Where ChatGPT actually helps with legal research
Most of the trouble lawyers get into with ChatGPT comes from one mistake: asking it to find case law. ChatGPT is a language model, not a legal database. It does not have a reliable index of decisions, and when pressed for citations it will invent them with the same calm confidence it uses for the right answer. That is the Mata v. Avianca pattern, and the sanctions have kept coming since.
The fix is not to swear off ChatGPT for research. The fix is to use it for the parts of legal research where it actually works, and to use a citation-grounded product for the case lookup. ChatGPT is strong at framing a question, summarizing cases you give it, drafting a research memo from verified sources, and pressure-testing a position. It is unreliable at finding cases, confirming current law, and naming the right statute on its own.
This guide is for the solo or small-firm lawyer who already pays for ChatGPT (or Claude, or Gemini, the workflow is the same) and wants to get research value out of it without ending up in a sanctions order. We will walk the workflow, show the exact prompts, and end with a list of things to never ask the model to do. For the broader picture of where AI fits inside a small practice, see the pillar guide on AI for law firms.
The five-step workflow that actually works
The output of legal research is a memo, a brief section, or a focused advice email. Treat ChatGPT as the assistant that compresses every step except the case lookup. The cases come from a real database. Everything around the cases can run through the chat window.
1. Frame the question with ChatGPT before you search
The biggest single time savings I see at small firms is in the framing step. A vague question runs slow and returns noise. A precise legal question, broken into sub-issues with the controlling rule called out, returns research you can use.
Run this prompt before you touch a research database:
Act as a senior legal researcher. I am preparing a memo on the following question: [ROUGH QUESTION IN PLAIN ENGLISH] The jurisdiction is [STATE OR FEDERAL CIRCUIT]. The relevant facts are: [3-5 RELEVANT FACTS] Reformulate this as a precise legal question or a set of sub-questions a court would actually decide. Identify the controlling statute or rule if there is one. List the elements I will need to address. Do not cite cases. I will pull case law myself.
The “do not cite cases” line matters. Without it the model will offer up cases that may or may not exist. Strip that path off entirely. The output of this step is a research plan, not a legal authority.
2. Use a citation-grounded tool for the case lookup
Once you have a clean question, switch to a tool that searches a real case database. The small-firm-relevant options in 2026: Lexis+ with Protégé (the product formerly known as Lexis+ AI; renamed in February 2026), Westlaw Precision AI, and CoCounsel (Thomson Reuters; absorbed Casetext in 2023). All three search real case data and ground their answers in actual citations.
If your firm has neither Lexis nor Westlaw, the citation-grounded option still has to come from somewhere. Free Google Scholar plus a careful read works for many practice areas; some bar associations also offer Fastcase or vLex access included in dues. Whatever the source, the case lookup never runs through ChatGPT alone. For a deeper comparison of the legal-database AI tier, see AI legal research tools.
3. Verify every case before you read further
Even the citation-grounded tools can return a real case with a wrong summary. Take five minutes per case before relying on it.
- Pull the case directly from the database, not from the AI summary’s link.
- Read the procedural posture and the headnote that matches your issue.
- Shepardize or KeyCite for negative treatment.
If a case does not check out, drop it. A research memo with three solid cases beats one with eight you have not verified.
4. Drop verified cases into ChatGPT for synthesis
This is where ChatGPT earns its place. Once you have a verified short list of cases, paste them into ChatGPT (full text where you have it, headnotes plus relevant excerpts where you do not) and ask for the synthesis:
Act as a legal research associate. Draft a research memo on the following question: [QUESTION FROM STEP 1] Use ONLY the cases I provide below. Do not introduce new citations. If a case I provide does not address the question, say so explicitly. For each case, identify: (1) the holding relevant to the question, (2) the procedural posture, (3) any limiting language. Then synthesize the cases into a position the firm can take, noting any conflicts in the authority. Output structure: 1. Question presented 2. Short answer 3. Discussion (case by case, then synthesis) 4. Open issues that need more research Cases: [PASTE CASE TEXT OR HEADNOTES PLUS RELEVANT EXCERPTS]
The “use ONLY the cases I provide” line is the guardrail. The model will sometimes still try to volunteer additional citations. When that happens, push back: “Remove any citation I did not provide.” The first draft of the memo is rarely the final draft, but it gets you most of the way in about ten minutes.
5. Pressure-test the position before you send
The final step is the one most lawyers skip. Once you have a draft memo, paste it back into ChatGPT and ask for the opposing argument:
You are opposing counsel preparing to attack the position in the memo below. Identify: 1. The strongest counterargument 2. Any case in the memo that could be distinguished on its facts 3. Any controlling authority I may have missed (note the issue, not a specific citation; I will look it up myself) 4. Any factual assumption in the memo that, if wrong, would change the answer Memo: [PASTE YOUR DRAFT MEMO]
This is the highest-leverage prompt in the whole workflow. It costs five minutes and frequently surfaces a weakness in the position before a partner or a judge does.
