What Ethics Rules Govern Defense Attorney AI Court Filings?

Authority: American Bar Association
Rule type: ethics opinion
Jurisdiction scope: US national (ABA Model Rules)
Effective date: Jul 29, 2024
Source text: Read primary rule text ↗

Defense attorneys must apply existing duties of competence, confidentiality, client communication, fees, candor, and supervision to AI-assisted filings and verify all cited authorities before signing.

Last verified: August 2, 2026. Regulation & Ethics analysis; legal information, not legal advice. No named legal reviewer was provided. Primary PDFs identified but not crawled for this article — including ABA Formal Opinion 512, certain Florida materials, and the primary Crabill, Farris, and Couvrette records — should be checked against the primary documents before any direct quotation is published.

For defense attorney AI court filings ethics rules, the answer starts at the signature line. The lawyer who signs or submits the filing is not merely promising that the citations survived a search-box check. The governing framework is familiar professional-responsibility law applied to AI conditions, plus state and tribunal rules that may add filing-specific attestations or disclosure duties.

ABA Formal Opinion 512 is the baseline map, not the whole operating manual. It organizes existing duties of competence, diligence, communication, fees, confidentiality, candor, meritorious claims, and supervision around generative AI use; the binding detail then depends on the jurisdiction and the tribunal where the filing lands.[1]

Legal pleading beside a laptop showing abstract AI-generated text with layered review sheets and a courthouse column in the background

The obligations tracker before the filing leaves counsel’s desk

A defense filing often passes through more hands than the docket reflects: associate, investigator, paralegal, intern, vendor platform, research database, supervising partner, appointed counsel voucher review, and finally the attorney of record. The rules below matter because the signature compresses that chain into one professional representation.

Duty	Baseline source	State or court layer that may sharpen it	Filing implication for defense counsel	Enforcement record that shows the risk
Competence with the tool and the legal work	ABA Formal Opinion 512 maps generative AI use to Model Rule 1.1, including technology competence, and requires a reasonable understanding of the tool’s capabilities and limits.[1]	WSBA Advisory Opinion 2025-05 treats competence as one part of a seven-duty AI framework for Washington lawyers.[2] Illinois ARDC guidance likewise frames AI use through Rule 1.1 technology competence.[3]	The lawyer must understand enough about the AI system and the task to decide what can be delegated to the tool, what must be checked, and whether the final legal position is supportable.	United States v. Hayes and United States v. Farris show that failed AI verification is treated as a professional failure, not a mere drafting mishap.[4][5]
Diligence	Formal Opinion 512 ties AI use to Model Rule 1.3.[1]	State ethics opinions apply the ordinary diligence duty to the speed and confidence that AI tools can create.[2]	A fast draft does not excuse missed controlling authority, unexamined factual assertions, or late correction once a hallucination is discovered.	In Farris, CJA counsel was removed and denied compensation after fabricated authorities appeared in appellate briefing.[5]
Client communication	Formal Opinion 512 maps AI use to Model Rule 1.4, including circumstances where the client needs to know how AI will be used in the representation.[1]	WSBA’s opinion includes communication as a separate AI duty.[2]	Defense counsel should decide whether the client must be told that AI will be used, especially where the use affects confidentiality, cost, strategy, or a material step in the representation.	The risk is practical as well as disciplinary: a client cannot give meaningful instructions about strategy, plea posture, or disclosure tolerance if counsel never explains the AI use that affects those choices.
Fees and billing	Formal Opinion 512 applies Model Rule 1.5 and warns that lawyers may not charge clients for time spent learning to use a generative AI tool.[1]	WSBA includes fees in its AI-duty framework.[2]	If AI reduces drafting time, billing must reflect the work actually performed and the value permitted by the fee agreement and governing rules. Training time belongs to the lawyer, not automatically to the client.	Fee consequences can arise even when the filing problem appears to be citation-related; Farris is the stark example in the appointed-counsel context because compensation was denied.[5]
Confidentiality	Formal Opinion 512 applies Model Rule 1.6 to AI use.[1]	WSBA Advisory Opinion 2025-05 gives the concrete criminal-defense example: entering a client name, plea negotiations, and privileged communications into a public AI tool can violate RPC 1.6 and create privilege-waiver risk.[2]	Before any prompt is entered, counsel must decide whether the platform receives, stores, trains on, discloses, or can be compelled to produce client information. Redaction alone may not solve the problem if the remaining facts identify the client or strategy.	This is the defense-specific danger that citation-only advice misses. A fabricated case can be corrected; disclosed plea negotiations may not be recoverable in the same way.
Candor, meritorious claims, and truthfulness	Formal Opinion 512 links AI filing output to Model Rules 3.1, 3.3, and 8.4(c).[1]	Florida Rule 2.515(d)(2), effective June 15, 2026, adds a statewide signature attestation for lower-court filings: the signer represents that cited authorities exist and are accurately cited, with sanctions including reprimand, contempt, striking, dismissal, costs, and fees.[6] New York Part 161, effective June 1, 2026, permits AI use without disclosure but requires careful review and independent assurance that the filing contains no fabricated cases, statutes, or other material.[7]	The signed filing must be independently checked for real authority, accurate quotation or paraphrase, and supportable factual use. Florida makes that verification part of the signature representation; New York imposes a review-and-assurance duty without a general disclosure requirement.	Hayes, Farris, Coleman, and Couvrette are the filing-trust cases courts and ethics writers now use to explain why unsupported AI output is not harmless.[4][5][7][8]
Supervision	Formal Opinion 512 applies Model Rules 5.1 and 5.3, treating AI-related work as part of the lawyer’s supervisory responsibility.[1]	WSBA includes supervision in its seven-duty structure.[2]	The signing lawyer must supervise junior lawyers, staff, interns, and AI-enabled workflows. Blaming the person who pasted the output is not a defense to the supervising lawyer’s duty.	People v. Crabill is the hard example: the lawyer received a 90-day suspension after ChatGPT-fabricated citations appeared and responsibility was shifted toward an intern.[9]
Tribunal disclosure	Formal Opinion 512 should not be treated, on the available materials, as imposing a categorical duty to disclose AI use to every tribunal.[1]	Disclosure duties come from court rules, standing orders, or judge-specific instructions. The research record identifies a 300-plus judge standing-order layer, and those orders may require disclosure even where state ethics guidance does not.[7]	Counsel must check the forum before filing: statewide rule, local rule, individual judge order, administrative order, and any case-management instruction. A no-disclosure default in one forum does not travel.	The filing consequence is avoidable: the lawyer may have verified every citation and still violate a judge’s AI standing order if disclosure was required.

