The AI governance conversation usually runs on hypotheticals. A menagerie of projected risk, theoretical exposure, and “what could go wrong.” The legal industry doesn’t have that luxury anymore. It has a body count.
Court sanctions for AI-generated fake citations have now topped 1,031 cases globally, and the pace is accelerating: 30 to 50 new cases every month. More than half (518 and climbing) have come from US courts since January 2025.
This isn’t a handful of embarrassing one-offs anymore. It’s a pattern with a growth rate.
Penalties are escalating
The penalties are escalating along with the case count. A federal judge in Oregon issued the largest AI hallucination sanction in American legal history this year: $110,000, after two lawyers submitted 23 fabricated citations and eight invented quotations in a single filing. Courts in Alabama, Nebraska, Georgia, and Mississippi have handed down sanctions, dismissed cases, and in at least one instance, barred an attorney from filing in that court again without co-counsel sign-off.
The pattern behind these cases is nearly always the same: a lawyer used a general-purpose AI tool to draft a brief or research memo, the tool generated citations that looked completely plausible, and no one verified them before they went into a filing that a judge or opposing counsel would read. The technology didn’t fail dramatically. It failed quietly, in a way that was indistinguishable from correct work until someone checked.
That’s the part that should concern every firm, not just the ones that have already been sanctioned. Let’s be clear: the failure mode is invisible without verification.
The absence of verification is the actual liability, not the AI tool itself.
This is exactly the gap iBridge’s document intelligence and eDiscovery work is built to close. The firms getting sanctioned aren’t failing because they used AI. They’re failing because they used AI without a verification layer between the model’s output and the court record. Raw generative AI, bolted onto casework with no human checkpoint, will eventually produce something that looks right and isn’t.
Human-supervised document intelligence is the alternative: AI accelerates the drafting, research, and review work, but every citation, every extracted fact, and every generated summary passes through a verification step before it reaches a filing. The speed of AI without the discipline of verification isn’t a shortcut. It’s how you end up in front of a judge explaining why your brief cites cases that don’t exist.”
AI governance doesn’t fail by accident. It fails on schedule.
This is the same argument I’ve made about AI governance broadly, just with sharper teeth. Organizations that treat AI oversight as a formality are the ones showing up in sanctions orders. Organizations that build verification into the actual workflow are not. The difference isn’t the AI tool each side chose. It’s whether anyone checked its work before it mattered.
1,031 cases and climbing is what happens when an entire industry adopts a capability faster than it builds the discipline to supervise it. The lag doesn’t stay theoretical forever. Eventually it shows up in a sanctions order with your name on it.