Court Decision Shines Light on AI

The government cannot prosecute possession of obscene AI-generated child pornography. This was the conclusion of a reluctant Seventh Circuit last week in a case that tested the Justice Department’s authority in an AI world. Although shocking, the result was predictable, as the judges’ hands were tied by decades-old Supreme Court precedents. The decision invites the justices to revisit these older cases, which have aged poorly and handicap prosecutors seeking to stamp out the scourge of child sexual abuse imagery (CSAM). This is the government’s first significant appellate decision involving AI-generated CSAM. Defendant Steven Anderegg allegedly produced several hyperrealistic images of prepubescent children engaged in explicit sexual acts and sent them via Instagram direct message to a minor’s account. Anderegg was charged with producing, distributing, and possessing CSAM in violation of various federal laws. The trial court granted Anderegg’s motion to dismiss the possession charge, and the Seventh Circuit upheld the dismissal. As I predicted a few years ago, cases like Anderegg’s are shaped by the confluence of several Supreme Court decisions going back nearly half a century. The Court has long held that obscenity is not protected by the First Amendment. But there is no nationwide definition of obscenity; under Miller v. California, the term is defined by local community standards. Moreover, Stanley v. Georgia held that the government could not prosecute over mere possession of obscene materials in one’s own home, as the government cannot “control the moral content of one’s thoughts.” The Court has made exceptions for CSAM: In Ferber v. New York, the Court held that CSAM could be prosecuted whether it is obscene or not, and in Osborne v. Ohio, the Court extended this to mere possession of CSAM, because such material constitutes a permanent record of criminal abuse of a child. But importantly, in Ashcroft v. Free Speech Coalition, the Court limited Ferber and Osborne to imagery involving actual children. This means that nonobscene virtual CSAM is protected by the First Amendment, and possession of even obscene virtual CSAM is protected under the ruling made in Stanley. I remain convinced that the Supreme Court should reconsider Ashcroft in the AI age. Judges Milton C. Lee and Joshua P. Kolar agree, writing in a concurrence that “given the significant advances in computer-generated image technology, particularly in the field of artificial intelligence, we would benefit from additional guidance from the Supreme Court regarding the intersection between the First Amendment and virtual CSAM.” Ashcroft was decided in 2002, when virtual images were both difficult to generate and low quality compared to real pictures. Today, AI-generated imagery indistinguishable from reality can be produced quickly and inexpensively. While companies like OpenAI and Midjourney have safeguards to prevent misuse, these safeguards can be bypassed for models residing on a user’s local machine, as Anderegg allegedly did with Stable Diffusion. This increased quality and quantity of AI-generated CSAM make it harder to prosecute actual CSAM cases. To detect CSAM, platforms and other entities rely on hash values, unique digital fingerprints that allow systems to compare files without looking at images. The National Center for Missing and Exploited Children (NCMEC) maintains the nation’s authoritative database of known CSAM hash values. In 2024, 4,700 AI-generated CSAM images were reported to NCMEC. By 2025, that number had risen to over 400,000 images. And it is likely to continue: One study estimates that in an eleven-day period, users prompted Grok to create over 3 million sexualized images, over 23,000 of which constituted CSAM. This explosive growth in AI-generated CSAM makes it harder for NCMEC and law enforcement to detect actual CSAM, particularly in cases where an abuser has used AI to make a real image look computer-generated. It also makes it harder to convict perpetrators, as defendants can create reasonable doubt by suggesting actual images are AI-generated. In Ashcroft, three justices would have found that this burden on prosecutors in actual CSAM cases was sufficient to uphold a virtual CSAM ban. A fourth, Justice Clarence Thomas, found this argument persuasive but speculative, as no defendant had successfully raised a “computer-generated” defense in 2002. The AI revolution makes such a defense inevitable. The Supreme Court has shown an interest in revising its early-internet-era speech decisions. Last year it upheld a Texas age verification law for online pornography despite effectively killing similar federal laws in 1997 and 2004. In doing so, the Court explained that the “technology of the internet has continued to evolve at a rapid pace” and now makes available “vast libraries of video content—both benign and obscene—at almost any time and place, with an ease that would have been unimaginable at the time” of those earlier decisions. The advent of AI-generated CSAM at scale was similarly unimaginable to the Ashcroft court and, at a minimum, warrants reconsideration of the rationales underlying that decision.