The best-documented area of legal AI, and the one where the reasoning is more instructive than the penalties.
Start a conversation with the AI Adoption Concierge, already scoped to courts & sanctions. Pick a starting point, or describe your situation directly.
The judicial response to AI has run on three tracks. Individual judges issued standing orders, beginning in 2023 and proliferating unevenly — some requiring disclosure of any generative AI use, some a certification that citations were verified by a human, a small number prohibiting it. Courts sanctioned lawyers who filed fabricated citations, with reported outcomes escalating from four figures to five, to removal from cases, bar referrals and at least one reported indefinite suspension. And rules committees began considering whether existing evidence and procedure rules are adequate, so far mostly concluding to study further. For a practising lawyer the operational content of all this is small and boring: check the judge's standing order, and verify every citation. The reasoning underneath is where the useful material is.
Patterns that recur across reported decisions, rather than a list of penalties.
Courts have not accepted delegation to a tool, to co-counsel, or to a trusted internal system as an answer. The signature on the filing carries the obligation.
Several decisions turned sharply on what happened after discovery — whether the problem was disclosed or minimised, and whether the court found it rather than being told.
Reported outcomes have been markedly worse where the same conduct recurred after a warning.
At least one reported matter involved a firm that had an AI policy which the lawyers did not follow. Having the document did not resolve it.
Reported outcomes have included signing and supervising lawyers, not only the person who generated the text.
Courts have largely proceeded under certification and candour obligations rather than new AI-specific rules.
What this implies for filing practice.
This is the area of AI practice with the most direct professional consequences on record.
Existing rules were drafted for inaccurate AI output. In 2026 courts in two countries addressed a different thing — text concealed inside a filing, formatted to be invisible to a human reader, instructing any AI that processed the document. One court reached it through the good-faith certification because nothing addressed it directly. Firms whose review pipelines ingest adversary-authored documents should read this as a live gap.
The judge's own page on the court website, checked per matter. There is no reliable central register — several trackers exist and none is complete, because orders are issued, amended and withdrawn individually. Requirements vary in kind as well as strictness: some ask for disclosure that AI was used, some for a certification that a human verified every citation, and those are different obligations with different drafting consequences. Where a firm appears before the same judges repeatedly, keeping a local list and refreshing it periodically is more reliable than checking a national tracker.
Courts have not treated it that way. The consistent position across reported decisions is that the certification obligation attaches to the person signing the filing, and that reliance on a tool — including a paid, legal-specific tool — does not discharge it. This matters because the tools that fabricate are not only the consumer chatbots: independent testing of purpose-built legal research products has found meaningful error rates, including the harder failure of a real authority cited for a proposition it does not support.
Take professional-responsibility advice in the relevant jurisdiction immediately — this page is not that advice. What the reported record suggests, without being a rule, is that outcomes have differed substantially depending on whether the lawyers disclosed promptly and accepted responsibility, or whether the court discovered it and the initial response minimised it. Firms that have thought about this before it happens tend to respond faster, which is the argument for deciding the escalation route in advance rather than during.
Proposals have been under consideration rather than adopted. A proposed federal evidence rule addressing machine-generated evidence offered without a sponsoring human expert went out for public comment, and in mid-2026 was reported to have been returned to committee for further study alongside a separate deepfake-authentication proposal. Some states have moved faster on adjacent questions, including requirements to disclose AI-generated or altered material offered as evidence. Treat this as unsettled and moving, and check the position in the jurisdiction rather than assuming a national answer.
Describe how filings are checked today. The Institute will help you close the gap.