When a federal appeals court declares a criminal statute unconstitutional, the headline practically writes itself. That is exactly what happened when a Seventh Circuit panel addressed a federal prosecution involving images generated entirely by artificial intelligence. On first read, it looks like a landmark defense victory: the court concluded that a federal statute could not constitutionally reach one defendant’s private, in-home possession of obscene virtual material.
But headlines are lousy at handling qualifiers, and this opinion is stuffed with them. The ruling is as applied. It covers one count out of four. It arrives with the panel practically apologizing for its own reasoning. And two of the three judges wrote separately to ask the Supreme Court to come fix the doctrine they had just been forced to apply.
This piece walks through what the court actually decided, why it decided that way, and — more importantly for anyone tracking where this area of law is heading — what the opinion quietly tells the next prosecutor to do differently.
The Case in Its Actual Posture
The defendant, Anderegg, faced a four-count federal indictment. Counts one through three alleged producing and distributing visual depictions of minors in sexually explicit conduct, plus transferring such material to a minor under sixteen. Count four was a straight possession charge.
He moved to dismiss the entire indictment. The chief judge of the Western District of Wisconsin granted that motion as to the possession count and denied it as to everything else. The district court held that the possession provision was unconstitutional as applied under Stanley v. Georgia, the 1969 Supreme Court decision recognizing a First Amendment right to possess obscene material privately at home.
The government then filed an interlocutory appeal. That term matters. Ordinarily, neither side can appeal until a case reaches final judgment. An interlocutory appeal is a narrow exception permitting review of an interim pretrial ruling. So the Seventh Circuit was looking at count four and nothing else. Counts one, two, and three survived and are headed to trial.
The practical takeaway: the defendant won something real, but he is still facing production, distribution, and transfer charges that carry serious exposure. Dismissing the possession count changed very little for him personally. It changed a great deal for the law.
How This Investigation Actually Started
There was no elaborate federal task force here. In October 2023, Meta filed a report with the cyber tip line at the National Center for Missing and Exploited Children. An Instagram user had sent a direct message containing apparently AI-generated images to an account belonging to a minor.
The detail that makes the case memorable: according to the opinion, the message history showed the sender explaining to the minor how he made the images — by typing text prompts into Stable Diffusion, a text-to-image generative model that runs on ordinary consumer hardware.
Investigators obtained a search warrant covering the home, a laptop, two cell phones, and other devices. The government alleged the defendant installed Stable Diffusion, added software extending what it could render, used curated prompts, and refined output using negative prompts to exclude depictions of adults.
A negative prompt is simply an instruction telling the model what not to generate. If you tell an image model “never depict adults,” everything it produces skews younger. That allegation — deliberately engineering adults out of the output — is what the government says demonstrates intent.
The Concession That Carried the Whole Case
Buried in footnote two is the sentence everything else rests on. The government conceded that the charged images do not depict an actual child and cannot be linked to any actual child. The footnote cites the oral argument recording. In open court, prosecutors said these are not real children.
That concession is load-bearing because the entire architecture of child pornography law is built on the presence of a real child. Take the real child out and the constitutional analysis changes completely.
Four Supreme Court Cases That Frame Everything
The panel walked carefully through the precedent, and the map is worth having.
1. Stanley v. Georgia (1969) — The Home Possession Shield
Police found three reels of obscene film in a man’s home and he was convicted of possession. The Supreme Court reversed, holding that state power does not extend to mere possession in the privacy of one’s own home. The quotable line: if the First Amendment means anything, the state has no business telling a man sitting alone in his house what books he may read and what films he may watch.
2. New York v. Ferber (1982) — The Real Child Carve-Out
Here the Court removed child pornography from First Amendment protection whether or not it meets the obscenity test. The reasoning is grounded in the real child: the material is a permanent record of the child’s participation in the harm, and the child’s trauma is worsened by its circulation. That logic is not arbitrary — it follows from an actual victim.
3. Osborne v. Ohio (1990) — The Market Rationale
Osborne possessed photographs of actual adolescents at home. Stanley did not save him. The Court explained that Ohio was not policing his mind paternalistically; it was trying to destroy the market for the exploitative use of children. Real kids, real market, real interest in cutting off demand at every link in the chain.
4. Ashcroft v. Free Speech Coalition (2002) — Virtual Material
Congress passed the Child Pornography Prevention Act in 1996, extending the ban to computer-generated images and images that merely appear to show minors. The Supreme Court struck that down. Where no real child is involved, the child-protection rationale simply does not apply. Virtual material is not intrinsically related to the sexual abuse of children the way the material in Ferber was.
