3 of 3 runs) on the research plan, 4.33 of 5 (1 of 3 runs) and 4.67 of 5 (2 of 3 runs) on the NDA, and 4 of 5 (0 of 3 runs) each on intake. Gemini ran through Google’s Antigravity CLI with a tool-less agent, not the Gemini app, so app results may differ there too.

Both caveats matter. The command-line API harness had tools, web and memory off, so app results may differ. GPT-6 Astra graded the Claude and Gemini outputs, while Claude Opus 5.5 graded GPT-6 Astra outputs, so GPT’s raw scores are not directly comparable with the others; the judges were not equally strict: on 30 sampled Claude runs, GPT-6 Astra passed 93% of criteria while the swap judge, Claude Opus 5.5, passed 98%; both judges passed 97% on 30 sampled GPT runs. On 10 sampled runs each, GPT-6 Astra passed 88% of criteria for Claude Fable 5.1 (swap judge 100%) and 92% for Gemini 4 Argon (swap judge 98%). Part of the gap comes from the grader. See the methodology for the design.

The research workflow helps separate issues, missing facts and search terms. The test does not establish that its proposed elements are current law in your jurisdiction. ABA Opinion 512 requires review of outputs, including citations, before court submission under Rules 3.1 and 3.3; footnote 47 directs lawyers to check local court rules on AI. California’s guidance likewise says candor cannot be delegated to AI.

Adapted from a tested prompt:

You assist a US litigation associate starting legal research.
Using the material below, produce:
- Legal issues framed as questions.
- Elements or factors, as generally understood.
- Facts still needed from the client.
- Primary sources and databases to search, with search terms.
Do not name any case, statute section or citation.
Where authority is needed, write [AUTHORITY TO FIND].
State that every point requires verification in primary sources
by a licensed attorney.
Material:
[FICTIONAL FACT PATTERN AND JURISDICTION]

Start with fictional material. Removing names from a real matter does not settle the confidentiality assessment.

How well does ChatGPT redline an NDA?

Our test supports using it for a playbook-based issue list with proposed replacement language. It does not establish reliability for unrestricted contract review or a finished redline.

The useful first pass is a table connecting each clause to your client’s stated position, with missing provisions listed separately. Counsel must decide whether each proposed change fits the transaction and works with the rest of the agreement. Do not treat a generated risk rating as a legal conclusion.

The test did not measure performance without a playbook. Supply the negotiating positions rather than asking the model to decide what is acceptable. California’s guidance requires the lawyer to retain control of strategic decisions and to review, verify and correct outputs.

Adapted from a tested prompt:

You assist counsel for the party identified in the playbook.
Review the NDA clauses against that playbook.
Output a table:
Clause | Issue | Risk (High/Med/Low) | Proposed replacement language
Then list provisions missing from the NDA that the playbook requires.
Keep replacement language concise and consistent with the playbook.
Do not assume facts. Flag decisions for the supervising attorney.
Material:
[FICTIONAL NDA CLAUSES AND PLAYBOOK POSITIONS]

Before using an actual NDA, assess the information it reveals about the representation, including parties and deal terms. Apply the confidentiality checks below. If you also use Claude, consult the NDA review skill as a separate workflow resource.

Can ChatGPT help with intake and client letters?

It can organize intake material and draft correspondence for attorney review; only the intake workflow has supplied test results here. A generated conflicts list is preparation for a conflicts check, not clearance to accept a matter.

For intake, ask for named people and entities, their roles, dates, missing documents and an acknowledgment draft. Compare the extracted list with the source message, then use the firm’s conflicts process. The prompt cannot establish that the prospect supplied every relevant name.

Deadline flags also require attorney review. The tested prompt expressly avoids stating a specific limitations period as fact. Review the acknowledgment’s wording rather than assuming a prompt instruction determines its legal effect.

Adapted from a tested prompt:

You assist a US law firm's intake team. From the message, produce:
- A neutral matter summary, max 120 words.
- Every named person and entity, with their role, for conflicts checking.
- Key dates and potential deadlines requiring urgent attorney review.
  Do not state a specific limitations period as fact.
- Documents to request.
- A short acknowledgment draft that gives no legal advice,
  avoids creating an attorney-client relationship and states
  that the firm has not agreed to represent the prospect.
Message:
[FICTIONAL PROSPECTIVE-CLIENT MESSAGE]

Do not assume prospective-client information is safe to paste. Assess applicable confidentiality duties and the tool’s terms before using an actual intake message.

