AI used to identify miscreant judge

As soon as we flagged an unnamed federal judge having sex in chambers as part of an extramarital affair with a “high ranking law enforcement officer,” everyone asked the immediate follow-up, “how is a federal judge unable to afford a hotel room?” Followed soon after by, “who is the anonymous judge?” Because despite the severity of the allegations — an affair that raised serious blackmail risks, attending openly partisan events, and lying to investigators when caught — the Eleventh Circuit and the Judicial Conference both concealed the judge’s identity. They even adjusted the very minor sanction to allow the judge “to word the letters of apology vaguely so as to ensure that a letter could not be ‘used against [the Subject Judge] in some way.’”

Within 45 minutes of publishing our article, we worked out with a very high degree of confidence that it was Judge Eleanor Ross of the Northern District of Georgia. That said, we didn’t have a source with first-hand knowledge to confirm the story. While we made coy allusions, we were not prepared to publish this without more.

Bloomberg Law News has just confirmed with a source directly familiar with the investigation that it is Judge Ross.

The Eleventh Circuit thought it had been so clever in anonymizing its report. The reports don’t include a name or a district, and refer only to “Subject Judge” throughout. The reports even assiduously avoid identifying the judge by gender, proving that even conservative judges can figure out how pronouns work with minimal effort. And yet the reports failed to obscure a number of details that made working out the judge’s identity possible. This is where the story is going to pivot and we’re going to talk about AI and how the federal courts are not ready for it.

As soon as I read the reports, I zeroed in on a footnote revealing that the district judge involved hired clerks for two-year staggered terms. That’s not the norm and narrowed down the possible judges out of the gate. But as I tried to scour OSCAR, a former Above the Law columnist reached out to remind me that we live in 2026 and AI exists. Handing the reports into two different AI models and turning on all the “deep research” modes, the bots churned for several minutes comparing the reports to publicly available information. Both models delivered lengthy reports reaching the same conclusion. So how did these models do it?

In addition to the two-year clerk fact, the models instantly filtered out the entire state of Florida. The official reports are littered with references, in varying contexts, to the office of “District Attorney.” Florida uses “State Attorneys” for its local prosecutors. After that, the bots noted that the sanction barred the judge from ever serving as chief judge of their district — meaning the judge was not senior status and not currently the chief judge. The report indicates that investigators spoke with clerks dating back to 2020, disqualifying anyone elevated after that. Discussing the judge attending a DA’s primary victory party, the bot pointed out that the judge had claimed to know the candidate based on their time at the office, narrowing the scope to judges with state prosecutorial experience who overlapped with a sitting DA who won a primary. And had martinis at the victory party. The AI models decided that matched with Atlanta’s Fani Willis.

Once it narrowed the list down, the bot also searched the dockets of possible judges to match the claim in the reports that the high-ranking law enforcement officer did not materialize into a conflict because no cases involving that police department showed up on the judge’s docket.

For good measure, the bot went ahead and took a guess at the officer’s identity too.

In about 10 minutes of work, the AI unraveled all the work these judges put in to keep this confidential. With nothing but a couple of published court documents and the open web. In the time someone might brew a cup of coffee, the most basic possible workflow defeated the Eleventh Circuit’s entire anonymization strategy.

In 2000, Harvard professor Latanya Sweeney famously showed that 87% of Americans could be uniquely identified by the combination of ZIP code, gender, and date of birth. The bombshell wasn’t that re-identification was possible — statisticians had been saying that for years — it was that institutions kept releasing “anonymized” data as if a name redaction can ward off identification. A recent study from ETH Zurich and Anthropic updated this for the generative AI era and found that LLMs can deanonymize pseudonymous online users at roughly $1 to $4 per match, correctly identifying 67% of Hacker News users from their writing alone after every direct identifier had been stripped. What used to take a dedicated human investigator hours takes a model minutes. What used to be “practically obscure” is now translucent at best.

If we take the reports at their word, the judges earnestly did not want to unmask the judge involved. That means a whole lot of federal judges are laboring under the misconception that they’re more clever than AI. While most of the unflattering commentary around this affair will be directed at the subject judge, save a good deal of ire for the judges running this probe who seem to have zero clue how AI works. Anyone with a $20 subscription and a rudimentary sense of how this technology works could replicate this result.

For lawyers, this should underscore the importance of getting up to speed on this technology. Whether you’re trying to maintain your client’s confidentiality or hoping to crack a mystery buried in the diligence materials, if you’re not using AI for these tasks, you’re failing. We have a lot of fun pointing and laughing when lawyers screw up redactions. We’re now at a point where the whole redaction strategy must be rethought to guarantee otherwise innocuous breadcrumbs don’t give AI enough material to decode the whole thing. The duty of competence has had a comment about technology for over a decade. We spent most of that decade arguing about whether that means lawyers need to know what “the cloud” is. We are well past that.

The legal AI conversation in 2026 has been dominated by the hallucination beat — Gordon Rees, Sullivan & Cromwell, Butler Snow… Gordon Rees again. But this obscures the other half of the AI story: competence isn’t just about not using AI badly, but also includes knowing how it can be used effectively.

Earlier: Federal Judge Had Sex In Chambers Bringing New Meaning To Gavel Bang

Joe Patrice is a senior editor at Above the Law and co-host of Thinking Like A Lawyer. Feel free to email any tips, questions, or comments. Follow him on Twitter or Bluesky if you’re interested in law, politics, and a healthy dose of college sports news.