Can AI Draft Legal Documents? An Honest Answer by Document Type (2026)

Yes, AI can draft legal documents, and for several document types the first draft is already good enough to save real hours. But "draft" and "file" are different verbs. Whether an AI first draft is usable depends almost entirely on the document type and on one discipline: you verify every authority before your name goes on it.

The useful answer is not "yes" or "no." It is "which document, and did you check the law?" A demand letter and an appellate brief are not the same risk. Sort the work that way and AI becomes a genuine accelerant instead of a sanctions headline.

The short answer, by verb

AI is good at drafting and bad at deciding. It assembles language fast and mirrors a format it has seen thousands of times, but it cannot tell you whether the case it just cited is real, still good law, or on point. A general-purpose model produces citations the same way it produces prose: by predicting plausible text, not by looking anything up.

So AI drafts; the lawyer verifies, edits, and signs. The output is a first draft from a very fast junior associate who is confident, tireless, and occasionally makes things up. You would never file that associate's memo without reading it. Same rule here.

Where an AI first draft is already good

For documents driven by facts and structure rather than controlling authority, an AI first draft can carry most of the load:

• Internal research memos. AI is strong at organizing an issue and producing a readable first pass. The structure of a good legal memo is exactly the pattern it handles well, as long as you supply the facts and verify the law it pulls in.
• Demand letters. Tone, chronology, and the ask are mostly a function of your file. AI drafts a clean version fast; you confirm the facts and check any legal claims.
• Engagement and retainer letters. Scope, fees, and standard terms are near-templated work. AI gives you a solid base to adapt to the matter and your state's rules.
• Routine and procedural motions. Extensions, continuances, and unopposed procedural motions are formulaic. The draft is usually 80% there; the review is local rules and any cited authority.
• Discovery shells. Interrogatory and request-for-production frameworks come out well, then get tailored to the actual claims and defenses.

The common thread: the risk in these documents is wording and completeness, not fabricated law, which makes the review fast and the time savings real.

Where an AI first draft is dangerous

The danger rises the moment a document turns on authority you have to get right:

• Anything with citations. The instant a draft cites cases, the failure mode changes from "awkward sentence" to "hallucinated precedent." This is the use that produced every sanctions order.
• Jurisdiction-specific arguments built from the model's memory. Ask a general model for the standard in your state and it will answer with total confidence and no idea whether it is current, or whether it is quietly describing another jurisdiction's rule.
• Anything signed under Rule 11. Federal Rule of Civil Procedure 11 certifies that, after an inquiry reasonable under the circumstances, your legal contentions are warranted by existing law. An unread AI draft is the opposite of a reasonable inquiry, and the sanctions Rule 11 authorizes fall on whoever signs.

None of this means "don't use AI here." It means the draft is only a starting point, and verification is not optional.

The document-by-document breakdown

Readiness is about how much you can trust the first draft. Every row still ends with a lawyer.

Generation from memory versus drafting grounded in your sources

This distinction separates a helpful draft from a career risk, and most coverage skips it.

Generation from memory is what a general chatbot does. You ask for a motion, and it writes citations from statistical patterns in its training data. Nothing is looked up. A citation to a case that does not exist reads exactly like one to a case that does, because to the model there is no difference. That is not a bug you can prompt away; it is how the technology works.

Retrieval-grounded drafting flips the order. The tool searches a real database first, pulls the actual statutes and opinions, and drafts only from what it found, with a link back to each source. If it cannot find authority for a point, it says so instead of inventing one.

This is the design choice behind CaseRead. It searches your firm's own case files and the public law together, across all 53 US jurisdictions, and every citation in an answer traces to a source the system actually retrieved. If a citation cannot be verified, it is flagged rather than asserted. That does not remove your duty to check; it means you are checking real sources instead of hunting for phantoms. When choosing a tool, "does it retrieve or does it generate?" is the first question to ask — the retrieval-versus-generation distinction the whole legal-AI vocabulary turns on — and our guide to the best AI legal research tools walks through how to tell.