For client letters, a proposed use is turning attorney-approved notes into plain-English status correspondence. We did not score this task. Check for unsupported conclusions, dates, promises and omissions rather than relying on fluent wording.

Untested client-letter prompt:

Draft client correspondence using only the attorney-approved notes.
Explain the matter's status in plain English.
List next steps, information needed from the client and supplied dates.
Do not add legal conclusions, deadlines or predicted outcomes.
Where information is missing, write [ATTORNEY TO CONFIRM].
Do not promise a result. Flag ambiguities separately for counsel.
Tone: [DESIRED TONE]
Notes:
[FICTIONAL STATUS NOTES, NEXT STEPS AND DATES]

Review the letter against the notes before sending it. ABA Rule 1.4 may require communicating AI use even where Rule 1.6 does not; California’s guidance calls for considering disclosure of intended use, method, benefits and risks. Find additional drafting resources in our prompt library.

Which ChatGPT plan should a law firm choose?

Choose only after checking the plan’s current terms against the intended work; the verified prices alone do not establish which plan suits a firm. Free provides a no-cost starting point for fictional practice exercises, but we have not verified a feature or usage-limit comparison here.

Prices and plans checked 2026-10-19 on OpenAI’s official pages. USD list prices for a US visitor, excluding tax.

OpenAI’s system card dated 2026-10-07 announces the October GPT-6 Sol and GPT-6 Luna Chat models across free and paid plans, replacing older GPT-5.6 Sol and GPT-5.6 Luna. Which model each plan uses by default is not verified. The tested API model, GPT-6 Astra, is distinct from those Chat models.

GPT-6.1 Sol is in Work and Codex only, not Chat, and is rolling out to Plus, Pro, Business, Enterprise and Edu. Do not assume a subscription reproduces our API test setup.

For readers in the UK, Australia and India, local prices, taxes and professional rules differ: check the official country pricing page and local guidance. Our ChatGPT overview for professionals provides a separate product reference.

We have not verified either vendor’s data terms for this page, including retention, training on inputs or business protections. A paid plan’s name is no basis for assuming confidentiality safeguards.

What does ABA Opinion 512 require before entering client information?

Before entering information relating to a representation, evaluate whether others inside or outside the firm could access it. For self-learning tools, obtain informed client consent first.

ABA Formal Opinion 512, dated July 29, 2024, grounds this in Rule 1.6. It identifies the risk that one client’s information could surface in another matter and says general boilerplate engagement-letter provisions do not provide sufficient informed consent.

Use this review sequence:

• Read the actual plan’s Terms of Use, privacy policy and related contracts, or consult someone who has. Rule 1.1 requires a reasonable understanding of capabilities and limitations, not technical expertise.
• Identify whether the proposed input contains information relating to a representation. Idea generation without such information does not require informed client consent under the opinion.
• Assess disclosure risks and whether the tool is self-learning. Where consent is required, explain the proposed use and risks sufficiently for an informed decision.
• Consider communication separately. Rule 1.4 may require disclosure even where Rule 1.6 does not. California’s Rules 1.2 and 1.4 guidance calls for considering disclosure of intent, method, benefits and risks.
• Set firm policies and supervise lawyers and nonlawyers. ABA Rules 5.1 and 5.3 address supervision and training; California’s guidance requires policies and periodic training. Its 2026 update addresses agentic AI at the California Supreme Court’s request: substantive legal determinations and filing pleadings require meaningful lawyer supervision.
• Check billing. Under ABA Rule 1.5, keeping the same flat fee after substantial AI time savings may be unreasonable. General office overhead is not billable, and an AI expense cannot be surcharged without disclosure. California requires informed written consent for AI-cost markups; hourly bills must reflect actual time, including prompting and review.

Limits of this guide: “Tested” means at least 2 completed, graded runs, not professional sign-off. Scores are machine-graded API results subject to judge asymmetry, not app-performance guarantees. A licensed attorney in your jurisdiction must verify applicable law, citations, deadlines, conflicts, contract language, correspondence, local court rules and professional duties. We have not verified vendor data terms. This is not legal advice or a promise of compliance.

FAQ