Together with Cyndie Demeocq at U of Edinburgh, today I’m releasing a paper analyzing 71 federal criminal cases, filed around the US from 2022-2025, that somehow involve AI-generated or -morphed CSAM or child obscenity. It's available on SSRN here.
We’re also releasing the dataset of cases (with an appendix of 2026 cases) on GitHub. Our hope is that this data will be useful to others. I believe this is the largest public collection of federal AI-CSAM cases to date (though I assume there’s an internal list at DOJ).
Our top-line finding: existing laws are working just fine in AI-CSAM prosecutions, at least in the cases prosecutors chose to bring. (We can’t know what cases they decided not to bring, or why.) Most defendants pled guilty (as usual); the few who went to trial all got convicted. Prison sentences are longer than average for CSAM or obscenity cases. Out of 71 cases studied, only half a dozen are still pending.
What about the First Amendment? There has been lots of chatter lately about the Seventh Circuit’s recent ruling in a case called U.S. v. Anderegg. But we show that case is an outlier. Only 8 defendants even raised 1A defenses at all, Anderegg’s the only one who won, and even he only won on one count; he still has to face three others. The Seventh Circuit held that the 1A protects private possession at home of fully AI-generated obscenity of imaginary kids. But the cases we study (even the Anderegg lower court) are clear: everything else – production, downloading images, possession outside the home, morphed images of real kids – is NOT protected. So: We do not need the Supreme Court to overrule its existing precedent. We do not need a new federal law. If you’re mad about the Seventh Circuit ruling and you think people who engage with AI-CSAM belong in prison, this paper has some great news for you.
Since federal law is already bringing AI-CSAM offenders to justice (at least in the cases DOJ chose to bring), our recommendations focus instead on (1) prevention (by AI model developers & AI model hosting platforms) and (2) begging prosecutors to please put more detail in court filings so I can tell if a case is about AI or not.
Finally: screw PACER fees. I spent over $1000 of my research budget on this project. That’s at $0.10/page – it’ll go up to $0.12 on Jan. 1. I’m not eligible for the fee exemption for research because it’s conditioned on not posting the court records you get on the internet. No way. CourtListener forever.
Please read the paper, please make use of the dataset, please send them to the federal criminal defense lawyer or prosecutor in your life, please send me cases I missed. And if you’re an articles editor at a law journal who’s still looking to fill all your spots for the August submissions cycle: this (very timely) piece is yours for the taking.