Every few weeks, another headline: an attorney sanctioned for citing a case that doesn’t exist. A brief built on quotes no judge ever wrote. A filing that unravels the moment opposing counsel tries to pull the source.

Each one gets treated like an isolated embarrassment: one lawyer, one bad afternoon, one AI tool gone wrong. It isn’t isolated. It’s a continued trend and legal and compliance teams should be paying attention to where it’s headed.

The pattern, in numbers

A tracking database maintained by a research fellow at HEC Paris’s Smart Law Hub is now the most comprehensive public record of this problem. As of early 2026, it had logged more than 1,200 documented cases worldwide of generative AI producing hallucinated content that made it into a court filing, up from roughly 200 a year earlier. New cases are being added at a rate of five to six per day.

Courts aren’t just noticing. They’re pricing it in. Tracking data compiled in early 2026 put total sanctions for AI-fabricated citations at an estimated $145,000 in Q1 alone, and some jurisdictions have moved past case-by-case discretion entirely. Oregon’s Court of Appeals established a per-infraction sanctions formula in late 2025; a federal court in the same state applied it this year to a case with 15 fabricated citations and eight fabricated quotations across three briefs, dismissing the underlying claims with prejudice and fining the lead attorney $15,500.

A few more data points worth knowing, regardless pf which side of the filing you stand on:

• A federal appellate panel sanctioned two attorneys after finding more than two dozen fake citations in a single brief, plus misrepresented quotations and citations that didn’t even support the arguments attached to them.
• The Tenth Circuit sanctioned an attorney $1,000 and referred the matter to her state’s disciplinary authorities.
• A New York appellate court ordered over $10,000 in sanctions after fabricated citations and inaccurate quotations surfaced in a civil brief.

It’s not just attorneys, either. A notable share of documented cases (well over half, by some counts) involve people representing themselves, often because they turned to AI as a substitute for counsel they couldn’t afford. The tool built to close the access-to-justice gap is, in a meaningful number of cases, actively widening it. Even judges’ chambers haven’t been immune. Two federal judges were compelled to retract rulings in 2025 after AI-related errors surfaced in draft orders, one traced to a law clerk’s unauthorized use of a generative tool.

Why this keeps getting worse, not better

Three things are pulling in the same direction at once. Adoption moved faster than verification habits did. The tools got convincing well before the profession built the discipline to check them. Time and billing pressure reward speed, and a fabricated citation reads exactly like a real one until someone opens the reporter. And most firms still don’t have a standard, enforceable verification step built into AI-assisted work the way privilege review or conflict checks are built into everything else.

Some courts are responding with disclosure rules. A growing number now require attorneys to certify that AI-generated content was checked before it’s filed. Rules like that only work if someone is actually doing the checking.

The takeaway: verification isn’t optional anymore, at any stage

Here’s the thing courts keep saying, in sanctions order after sanctions order: “I trusted the tool” is not a defense. Anything AI touches (a citation, a summary, a quote) needs a human verification step before it leaves the building, and that step needs to be demonstrable, not assumed.

That principle doesn’t stop at legal research. It’s the same principle ModeOne was built around for the evidence itself.

A hallucinated citation is a problem because nobody can trace it back to something real. An undocumented, unverifiable chain of custody creates the exact same problem for collected mobile data and manual, ad hoc collection logs are increasingly being flagged industry-wide as failures waiting for a judicial challenge. The issue was never really about collecting data fast. It’s about being able to show, later, precisely what was collected, from where, from whom, and under what authorization. And increasingly important, in a way that holds up the moment someone checks.

That’s the gap between “fast” collection and fast and defensible collection. It’s the difference between a full-device dump nobody can fully account for and a targeted, participant- and date-filtered extraction with a documented, repeatable process behind it. As courts sharpen their scrutiny of AI-assisted work product, that same scrutiny extends to how the underlying evidence was gathered in the first place, “We collected everything, just in case” is not going to be a stronger answer than “we trusted the AI.”

The hallucination problem isn’t going away. The case for building verification and defensibility into every stage of a matter (not just the brief, but the evidence behind it) has never been stronger.

ModeOne collects targeted mobile data (texts, calls, chats, photos, videos, emails, location data, contacts, apps, media) remotely, same-day, with a documented chain of custody built in from the first byte. No shipped devices, no onsite technicians, no guessing what got pulled off a phone down the road.

See what defensible, remote, targeted collection looks like for your team. Book a demo →