ABA Formal Opinion 512 is the starting drawer

Formal Opinion 512, issued July 29, 2024, is important because it refuses to invent a separate AI ethics code. It instead asks the ordinary Model Rules questions: Did the lawyer understand the technology enough to use it competently? Did the lawyer preserve confidentiality? Was the client adequately informed? Was the fee reasonable? Did the lawyer supervise the people and systems used in the work? Did the filing mislead the tribunal?[1]

That structure is useful for defense lawyers because AI risk in litigation is not confined to hallucinated citations. A suppression motion, bail submission, sentencing memorandum, or civil defense brief may contain client confidences, investigative facts, negotiation history, expert theories, or strategic concessions. The filing can be technically well cited and still be ethically defective if the prompt disclosed protected information or the client was billed for the lawyer’s tool training.

The point is also hierarchical. ABA guidance helps identify the duties, but it does not answer every filing question in every courthouse. State ethics opinions may interpret the lawyer’s duties under state rules. Statewide court rules may attach representations to the act of signing. Individual judges may require an AI certificate, disclosure, or other case-specific assurance. Those are different sources with different force.

Confidentiality is not an abstract AI issue in defense work

WSBA Advisory Opinion 2025-05 is valuable because it does what many AI summaries do not: it gives a criminal-defense confidentiality illustration. In that example, a defense attorney uses a public AI tool and inputs a client’s name, plea negotiations, and privileged communications. WSBA treats that conduct as a violation of RPC 1.6 with privilege-waiver risk.[2]

That example should change the order of review. Citation verification is a late-stage filing control. Confidentiality review happens before the first prompt. If the tool is public, if the terms permit retention or training, if the matter is sealed, if facts identify the client despite redaction, or if the prompt includes plea posture or privileged communications, the filing problem has already begun before the draft exists.

Do not paste client names, witness identities, plea communications, privileged facts, sealed materials, or strategy into an AI system unless the lawyer has confirmed that doing so is permitted under the governing confidentiality rules and the tool’s actual data practices.
Do not assume that redaction is sufficient where the remaining facts identify the client, the prosecution, the incident, or the negotiation posture.
Do not let procurement language substitute for legal judgment. A vendor promise may be relevant, but the lawyer still owns the Rule 1.6 decision.

For a defense lawyer, this is not only a discipline issue. Confidentiality failures can affect liberty, plea leverage, impeachment strategy, cooperation risk, protective orders, and the client’s willingness to communicate candidly with counsel.