Subscription tier and confidentiality
ChatGPT prompts are not privileged. Free-tier and Plus inputs may be used to train future models unless you turn off training in settings, and even then OpenAI retains conversations for a period for safety review. For client-related research, the right tier is ChatGPT Team or Enterprise, which contractually excludes inputs from training and provides data residency and SOC 2 Type 2 controls. Claude for Work and Microsoft Copilot for Business sit at similar tiers.
Before pasting any client-identifying material into a chat window, confirm three things: the tier excludes your inputs from training, the firm has a data processing agreement with the vendor, and your engagement letter does not bar third-party AI processing. The American Bar Association’s Formal Opinion 512 sets the framework: competence, confidentiality, candor, supervision, and reasonable fees. None of those rules ban ChatGPT, but each of them constrains how you use it. Confirm with your bar before deciding what your firm’s posture should be.
Tools to consider for the synthesis step
The synthesis tool can be any general-purpose model. Three real options for small firms in 2026:
- ChatGPT (OpenAI). Plus is $20 per user per month; Team starts around $30 per user per month with admin controls and training opt-out by default.
- Claude (Anthropic). Pro is $20 per user per month; Claude for Work starts around $30 per user per month. Claude has a longer context window than ChatGPT Plus, which matters when you are pasting full case text.
- Gemini (Google). Bundled into Google Workspace tiers your firm may already pay for. Worth a look if you have not already standardized on one of the others.
Pick one and standardize the firm on it. Switching back and forth defeats the muscle memory that makes any of this fast.
What to watch for
Hallucinated citations
The Mata v. Avianca order from 2023 is the canonical warning, but the sanctions did not stop. The Second Circuit referred a lawyer for discipline in Park v. Kim in 2024 over fabricated citations in an appellate brief, and trial-court sanctions for invented cases have continued through 2025 and into 2026. Any time you ask a general-purpose model for a case, statute, regulation, or quotation, treat it as a tip you have to verify, not a result. The “do not cite cases” line in the framing prompt and the “use ONLY the cases I provide” line in the synthesis prompt exist to keep the model out of citation-generation mode.
Stale law
ChatGPT’s training data has a cutoff. Anything that changed close to or after that cutoff (new statutes, recent appellate decisions, regulatory updates) is the model’s weakest area. Web-search modes help, but they are not a substitute for a current case database. For anything that turns on recent law, run the case lookup through Lexis or Westlaw and treat the model output as commentary.
Confidentiality slips
The “I will just paste the deposition transcript into ChatGPT to summarize” reflex is where most confidentiality issues start. (See AI deposition summary for the same workflow applied to transcripts.) Default redactions before any paste: client names, opposing parties, third-party witnesses, account numbers, tax IDs, addresses, and medical record numbers. If the analysis truly needs the names, run it on an enterprise tier under a data processing agreement, not on a personal account.
Overconfident summaries
Models present everything in the same well-formatted voice. A correct summary and a wrong one read identical. Treat the polish as a styling decision, not a signal of accuracy. Verification is the only protection.
FAQ
Can I just ask ChatGPT to write a research memo from scratch?
You can, and the result will be authoritative-sounding text with a real risk of fabricated cases. Use the workflow above instead: framing through ChatGPT, case lookup through a citation-grounded tool, synthesis back through ChatGPT with verified cases pasted in.
Is ChatGPT enough, or do I still need Lexis or Westlaw?
For framing, synthesis, and pressure-testing: ChatGPT is enough. For case lookup and Shepardizing or KeyCite: no. The two categories of tool do different jobs. If your practice involves frequent novel research, the Lexis or Westlaw subscription is the line item that actually protects the firm. For a vendor-by-vendor comparison, see Best AI legal research tools.
What about Claude or Gemini? Are they better for legal research?
Different strengths, same category. Claude tends to follow the “do not invent citations” instruction more reliably in my testing. Gemini’s web grounding is stronger when you need recent commentary. Neither replaces a citation-grounded research product. Pick one general-purpose model, standardize on it, and learn its quirks.
What is the single biggest mistake to avoid?
Asking the model for case citations. Every other failure mode is downstream of that one.
Do I need to disclose AI use to clients or courts?
Court rules vary by jurisdiction and judge. A handful of federal judges have standing orders requiring disclosure of generative AI use in filings. State bar opinions are inconsistent. Confirm local rules and your bar’s guidance before each filing where AI played a role beyond clerical work. ABA Formal Opinion 512 leans toward disclosure when AI materially affected the work product.
Related on Business AI Workflows
- AI for law firms. The pillar guide covering workflows, risks, and a 30-day adoption plan.
- AI legal research tools. The citation-grounded category that pairs with ChatGPT in this workflow.
- AI deposition summary. Applying the same paste-and-prompt pattern to deposition transcripts.
- Best AI legal research tools. Vendor-by-vendor comparison of the citation-grounded research products.