The review workflow: you are the editor of record

Treat every AI draft as something to edit, not something to accept. A workflow that holds up:

1. Ground the draft in your matter, not the open internet. Give the tool the actual documents, the record, the operative facts. A draft built from your file is far easier to trust than one improvised from nothing. If you are tempted to paste client material into a consumer chatbot, read why that is a privilege problem first.
2. Verify every authority. Open each cited case and confirm it exists, says what the draft claims, and is still good law. A structured citation-verification workflow makes this a fast, repeatable habit instead of a scramble.
3. Edit as a lawyer, not a proofreader. Confirm the argument is actually yours, the facts are right, and nothing was smoothed over with a confident sentence no source supports.
4. Sign only what you would defend. Your signature is the certification. If you could not stand behind a citation in front of the judge, it does not go in the document.

The AI compresses steps one and three. It cannot do step two for you, and it never does step four.

What the sanctions cases actually teach

Two 2026-era orders make the stakes concrete, and both point at the same lesson.

In Garner v. Kadince, 2025 UT App 80, the Utah Court of Appeals issued what is widely described as the first published appellate opinion sanctioning AI-fabricated citations. Counsel filed a petition citing a case, Royer v. Nelson, that existed nowhere but in a chatbot's output. The attorney was ordered to pay the opposing party's fees, refund the client, and donate to a legal-aid nonprofit, and the court wrote that it "takes the submission of fake precedent seriously."

Months later, the Sixth Circuit sanctioned two attorneys $15,000 each over briefs with more than two dozen fake or misrepresented citations, plus double costs and the appellees' fees. Its framing is the part to remember: no filing should contain citations, however generated, that a lawyer has not personally read and verified. The tool is beside the point.

AI did not create a new obligation; it created a fast new way to violate an old one. The rules on using AI in legal research have not changed. Competence and candor still mean you verify before you file.

The bottom line

Can AI draft legal documents? Yes, and you should, on the documents where a first draft saves time and review is manageable: memos, demand letters, engagement letters, routine motions. Be far more careful with anything that cites authority, and never sign a filing whose citations you have not personally read. The dividing line is not the software; it is whether the draft is grounded in real sources and verified by a lawyer.

Before any AI-drafted document leaves your desk, run it through the Hallucination Shield: paste the text and every citation is checked for existence and support, free, no signup. It is the two-minute habit that keeps your name behind work you can defend.

Frequently asked questions

Can AI draft legal documents? Yes, and for some document types the first draft already saves real hours: internal research memos, demand letters, engagement letters, and routine procedural motions. Where it fails is anything that turns on jurisdiction-specific authority or citations, because a language model will invent cases that read perfectly and do not exist. Treat the output as a first draft from a fast junior associate, never a finished filing.

Is it safe to draft legal documents with AI? It is safe to draft with AI and unsafe to file AI output unread. The sanctioned filings of 2025 and 2026 share one pattern: a model generated citations from memory and no one checked them. Federal Rule of Civil Procedure 11 makes the signing attorney responsible for a reasonable inquiry into the law, whatever tool produced the draft. Ground the draft in sources you can read, then verify every authority before you file.

Which legal documents can AI draft well? AI first drafts are strongest on documents driven by facts and structure rather than controlling authority: internal research memos, demand letters, engagement and retainer letters, routine motions like extensions and continuances, and discovery request or response shells. These still need review, but the review is fast because the risk is wording and completeness, not fabricated law. The more a document depends on cited holdings, the less you should trust an unverified draft.

Can AI draft a legal brief with citations? It can produce the prose, but you cannot trust the citations without checking every one against the actual opinion. This is the single most dangerous use of AI drafting. A general-purpose model generates citations that mimic real reporters and pin cites while pointing to cases that do not exist or do not hold what the draft claims. Use a tool that retrieves and links real authority, then read each case yourself before it enters a filing.

What is the difference between AI generating and retrieving citations? Generating means the model produces a citation from patterns in its training data, so it can output an authority that never existed. Retrieving means the tool searches a real database, pulls actual documents, and cites only what it found, with a link back to the source. Generation-from-memory is what produced the fake-citation sanctions. Retrieval-grounded drafting is the only version safe for work headed to a court or a client.

Who is responsible if an AI-drafted document has a fake citation? The signing attorney, without exception. In Garner v. Kadince, the Utah Court of Appeals sanctioned counsel for a brief citing an AI-hallucinated case, and the Sixth Circuit fined two lawyers $15,000 each for fabricated citations, stressing that responsibility attaches however a citation was generated. Rule 11 and the ethics rules put the duty of verification squarely on the person who files.