The jurisdiction layer changes the filing obligation

Layered authority structure with translucent tiers and a scales of justice symbol

By Q3 2026, the filing lawyer has to separate at least three layers: ethics guidance, statewide filing rules, and judge-specific orders. Treating all of them as a generic “AI disclosure rule” is how lawyers miss the rule that actually binds the document they are about to submit.

Florida: signature attestation, not merely good practice

Florida Rule 2.515(d)(2), effective June 15, 2026, makes the signer’s certification the center of the rule. For lower-court filings, every attorney or pro se signer represents that cited authorities exist and are accurately cited. The reported sanctions available for violation include reprimand, contempt, striking, dismissal, costs, and fees.[6]

The Florida materials also matter for what they remove from the local-rule maze: a separate administrative order bars local circuits from imposing their own AI-disclosure rules.[6] That does not make AI use invisible. It means Florida’s statewide attestation is the filing control counsel must build around, while still checking any judge-specific or case-specific order that is valid in the matter.

For the defense signer, the operational consequence is plain. A junior lawyer’s research memo, a paralegal’s cite list, or an AI-generated first draft does not dilute the representation made by the signature. If the cited case does not exist or the quotation is wrong, the signer has made the defective filing representation.

New York: permitted use without general disclosure, but careful independent review

New York Part 161, effective June 1, 2026, takes a different shape. The NYSBA discussion describes it as permitting AI use without disclosure, while requiring the attorney to carefully review and independently ensure that the filing contains no fabricated cases, statutes, or other material. Judges retain discretion.[7]

That distinction is easy to lose. New York’s rule, as described in the available source, is not a general confession requirement every time AI touches a draft. It is a review obligation, backed by judicial discretion. A defense lawyer who files in New York should therefore document the independent review, not invent a disclosure where no rule requires one — unless the judge’s order, local practice, or case instruction says otherwise.

Judge standing orders: the layer that defeats national shortcuts

The judge-standing-order layer is now too large to treat as an exception. The research record identifies more than 300 judge standing orders addressing AI use.[7] Some orders require disclosure. Some require certification. Some focus on verification. Some may apply only to particular chambers, case types, or filings.

This is where a national ethics summary can mislead a filing lawyer. Formal Opinion 512 may provide the baseline duties; New York may permit use without general disclosure; Florida may use a statewide signature attestation. None of that answers whether the judge in the actual case has entered a standing order requiring a separate AI disclosure or certificate.

Forum question	What defense counsel should verify before filing
Which tribunal receives the filing?	Federal, state, appellate, trial, administrative, arbitration, and specialty courts may impose different AI controls.
Which statewide rule applies?	Check rules like Florida Rule 2.515(d)(2) and New York Part 161 before relying on a general ethics article.
Is there an individual judge order?	Review standing orders, case-management orders, scheduling orders, and chambers instructions.
Does the order require disclosure or certification?	Separate a duty to verify from a duty to disclose. They are not the same obligation.
Who signs?	The attorney whose signature appears on the filing should be treated as the person making the filing representation, even if AI output or staff work entered the draft upstream.

Verification is necessary, but it is not the whole ethics analysis

Verification still deserves a hard floor. The Stanford RegLab study tested Lexis+ AI, Westlaw AI-Assisted Research, and Ask Practical Law AI and reported hallucination rates in the 17% to 33% range across the tools studied.[10] That finding is not a reason to rank products in a filing brief. It is a reason not to outsource professional judgment to the confidence of a legal research interface.

A defensible verification workflow should confirm that each cited authority exists, remains good law for the proposition used, comes from the jurisdiction represented, is quoted or paraphrased accurately, and supports the procedural posture of the filing. In criminal and civil defense work, that also means checking whether the authority is being used to support the actual relief requested, not merely a nearby legal proposition.

The enforcement record explains why courts are impatient with less. In Hayes, the Eastern District of California imposed a $1,500 sanction after AI-generated false authorities appeared, and the order was served on the California and D.C. bars and all district judges.[4] In Farris, appellate counsel appointed under the Criminal Justice Act was removed and denied all compensation after fabricated citations appeared in briefing.[5]

Those outcomes are not just punishment for bad research. They are damage to court trust. A defense lawyer asking for suppression, release, dismissal, resentencing, or reversal needs the tribunal to believe the filing’s factual and legal representations. AI hallucinations spend credibility that may be needed for the client’s substantive argument.

Supervision means the intern, associate, and tool are part of the same control system

Formal Opinion 512’s supervision point is easy to understate. If AI-assisted work flows through nonlawyer staff, junior lawyers, contractors, or an intern, the lawyer responsible for the matter must have a process that catches the foreseeable failure modes. “The intern did it” is exactly the answer supervision rules are designed to reject.

