Can You Use ChatGPT for Legal Work?

Can You Use ChatGPT for Legal Work?

Every few months a story goes round that ChatGPT has stopped answering legal questions.

It gets shared, OpenAI denies it, and the underlying question never actually gets answered: what do the rules say, and what do they require of you if the work you are doing is legal work?

The answer is in OpenAI's own documents, and it is more specific than the headlines and more demanding than most people assume. It is also notable for what it leaves out.

Can You Use ChatGPT for Legal Work? The Short Answer

Yes. Nothing in OpenAI's policies prohibits using ChatGPT on legal matters, and the idea that the product refuses legal questions is not what the policies say.

What they do say is narrower and more interesting. OpenAI's usage policies restrict "tailored advice that requires a license, without appropriate involvement by a licensed professional." The restriction is not on the subject matter.

It is on handing out licensed advice with no licensed person in the loop. Read carefully, that is a rule about who is involved, not about what you may ask.

There is a second clause that matters just as much. The policies also prohibit "automation of high-stakes actions in sensitive areas without human review," and legal is one of the named sensitive areas.

High-stakes actions are defined as those that materially affect a person's rights, safety, access to essential services, or other consequential interests. If you are building something that acts rather than drafts, that clause is the one that governs it.

So a lawyer using ChatGPT to draft, summarize or think out loud is squarely inside the rules. A product that answers the public's legal questions without a lawyer attached is not.

Most real-world use sits comfortably in the first category, which is why the panic is misplaced and the detail still matters.

What OpenAI's Policies Actually Say

The most useful thing about reading the source rather than the coverage is noticing the vocabulary. OpenAI's usage policies do not contain the phrase "legal advice" at all.

The entire restriction rests on the words "tailored," "requires a license" and "appropriate involvement."

The Licensing Clause

The clause is about professional gatekeeping rather than about topics. Asking what a clause means, how a doctrine works or how a filing is structured is not the restricted activity.

Offering an output to a third party as if it were advice from a licensed professional is. The policy adds a related line that is easy to miss: you may not state or imply that the AI itself holds professional credentials or authorization.

That second sentence is the one with practical teeth for anyone publishing a tool, a chatbot or a client-facing workflow.

The model cannot be presented as the professional. Someone with a license has to be meaningfully in the chain, and the output cannot pretend otherwise.

The Human-Review Clause

The automation clause is the sharper of the two, because it does not care how good the output is. It cares whether a human looked.

A workflow that reviews, classifies or decides on something consequential and then acts on its own conclusion is the thing being prohibited, and a legal matter is explicitly within scope.

One date is worth knowing. The version of the usage policies OpenAI currently publishes carries an effective date of November 20, 2026, with a changelog entry dated October 21, 2026 noting that it clarifies requirements for high-stakes uses.

The policy that governs you is the one in force when you act, so this is a page to read rather than remember.

The Vertical OpenAI Has Not Built a Product For

Here is the part no summary of this topic mentions, and it says more than the policies do.

OpenAI's service terms, updated October 16, 2026, address healthcare and finance by name. There is a section for ChatGPT for Healthcare, which tells customers to "always verify the information provided by ChatGPT for Healthcare and exercise independent professional judgment" and notes that a Business Associate and Healthcare Addendum applies.

There is a separate medical-use section stating that "Our Services are not intended for use in the diagnosis or treatment of any health condition." And there is a financial section stating plainly that "OpenAI does not provide financial or investment advice through Financial Services."

Three named regulated verticals. Two product-specific regimes, one explicit disclaimer, one contractual addendum.

And for law: nothing. No ChatGPT for Legal, no legal addendum, no disclaimer, no section.

That absence is the most honest answer to the original question. In healthcare, OpenAI has built the contractual apparatus a regulated buyer needs and told you where the lines are.

In law, it has left the question entirely to the professional rules that already govern you, which is exactly what its own preamble says it is doing: "Our rules are no substitute for legal requirements, professional duties, or ethical obligations that should influence how people use AI."

How to Use ChatGPT for Legal Work Without Creating a Problem

The practical approach follows directly from those two clauses.

1. Keep a licensed person in the loop on anything that leaves your hands. This is the policy requirement and it is also the only defensible posture. Treat every output as a draft prepared by a capable, unqualified assistant.
2. Verify every citation, authority and quotation independently. Not because the model is unusually bad at it, but because a fabricated authority is the one error that is both easy to produce and catastrophic to file.
3. Decide what you will and will not type in before you start. Client identities, matter specifics and privileged material are a policy decision for the firm, not a judgment call made at the keyboard at 7pm.
4. Match the plan to the sensitivity. The data commitments differ materially between consumer and business tiers, and that difference is the main reason to pay.
5. Do not automate the decision. Drafting, summarizing, comparing and outlining are all fine. Anything that reaches a conclusion about a person's rights and then acts on it needs a reviewer by policy, and needs one by professional duty regardless.
6. Write it down. A short internal note on approved uses, prohibited inputs and the review step is what turns all of the above from good intentions into something you can show a regulator.

