A court just showed Congress the AI loophole in child sex abuse material law

A federal appeals court has just told Congress, in unusually direct terms, that the law has fallen behind the technology it was written to police. On Aug. 25, a three-judge panel of the U.S. Court of Appeals for the Seventh Circuit ruled that federal prosecutors could not charge a Wisconsin man with a crime for privately possessing, inside his own home, obscene AI-generated images of children that the government itself conceded did not depict any real child. The panel said it reached that result reluctantly, bound by a quarter-century-old Supreme Court precedent decided before generative image models existed in any recognizable form. The ruling does not legalize producing, distributing, or sending such material to anyone, and it leaves three far more serious charges against the defendant intact. But it exposes, with unusual clarity, a hole in federal law that artificial intelligence has made wide enough to matter — and that only Congress or the Supreme Court can now close.

What the court actually decided

The case, United States v. Anderegg, involves Steven Anderegg, a software engineer who prosecutors say used the text-to-image model Stable Diffusion, modified with add-on components, to generate images of prepubescent children in sexually explicit scenarios and sent at least one to a minor on Instagram. He was indicted on four counts: producing, distributing, and transferring such material to a minor, plus a fourth count for simply possessing it at home. The district court dismissed only the possession count, and the government's interlocutory appeal challenged that single dismissal. Writing for the panel, Judge John Lee held that the possession charge, brought under 18 U.S.C. § 1466A(b)(1), was unconstitutional as applied to Anderegg specifically because the images could not be linked to an actual child and never left his house. The other three counts were not before the court and remain pending in the Western District of Wisconsin. Judge Lee did not hide his discomfort with the outcome:

Given the ever-accelerating march of imaging-generation technology, we have some misgivings about applying Free Speech Coalition here, but "unless we wish anarchy to prevail within the federal judicial system," we are dutybound to follow it.

A 2002 precedent the panel says AI has outrun

The precedent in question is Ashcroft v. Free Speech Coalition, in which the Supreme Court struck down a 1996 law banning "virtual" child pornography that used no real children, reasoning that such images lack the direct link to child abuse that justifies treating real CSAM as categorically unprotected speech. Combined with Stanley v. Georgia, the 1969 case establishing a First Amendment right to possess obscene material in one's own home, the two rulings leave a narrow but real opening: obscene, fully synthetic images of children, never shown to anyone, kept in a private residence. Congress tried to close part of that opening with the 2003 PROTECT Act, which added 18 U.S.C. § 1466A, criminalizing obscene depictions of minors "of any kind," explicitly including computer-generated images, whether or not the minor depicted actually exists. But the panel found that statute cannot survive an as-applied challenge on facts like Anderegg's, because the government's justifications — that such images groom victims, normalize abuse, or displace demand for real CSAM — are the same arguments the Supreme Court already rejected in 2002. In a concurrence, Judge Lee, joined by Judge Joshua Kolar, went further, urging the Supreme Court to revisit Free Speech Coalition altogether, noting that Justices Sandra Day O'Connor and Clarence Thomas warned in 2002 that computer-generated images might one day become "virtually indistinguishable" from real ones — and writing that "that day may have arrived."

A gap that runs through the states, too

The stakes extend well beyond one Wisconsin prosecution. Because Stanley and Free Speech Coalition are constitutional rulings binding on state as well as federal law, the same reasoning could unravel state statutes built on the same theory. According to a survey published this summer by the National Association of Attorneys General, at least 30 states have amended their child-pornography statutes to explicitly cover AI-generated material, and the Anderegg ruling is now the clearest signal those laws face the same constitutional vulnerability for purely private, non-distributed possession. Legal analysts tracking the case, including the Volokh Conspiracy at Reason, have flagged the decision as likely to be cited in similar challenges nationwide, given that no other circuit has yet confronted the identical question. The National Center for Missing and Exploited Children, which runs the federally designated CyberTipline that first flagged the Anderegg images through a Meta referral, has separately reported that AI-generated reports to its tipline are straining an operation built to identify real, at-risk children — a burden the Seventh Circuit's ruling does nothing to relieve and arguably compounds, since it narrows what law enforcement can charge once such images are found.

The fix belongs to Congress, and it already has a draft

Judge Lee's opinion is, in effect, an invitation the legislative branch should accept rather than defer. A bipartisan group of senators — John Cornyn, Richard Blumenthal, Mike Lee, and John Kennedy — introduced the ENFORCE Act last October specifically to align penalties for AI-generated CSAM and obscenity with those for material involving real children, with early backing from NCMEC, RAINN, and the National District Attorneys Association. That bill, along with a companion House measure, has sat without a floor vote while the Seventh Circuit's ruling makes the case for passing one more urgent, not less. Congress cannot simply reenact the 1996 ban Free Speech Coalition struck down; the Supreme Court was explicit that a statute reaching beyond obscenity, with no evidentiary link between synthetic images and real-world harm, will not survive First Amendment scrutiny. But it can do what the Seventh Circuit's concurrence effectively asked for: build a record. Evidence that generative models are trained in part on real CSAM, that AI-generated reports are consuming the same investigative capacity used to identify actual child victims, and that synthetic and real material are treated by offenders as interchangeable would supply exactly the "significantly stronger, more direct connection" the Supreme Court said was missing in 2002. Absent that record and a statute built on it, the current federal law will keep losing the narrow but real category of cases the Seventh Circuit just carved out — not because judges want it to, but because the Constitution, as written a generation before generative AI existed, currently allows nothing else.