People v. Crabill is the useful record here because the accountability problem is visible. The lawyer received a 90-day suspension after ChatGPT-fabricated citations appeared, and the available reports describe an attempt to shift blame toward an intern.[9] The docket identifiers for Crabill are inconsistent across secondary materials, so any publication should verify the primary disciplinary record before stating the docket number.

A defense practice does not need an elaborate AI committee to meet the basic point. It needs a responsible lawyer assigned to the filing, a rule for what may be entered into AI systems, a review path for AI-generated legal and factual content, and a record showing who checked the final authorities and confidential-information issues before signature.

Candor problems are expanding beyond fake citations

The earliest AI filing stories were often summarized as fake-citation cases. That label is now too narrow. The NYSBA discussion of State v. Coleman describes a defense motion built on ChatGPT-invented statements, not merely fabricated case law.[7] That matters because candor problems can arise from invented record facts, false descriptions of testimony, mischaracterized statutes, phantom quotations, and procedural assertions the lawyer never checked.

Couvrette v. Wisnovsky also belongs in the working file because judges now cite sanctions surveys to show that AI filing failures are not isolated curiosities.[8] The lesson is not that every AI mistake deserves the same sanction. It is that courts now know what these failures look like, and a lawyer’s explanation will be assessed against a growing public record of similar conduct.

Aggregate counts should be handled carefully. Different trackers use different scopes and dates: Illinois materials cited more than 490 global matters as of January 2026, while NYSBA discussed more than 1,000 U.S. matters in June 2026.[3][7] Those figures should not be blended into one master statistic. For filing counsel, their value is narrower: courts, regulators, and opposing counsel have enough examples to recognize the pattern.

What the signing defense lawyer should document

Documentation does not need to be performative. It needs to answer the questions that will matter if the filing is challenged: what tool was used, what categories of information were entered, who reviewed the output, what authorities were independently checked, whether the client had to be informed or consent, what court rule or standing order applied, and who approved the final filing.

Confidentiality decision: record whether client-identifying, privileged, sealed, or strategic information was excluded from AI prompts, or why a particular protected system was permissible.
Client communication decision: record whether AI use required client notice or consent under the facts of the representation and governing state rules.
Supervision decision: identify who used the tool and who reviewed the work before it entered the filing.
Authority verification: save the final cite-check path, including confirmation that authorities exist, remain valid, and support the proposition used.
Forum rule check: note the statewide rule, local rule, judge standing order, and any case-specific instruction reviewed before filing.
Billing treatment: separate chargeable legal work from nonchargeable time spent learning or configuring the AI tool.

This record is not a universal safe harbor. It is a way to make the lawyer’s actual professional judgment visible if the filing is questioned.

The jurisdiction-first answer

Before an AI-assisted defense filing reaches a tribunal, counsel should identify the governing forum, check state ethics guidance, check statewide court rules, check the assigned judge’s standing and case-specific orders, document confidentiality and supervision decisions, and then verify every authority and material assertion. Citation verification is one required step. It is not the ethics framework.

The lawyer who signs the filing should assume the court will treat that signature as the point where all upstream AI use, staff work, vendor output, and legal judgment become counsel’s professional responsibility.

References

ABA Formal Opinion 512, American Bar Association, July 29, 2024, link
WSBA Advisory Opinion 2025-05, Washington State Bar Association, 2025, link
Paste in Haste: The Fallout of AI Hallucinations in Court Filings and the New ARDC's Guide to Implementing AI, Illinois Courts / ARDC, January 2026, link
United States v. Hayes, FD.org, link
United States v. Farris, Norton Rose Fulbright, mid-2026, link
Florida Supreme Court Adopts Statewide AI Citation Rule for Lower Court Filings, Friedman Vartolo, link
Beyond the Mirage: Beware of Generative AI and Hallucinations, New York State Bar Association, June 2026, link
Parade of Horribles, NWSidebar, link
People v. Crabill, FD.org, link
Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, Stanford RegLab, link

Operationalizing workflow

No workflow has been explicitly linked to this obligation yet. See Workflows generally.

Illustrative cases

No illustrative case is currently tracked for this obligation. See Risk Digest for documented incidents generally.

← Back to Regulation

Report a correction or tip

Spotted an outdated figure, a misstated fact, or a ruling this regulation entry should reflect? Public comments are disabled for this content given the professional cost of a misreported case outcome, penalty amount, or rule text — use the structured correction channel instead.

Report a correction or tip for this record →