Applying the same discipline to how your firm gets found online? Get a free audit and see where the gaps are.

What Happens to What You Type In

For legal work this is the question that decides which plan you can use, and OpenAI answers it most clearly on its enterprise privacy page.

The headline commitment is that OpenAI does not train on business data, and that the models do not learn from that usage.

The same page says customers control how long data is retained on ChatGPT Enterprise, that OpenAI has been audited for SOC 2 compliance, and that data is encrypted at rest with AES-256 and in transit with TLS 1.2 or above. Those are the claims a procurement questionnaire is asking about, and they are stated in the vendor's own words.

Two things are worth noting about what the page does not say. It does not mention a zero-data-retention option, and it does not offer a legal equivalent of the healthcare addendum.

If your obligations require either, that is a contract conversation rather than something you can establish from a published page.

There is also an internal exposure that has nothing to do with OpenAI. The service terms note that Enterprise and Business administrators may be able to access, share and remove content, and can access logs.

Within a firm, that means the administrator is a person who can read the chat. For privileged material, who holds that role and under what policy is a genuine question, and it is one firms routinely forget to ask because they are busy worrying about the vendor instead.

Want the same clarity about where your organic traffic comes from? Request a free audit.

Where It Creates Work Instead of Saving It

The failure modes cluster in a few predictable places.

The first is confident wrongness on anything jurisdictional. Legal answers depend on where you are, which court, which year and which amendment, and a fluent answer that silently assumes the wrong jurisdiction looks identical to a correct one.

The further a question sits from a well-documented federal position, the more carefully the output needs checking.

The second is the review-time trap. If verifying the output takes as long as drafting it would have, the tool has not saved you anything and has quietly added a risk. This is worth measuring honestly per task type rather than assuming.

The tasks where the economics work tend to be the dull, high-volume, low-stakes ones: summarizing long documents you will read anyway, converting notes into structure, producing first drafts of routine correspondence, and explaining an unfamiliar area well enough that you know what to go and read.

The third is the billing question, and it is the one most likely to be decided by someone other than you.

Whether time saved can be billed, and whether the cost of the tool can be passed to a client, is a professional-conduct matter governed by your own bar or regulator rather than by anything in OpenAI's documents. Settle it with the people whose rules actually bind you before the first invoice, not after.

Spending on tools and unsure which ones are earning their keep? Start with a free audit.

ChatGPT vs the Alternatives for Legal Work

Three alternatives are worth weighing, and the general-purpose assistant does not win all of them.

Purpose-built legal AI. Tools built for contract review, diligence and litigation research are trained and evaluated on the task, cite into real authority databases, and come with the contractual terms a regulated buyer expects. They cost more and do less outside their lane. For the core billable work, that trade is usually the right one.

The incumbent research platforms. The established legal research providers have added generative features on top of authority databases they already maintain. The advantage is not the model, it is that the citations resolve to something real. For anything that will be filed, that difference is the whole argument.

Another general assistant. The competing general-purpose assistants are broadly comparable on drafting quality, and their published commitments on training, retention and enterprise administration differ in detail. If your decision is being made on data handling rather than output quality, read all of their terms rather than assuming the category behaves uniformly.

The sensible pattern for most firms is both rather than either: a general assistant for the thinking, drafting and summarizing that surrounds the work, and a specialist tool for the work itself.

What ChatGPT is genuinely good at here is the sprawling, unbillable, undifferentiated writing that fills a practice. What it should not be is the thing that decides.

Ready to find out what your firm's search presence is actually worth? Get a free audit.

Final Thoughts

You can use ChatGPT for legal work, and OpenAI's rules are not the obstacle people think.

They ask for a licensed professional to be appropriately involved, and for a human to review anything high-stakes and automated. Both are things a competent practice was going to do anyway.

The more telling fact is the one found by absence. OpenAI has built a named product, a disclaimer and a contractual addendum for healthcare, and an explicit disclaimer for finance, and has written nothing at all about law.

The duties in this vertical are being left where they already live, with you and your regulator, and no vendor page is going to discharge them.

If the data-handling question is what is actually holding you up, our guide to whether ChatGPT is safe to use goes through the retention and access detail in full.

Read it alongside your own professional rules, decide what may be typed in, write the decision down, and then get on with the work.