ChatGPT for Legal Research: What It Can and Cannot Do, and How to Use It Safely

You can use ChatGPT for legal research, but not as a source of case law. ChatGPT has no case law database and no citator, so it generates citations that look correct rather than retrieving citations that are correct. Use it for the thinking work: explaining an unfamiliar doctrine, outlining an argument, turning a holding you already have into plain English, or drafting a first pass at a memo. Then find and verify the actual authority in a tool that returns real, retrievable opinions with a still-good-law signal. Every lawyer sanctioned over AI in the last three years skipped that second step.

The question comes up constantly now, and the two loudest answers are both wrong. "Never touch it" ignores that ChatGPT is genuinely useful for a lot of legal work. "It is a research tool now" ignores what the model actually is. The useful answer is narrower and more practical: it depends entirely on which task you hand it.

Can you use ChatGPT for legal research?

Yes for understanding law, no for finding law. ChatGPT is a language model that predicts likely text, not a database that retrieves records. When you ask it for cases supporting a proposition, it produces text shaped like a citation, complete with a plausible case name, a real-looking reporter volume, and a page number. Sometimes that text happens to match a real case, because the case appeared often enough in its training data. Sometimes it does not, and nothing in the output tells you which situation you are in.

That is a structural limit, not a bug that gets patched. A citation is a factual claim about the world: this opinion exists, this court issued it, it appears at this page, and it holds this. A model generating fluent text has no mechanism for checking any of those four claims. Purpose-built legal tools solve it by searching an actual corpus of opinions and handing back what they found, which is a different operation entirely.

Why does ChatGPT make up case citations?

Because a citation is a highly patterned string, and producing patterned strings is exactly what the model is good at. "Smith v. Jones, 412 F.3d 1099 (9th Cir. 2005)" follows rules the model learned thousands of times over. It can generate a flawless one for a case that was never decided, in the same way it can generate a convincing phone number that belongs to nobody. Formatting correctness carries no information about existence.

The failure is worse than pure invention, though, and this is the part people miss. The more dangerous outputs are the ones that are partly true: a real case with a misstated holding, a real holding attached to the wrong case, or a quotation that never appears anywhere in the opinion. A quick existence check catches the fabricated case. It does not catch a real citation with an invented holding bolted onto it, which is why avoiding hallucinated citations means reading the opinion rather than confirming the case is real.

Does ChatGPT Pro or deep research mode fix the citation problem?

It helps and it does not solve it. Browsing and deep-research modes change the operation from "generate plausible text" to "search the web, read pages, then summarize with links," which is a real improvement: the model is now grounding its answer in documents it actually retrieved, and you can click through to check. For a question like "what is the general rule on non-compete enforceability in California," that is a reasonable way to get oriented.

What it still does not give you is the thing legal research actually requires. Web retrieval finds pages that rank well, not the opinions that control your jurisdiction. It has no reliable sense of whether an Appeals Court decision was published or issued as an unpublished summary disposition. Above all, it has no citator, so it cannot tell you that the case it just found was overruled in 2021. A confidently summarized, correctly linked, genuinely real case that is no longer good law will lose your motion just as thoroughly as a fake one. Our explainer on what a citator is covers why that verification layer is a separate product, not a feature of a search box.

What is ChatGPT actually good for in legal work?

Quite a lot, once you stop asking it for authority. The tasks it handles well share a feature: you supply the facts and the law, and it does the language work.

Read down that table and the boundary is obvious. ChatGPT is a strong writing and explanation tool that happens to know a lot about law. It is not a research database, and the sanctions all come from treating it as one.

What are good ChatGPT prompts for legal research?

The prompts that work are the ones that do not ask for authority. Give it the law and ask for language, or ask for orientation you intend to verify anyway:

• "Explain the elements of promissory estoppel in plain English, and list what a plaintiff has to prove." Doctrine, not citations. Verify against a real source before relying on it.
• "Here is the text of an opinion. Summarize the holding in one paragraph and tell me what facts drove the result." You supplied the opinion, so it cannot invent one.
• "Give me ten different ways to phrase a search for whether a non-compete binds an independent contractor." Search-term generation, checked by whether the searches work.
• "Here is my argument section. What is the strongest counterargument opposing counsel will make?" Adversarial thinking, no citation risk.
• "Rewrite this paragraph to be shorter and less hedged, keeping every citation exactly as written." Editing, with the citations explicitly locked.

Notice what is missing: "find me cases that say X." That single prompt is responsible for most of the trouble, and no amount of prompt engineering fixes it, because the model cannot search what it does not have. If you want that question answered, ask it of a tool built on real opinions. Our guide to searching case law with AI covers what that workflow looks like in practice.

How do I check whether a case ChatGPT gave me is real?

Search for the citation in a database of actual opinions, open the case, and read the part you intend to rely on. That is the whole check, and it takes about two minutes per case. If the citation returns nothing, the case does not exist. If it returns a case with a different name or from a different court, the model has blended two sources. If the case exists and says what you were told, you still are not finished: run it through a citator to confirm it has not been reversed, overruled, or superseded, which is the step covered in checking whether a case is still good law.

Do not verify a citation by asking ChatGPT whether the citation is real. It will usually agree with you, and sometimes it will apologize and produce a second fabricated case in place of the first. Verification has to come from outside the model.

Have lawyers been sanctioned for using ChatGPT?

Yes, and the volume has grown faster than most practitioners realize. The case everyone knows is Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023), where two attorneys submitted a brief citing six nonexistent decisions produced by ChatGPT. On June 22, 2023, Judge P. Kevin Castel sanctioned them $5,000 under Rule 11 and ordered them to notify each real judge whose name had been attached to a fabricated opinion.

Mata is no longer an outlier. A public database of AI hallucination cases maintained by legal researcher Damien Charlotin had logged roughly 1,600 court decisions worldwide by June 2026, more than a thousand of them in the United States, up from around 200 a year earlier. Courts have responded with fines, public reprimands, stricken filings, and referrals to disciplinary bodies. What is striking about the pattern is how ordinary the underlying mistake is: in nearly every case the lawyer treated model output as a finished product rather than a lead. We go deeper on the accuracy numbers in our piece on whether you can trust AI for legal research.

The firms that have stayed out of this have mostly done two unglamorous things: they wrote down which tools may be used for which tasks, and they actually trained everyone on the policy instead of circulating it once and assuming it stuck. Verification is a habit, and habits need to be taught.

ChatGPT or a legal research tool?

Both, for different jobs. Keep ChatGPT for explanation, drafting, and thinking out loud. Use a research tool that searches real opinions when you need authority you can put in front of a judge. The practical difference is what comes back: a general chatbot returns prose, while a research tool returns a case, a citation you can check, the holding, and a signal about whether the case still stands.

That is also the honest limit on legal AI generally. A tool that shows you the opinion and lets you verify it is doing something a chatbot cannot, but it does not remove your obligation to read the case. If you are weighing options, our breakdown of legal research software compares the categories, and citation checking explains the verification layer that general chatbots simply do not have.

Casesearch is a legal research tool, not legal advice. Always read the full opinion and verify citations before relying on them.