Courts Are Sanctioning Lawyers for AI Citations. The Fix Isn't Better AI — It's Honest AI.

In 2023, two attorneys submitted a brief in Mata v. Avianca, Inc. that cited six cases none of the judges could locate — because they were fabricated by ChatGPT. The court imposed sanctions. In 2025 and 2026, similar rulings followed across multiple federal courts: sanctions for AI-hallucinated citations, orders requiring attorneys to certify that cited cases exist before submission, and standing orders in multiple districts specifically restricting AI-generated filings without independent verification.

Each ruling produced the same popular response: AI legal tools need to get better at not making things up. This diagnosis is accurate as far as it goes. But it stops short of the deeper design question the sanctions actually surface: what should a legal AI tool do when it doesn't know something with sufficient confidence to stake a professional reputation on it?

The answer, it turns out, is the same answer a careful lawyer would give: say so, and stop.

What the Sanctions Actually Require

The courts issuing sanctions are not prohibiting AI-assisted legal research. They are requiring verification. The duty at issue in every AI sanctions case is the attorney's obligation under Federal Rule of Civil Procedure 11 to certify that legal contentions are warranted by existing law — not by what an AI said the law says.

ABA Formal Opinion 512 (July 2024), issued by the ABA Standing Committee on Ethics and Professional Responsibility, addressed this directly: competent use of AI tools in legal practice requires that attorneys independently verify AI-generated legal research before relying on it in any filing or advice. The opinion does not prohibit AI use. It establishes that the attorney, not the AI tool, bears the professional responsibility for the accuracy of the output.

The verification duty creates a specific design requirement for any AI tool intended for legal use. If an attorney must independently verify every AI-generated citation before filing, then an AI tool that produces confident-sounding citations without surfacing its confidence level is creating work, not removing it. The attorney still has to check everything — but now has to check against an AI output that was delivered as if it were reliable.

The Overconfidence Problem

Most AI systems are trained to produce answers. When a general-purpose language model is asked about a legal case it hasn't seen or a statutory provision that doesn't exist, it does what it was trained to do: it generates a plausible-sounding response. It fabricates citations the way it completes every other sequence — by predicting what comes next, based on patterns from training data. There is nothing in the base model's architecture that recognizes "I don't have reliable information about this specific case" as a separate class of response requiring special handling.

This is an engineering problem with a known solution. Systems can be designed to recognize when a query falls outside their reliable knowledge domain, to surface uncertainty rather than suppress it, and to decline to produce a specific citation when confidence falls below a threshold. These are choices. They require deliberately prioritizing accuracy over apparent helpfulness — and accepting that a tool that sometimes says "I can't find that" will feel less capable than a tool that always produces something.

In most domains, this tradeoff is acceptable. In legal practice, it is not. A tool that confidently produces a fabricated citation in a court filing does not just fail at its task — it creates professional liability for the attorney who used it.

Attorney-Client Privilege and Third-Party AI Tools

A separate, less-discussed issue the sanctions pattern raises is what happens to attorney-client privilege when legal analysis passes through a third-party AI tool. The question is not hypothetical. Every time an attorney enters client information into a commercial AI platform to generate legal research, analysis, or draft work product, that information is transmitted to a third-party server.

Whether this transmission destroys privilege depends on whether it falls within an existing exception — most commonly, the "agents of the attorney" doctrine, which extends privilege to third parties who assist the attorney in rendering legal services. The application of this doctrine to AI tool providers is unsettled. Analysis published by Parker Poe in May 2026 noted that courts have not definitively ruled on whether a commercial AI service constitutes a "privileged agent" in the way that a paralegal, legal assistant, or outside expert would — and that the analysis turns significantly on the confidentiality protections in the service agreement and how the information is used by the provider.

For consumer-facing AI tools used by clients without attorney direction, privilege is even less likely to attach. The privilege belongs to the attorney-client relationship; communications made outside that relationship for non-legal purposes generally are not protected. A client who uses a legal AI tool independently — without the involvement of an attorney — should not assume that those interactions are privileged.

What Honest AI Looks Like in Legal Practice

The sanctions pattern, the ABA guidance, and the privilege question together point to a consistent conclusion about what a professional legal AI tool should do differently from a general-purpose AI assistant.

Abstention is a feature, not a failure. When confidence is insufficient — because the query involves a novel area, a jurisdiction with sparse case law, or a specific statutory provision the tool hasn't reliably indexed — the correct output is an explicit statement of what the tool does and doesn't know, with enough information for the attorney to decide where to direct their own research.

This also means that a legal AI tool should distinguish between what it found and what it inferred. Surfacing a clear statutory exemption that appears directly in the text of the statute is a different confidence level than suggesting that an exemption "may apply" based on analogy to cases in a related area. Both answers can be useful. Neither is useful if they are delivered with the same apparent confidence.

The tools that will prove most valuable in professional legal practice are the ones designed around this principle: do less, confidently. Surface the exceptions and exemptions that are actually there, with citations that can be verified, and decline to speculate beyond that boundary. Every fabricated citation in a court filing was a moment when a tool should have stopped — and kept going instead.

That is the design problem Legal Exception is built to solve.

The content on this site is legal information, not legal advice. It does not create an attorney-client relationship and cannot substitute for consultation with a licensed attorney about your specific situation.