A lawyer's relationship with artificial intelligence now has a body of case law all its own, and most of it reads like a warning label. The tools are genuinely useful. They are also confidently wrong in ways that look exactly like being right, and judges have stopped treating that as a novelty. The encouraging part, if there is one, is that nearly every sanctioned filing failed at the same predictable step. Avoid that step and AI becomes what it should be: a fast first drafter you never trust on faith.

This is a guide to that step.

The sanctions docket is now its own body of law

What started as a curiosity in 2023, when lawyers in Mata v. Avianca filed a brief citing decisions an AI tool had invented, has become a recurring docket entry. Researchers tracking the problem have now catalogued well over a thousand filings worldwide containing AI-fabricated content, and into 2026 the pace has not slowed. The penalties have escalated alongside the volume. Five-figure sanctions stopped being remarkable; the Sixth Circuit imposed a combined $30,000 on two attorneys for a brief carrying more than two dozen fake citations, and a single federal matter in 2026 produced a sanctions-and-costs award north of $100,000 against one lawyer.

The consequences are no longer confined to embarrassment. Across 2025 and 2026, courts have struck filings with no leave to amend, revoked pro hac vice admissions, ordered attorneys to self-report to their state bars, and referred matters for discipline. At least one large firm became a repeat offender, updating its written AI policy after a first incident and then allegedly filing hallucinated authority again, which is the cleanest illustration of the core lesson available: a policy that does not change the workflow gating what actually goes out the door changes nothing.

For the plaintiff-side bar this is not an abstract risk. The firms most exposed are the high-volume, intake-driven practices, mass tort and personal injury shops where deadlines stack and drafting gets pushed down to the most leveraged person in the room. The most-cited plaintiff-firm cautionary tale of the cycle came from exactly that environment: a drafting attorney fined and stripped of a temporary admission after a brief went out with citations no one had checked. The tool was never the violation. The unverified signature was.

There is no separate "AI rule." There are the ones you already have

The instinct to wait for a dedicated set of AI rules misreads the problem. The ABA's first formal guidance on the subject, Formal Opinion 512 from July 2024, does not invent new duties. It maps generative AI onto the obligations already governing your practice, and every sanctions order to date has done the same.

Competence (Model Rule 1.1). Technological competence now includes a working understanding of how a generative tool produces output and where it fails. You need not be an engineer, but you cannot claim competence while treating the model as an oracle. The ABA is explicit that these tools generate plausible text without understanding its meaning, and are therefore not a substitute for a lawyer's independent professional judgment.

Candor and the signature (Model Rule 3.3 and FRCP 11). When you sign a filing you certify, under Rule 11(b), that you made a reasonable inquiry and that the legal contentions are warranted. The sanctioned filings did not fail because AI was involved; they failed because that reasonable inquiry never happened before the signature. Rule 3.3 separately forbids knowingly presenting false statements of law, and a citation to a case that does not exist, or one that holds the opposite of what you claimed, breaches both.

Confidentiality (Model Rule 1.6). The duty reaches essentially all information relating to a representation. Pasting client facts, draft theories, or privileged material into a consumer tool can expose that information through the tool's data handling. Opinion 512 advises securing the client's informed consent before inputting confidential information into a self-learning tool, and is pointed that boilerplate language buried in an engagement letter does not qualify.

Supervision and fees (Rules 5.1, 5.3, and 1.5). Managing lawyers must set clear policy on permissible AI use, and supervising lawyers must ensure associates and staff are trained on the tools and their failure modes. And the time AI saves is time you did not spend; a bill must reflect hours actually and reasonably expended, not the hours the work would have taken without it.

Check the forum before you draft

The federal bench has not settled on a uniform approach, which means the binding requirement is frequently whatever the individual judge or court has issued. Before AI touches anything destined for a filing, read the assigned judge's standing orders, because many now require a certification of whether AI was used and that any AI-assisted content was human-verified, and some demand disclosure of the specific tool. Check the court's local rules for a general order on generative AI, and confirm which rule governs the particular filing, since trial and appellate forums handling the same matter can impose different obligations. The absence of an AI-specific rule is not permission to skip verification; Rule 11 and the duty of candor apply in full regardless.

Where AI earns its keep, and where it bites

The safe uses share a trait. Either the lawyer can verify the output against something authoritative, or the output never asserts a fact or a holding in the first place. Drafting structure for a brief you will rewrite, tightening your own prose, summarizing a document you supplied and can check, pressure-testing an argument against its counterargument: these are low-risk because the model is not being asked to supply authority.

The dangerous uses share the opposite trait. The moment you ask a generative tool to find cases supporting a proposition, or you trust the citation, quotation, or pin-cite it returns, or you rely on its summary of what a decision held, you have handed the model the one job it cannot do reliably, and its fabrications are indistinguishable from the real thing. The most seductive of these is the verification shortcut: running one AI's citations through a second AI is not independent verification. It is two unreliable narrators agreeing with each other.

The verification discipline

Treat every factual and legal assertion the model hands you as an unconfirmed lead, and the rest is procedure.

First, pull every citation independently in Westlaw, Lexis, or another authoritative source. If you cannot find it, it almost certainly does not exist, and it comes out. Second, confirm the authority is not only real but still good law, because existence is not enough if the case was reversed, overruled, or quietly distinguished into irrelevance. Third, read the holding yourself, since the tools routinely misdescribe what a genuine case actually decided, and verify it stands for the proposition you are citing it for. Fourth, check every quotation and pin-cite against the source, because fabricated quotes attributed to real decisions are common. Fifth, where the model summarized a record document, compare its summary against the document rather than the summary. And throughout, keep a record of the review you did, because if a court later questions the filing, contemporaneous proof that you verified is the strongest protection available to you.

A note on confidentiality

Before anything client-related enters a tool, answer two questions: where does this data go, and who agreed to that. Know whether the tool trains on, retains, or shares your inputs, because an enterprise or zero-retention configuration behaves nothing like a free consumer version. Obtain specific informed consent before entering confidential information into a self-learning tool, framed as a real explanation of the risks rather than buried boilerplate. Default to de-identifying inputs when the task does not require the identifiers, and vet AI vendors the way you would any other outsourced service, because the duty to confirm a third party will protect client confidences extends to them.

Disclosure runs in two directions

To the court, follow the operative standing order or local rule precisely; if it requires a certification or a disclosure, provide exactly what it asks, and where no rule requires disclosure, none is generally mandated, though the verification duty is unchanged. To the client, disclosure is not required for every use, but you should tell the client when AI will meaningfully shape a significant decision in the representation, or when its use affects how they are billed.

The pre-filing checklist

Before your signature goes on anything AI touched, run the short version:

• Every cited authority opened and confirmed to exist in an authoritative database.
• Each cited case confirmed as good law, not reversed, overruled, or superseded.
• Each holding read and confirmed to support the proposition cited.
• Every quotation and pin-cite checked against the original source.
• Any AI-summarized facts compared against the underlying record.
• No confidential client data entered into a tool without informed consent.
• The assigned judge's standing order and the local rules on AI checked and satisfied.
• Any required AI certification prepared and accurate.
• The brief read end to end by a human who can vouch for every line.

The takeaway

Generative AI does not make brief-writing categorically riskier or safer. It relocates the risk to a single point, the gap between what the model produced and what the lawyer verified, and then it makes that gap invisible by dressing fabrication in the same format as fact. The discipline is not complicated and it is not new. You may use AI to help write the brief. You may never let AI be the reason you did not read the law. Your signature certifies your inquiry, not the model's, and the sanctions docket is, at bottom, a list of lawyers who forgot that.