Generative AI is no longer a novelty in legal practice. At conferences like Legalweek, the conversation has moved well beyond whether firms should pay attention to AI. The market has shifted to implementation. Law firms, courts, and legal departments are now actively testing, adopting, and integrating AI into legal workflows. But as AI adoption expands, so do the risks. Researcher Damien Charlotin’s database tracking legal AI hallucinations has now documented more than 1,300 cases globally and over 888 cases in the United States, reflecting a growing wave of incidents involving fabricated authorities, inaccurate quotations, and other false outputs in legal settings.
The profession is also past the point where the only concern is a chatbot inventing a fake case citation from scratch. The more serious problems are often subtler. Today, AI tools may generate legal propositions that sound persuasive, attribute them to real authorities incorrectly, or attach false quotations to legitimate citations. Those errors can be harder to detect and, for that reason, more dangerous.
That reality points to one of the clearest trends in the legal market: the competitive divide is no longer simply between organizations that have AI tools and those that do not. It is increasingly between organizations that understand how to use those tools competently and those that do not. Access is becoming common. Competence remains scarce.
One of the more persistent misconceptions in the market is that legal-specific AI tools entirely eliminate the weaknesses associated with general-purpose models. They do not. Specialized legal tools may reduce risk, but they do not make hallucinations, analytical errors, or source misattribution disappear. As a result, lawyers cannot assume that a product is reliable simply because it is branded for legal use.
This issue is not limited to practicing lawyers. Judges are increasingly using AI as well. A recent survey found that more than 60% of responding federal judges reported using at least one AI tool in their work, although only 22.4% reported using such tools weekly or daily. That suggests the judiciary, like the private bar, is in a transitional moment: AI adoption is broadening, but confidence and routine use are still developing.
AI use has also started to appear directly in judicial opinions. Judge Kevin Newsom of the Eleventh Circuit Court of Appeals discussed the use of AI tools for judicial interpretation in published concurrences in United States v. Deleon and Snell v. United Specialty Insurance Company, illustrating that judicial engagement with these technologies is no longer theoretical.
Judge Joshua Deahl of the D.C. Court of Appeals carried out a similar AI-supported analysis in a dissent in Ross v. United States in February 2025. At the same time, AI-related errors have surfaced in judicial work product as well. Reuters reported in 2025 on withdrawn rulings connected to erroneous citations and questions over AI misuse in opinions associated with Judges Julien Neals (District of New Jersey) and Henry Wingate (Southern District of Mississippi). These incidents underscore that insufficient AI training creates risk not only for lawyers, but for the judiciary itself.
All of this helps explain why one of the most important developments in legal AI is the growing need for practical, hands-on training. Written AI policies remain necessary, but policy alone is not enough. A firm may authorize tools such as CoCounsel, Lexis Protege, or Microsoft Copilot, yet still leave its lawyers unequipped to determine which tool is appropriate for which task, what outputs require heightened verification, and where ethical or professional responsibility issues are most likely to arise. This is where the market is heading: from tool acquisition to operational competence.
In that sense, legal AI’s next phase will not be defined merely by adoption rates. It will be defined by whether institutions can translate access into sound judgment, disciplined verification, and responsible use. The firms, courts, and legal organizations that will lead in this environment are not simply those with the most tools. They are the ones investing seriously in understanding those tools’ limitations and training their people.
