AI FactScan
You collect the case names, docket numbers, sanctions amounts, and quotes. You hand them to an AI and ask it to draft the article.
The draft comes back complete. It has structure. It has transitions. It names the cases with confidence.
Several details are wrong.
Not invented from nothing. The cases exist somewhere in the search results. But a sanctions amount gets attached to the wrong matter. A quote comes from a secondary summary, not the ruling itself. One description sounds right because it was assembled from nearby facts.
The model does not stop and say: you need the original order before publishing this. It gives the finished version.
That is the failure mode. AI is good at producing the shape of verified work before verification has happened.
Researchers sometimes discuss one version of this problem as reinforcement learning from AI feedback, or RLAIF. The plain version is easier to recognize: the model has learned to deliver the answer that looks finished, not the answer that has been verified. Completeness, fluency, and confidence often score well. Slow verification does not.
We caught the errors because someone went back to the primary sources.
The same failure mode is now showing up in courtrooms, where the cost is not a bad draft. It can become a filing, a sanctions order, or a lost motion.
AI lowers the barrier to filing. It does not lower the duty to check.
The appeal is easy to understand. If you cannot afford a lawyer and believe you have a legitimate claim, AI lowers the barrier enough to try.
That part is not irrational. Legal help is expensive, and many people already navigate courts without counsel.
A 2026 paper, The New Pro Se, found that federal civil pro se plaintiff filings rose after widespread generative AI adoption. It also found that filings with AI markers were dismissed more often, a useful reminder that AI can help someone draft a complaint without making the case stronger. The paper does not prove AI caused every filing. It does show why this issue matters: more people are using AI at the edge of a system that still runs on rules, records, and citations.
AI may help someone get words onto the page. The problem starts when those words include legal authorities.
The lawyer case: Mata v. Avianca
The standard example is Mata v. Avianca.
Lawyers submitted a brief that included fake cases generated by ChatGPT. The court sanctioned them and imposed a $5,000 penalty. The important part is not simply that ChatGPT made things up. It is that the brief treated those cases as if they had been checked.
The court did not ask whether the fake citations were convenient. It asked whether lawyers had met their duty before filing.
The pro se case: BFG v. Pierce RE Holdings
The same pattern is not limited to lawyers.
In BFG Corporation d/b/a Byline Financial Group v. Pierce RE Holdings, LLC, 2026 WL 1878426, filed June 30, 2026 in the Northern District of Illinois, the defendants represented themselves. In a motion to reconsider, they accused the court of error and relied on cases the court said could not be found or did not support the propositions cited.
The court described the citations as hallmarks of generative artificial intelligence. It also reminded the defendants that there is no pro se exception to Rule 11.
The motion to reconsider was denied. The plaintiff's motion for summary judgment was granted.
That is the part worth sitting with. The defendants were self-represented, but the filing still had to survive contact with the record.
What to check before anything goes to court
Start with the case citation. Search the exact case name and citation in a legal database, Google Scholar, CourtListener, or the court's own website. If the case does not appear, do not assume the search engine is the problem. Treat it as unverified until you can locate the actual opinion.
Then read the opinion itself. Do not rely on an AI summary, a search snippet, or a quote card. Confirm that the case says what your filing claims it says.
Check the court and date. A real case from the wrong court may be weak support. A real case from the wrong jurisdiction may not help you. A real case that has been overruled may hurt you.
Check quotations line by line. Fake legal citations are bad. Fake quotations from real cases are harder to catch and easier to overtrust.
Save the sources. Keep PDFs, screenshots, or links to the actual opinions you relied on. If the other side or the judge asks where a case came from, "AI gave it to me" will not be enough.
Where AI FactScan fits
AI FactScan is not a lawyer and does not tell you whether a legal argument is good.
It can help with an earlier layer: checking the sources inside an AI answer before you trust them. If an AI response gives you links, laws, agency pages, reports, or other web sources, AI FactScan helps flag sources that need a closer look while you are still in the chat window.
For court cases, you still need to verify the actual legal authority in a legal database or court source.
That is the point. AI can make drafting faster. It cannot take over the duty to check what gets filed.
A faster draft is useful only if the citations survive contact with the record.