Apply that map to the facts here and the outcome follows almost mechanically. No real child means Ferber and Osborne drop off the table. What remains is Stanley, which protects private possession, and Free Speech Coalition, which rejected the government’s justifications for banning virtual material.
Three Government Arguments and Why Each Failed
The Grooming Argument — and a Twenty-Three-Year-Old Brief
The government argued that offenders use this material to groom minors, and that risk exists whether the imagery shows real or computer-generated children. The panel did not dispute that grooming is real; it cited cases and congressional findings acknowledging it.
But the panel noted the Supreme Court flatly rejected that same argument in Free Speech Coalition: the government may not prohibit speech because it increases the chance an awful act will be committed at some indefinite future time.
Then came the part that made the hosts laugh. The panel set the government’s current brief beside the brief it filed and lost in 2001. The 2001 version argued that child pornography is used to seduce children into sexual activity and that computer-generated images can do the same. The current version argued that offenders commonly use similar images to groom minors. Same argument, essentially the same wording, twenty-three years apart. A footnote observed that instead of offering additional support, the government merely restated what it already lost on.
The Interstate Commerce Hook
The government argued this was not a lone man at home like Stanley, because conviction requires proving the material or the computer traveled in interstate commerce — and the defendant downloaded Stable Diffusion over the internet.
The panel was unimpressed. The obscene films in Stanley also moved in interstate commerce, and it did not matter. Once material leaves the public sphere and enters the home, governmental regulatory power is eclipsed by the individual’s First Amendment right to receive information and ideas regardless of its social worth. The commerce nexus ended up in a footnote as a jurisdictional afterthought.
Reading Stanley Narrowly
The last attempt argued Stanley protects only material depicting adults, leaning on language from Williams noting that government may criminalize possession of child pornography even though it may not criminalize possession of obscene material involving adults.
The panel called that an overreading of a single sentence. Stanley‘s reasoning never turned on the content of the material. It turned on the location — the home.
Why the Win Is Fragile
Here is the uncomfortable part for anyone tempted to celebrate.
It hangs on one concession. Expect future prosecutors to say they cannot determine whether the depictions correspond to real minors, rather than conceding the point outright.
Morphed images are a separate category. Free Speech Coalition itself distinguished morphing provisions in a footnote: morphed images implicate the interests of real children and sit closer to Ferber. Take a real child’s ordinary photograph and alter it into something explicit, and the analysis snaps back to Ferber and Osborne. This case came out the way it did only because the images were built from scratch with no real child anywhere in the chain.
The panel did not believe in its own result. Early in the opinion the judges wrote that they have concerns about the lines these cases draw but are not free to redraw them. Later came the money line: they have misgivings about applying Free Speech Coalition here, but unless they wish anarchy to prevail within the federal judicial system they are duty bound to follow it. They cited Hutto v. Davis for the proposition that precedent must be followed no matter how misguided lower court judges think it is.
That is compliance, not conviction.
The Concurrence Is a Roadmap for the Other Side
Two of the three judges wrote separately. Their concurrence does not defend the holding — it asks the Supreme Court to revisit it, noting that given significant advances in computer-generated image technology, the courts would benefit from additional guidance on the intersection of the First Amendment and virtual material.
It goes further. It quotes Justice O’Connor and Justice Thomas from Free Speech Coalition, including Thomas’s observation that if technological advances thwart prosecution of unlawful speech, government may well have a compelling interest in barring otherwise lawful speech to effectively enforce laws against material made through the abuse of real children.
Then it flags the argument the government did not make: a major dataset used to train image-generating AI contained hundreds of abuse images, and the market for AI-generated material may affect demand for the real material these models train on. Such a finding, the concurrence noted, would implicate many of the concerns raised in Osborne.
That is the crack in the wall. If a prosecutor can build a record connecting generated output back to training data derived from real abuse, the case stops being a Free Speech Coalition case and becomes an Osborne market case. The concurrence conceded the record here was insufficient to evaluate those complex issues — which is judicial shorthand for someone should develop this record next time.
The Familiar Legislative Cycle
The opinion narrates the last time this happened. Congress passed the CPPA in 1996. The Supreme Court struck it down in 2002. Congress went back to the drawing board and enacted the PROTECT Act in 2003.
The pattern is durable: lose, redraft, return with more findings. Expect new congressional findings about morphing, about training datasets, about whatever theory looks most likely to survive the next challenge.
Does This Touch the Registry?
No. This is purely a First Amendment possession holding. The court concluded that the specific possession provision is unconstitutional as applied and on its face where no real child is involved. It is not a registry case, not a residency restriction case, not a supervision case.
Actionable Takeaways
- Read past the headline to the posture. “Statute held unconstitutional” and “defendant walks” are very different sentences. Here, three serious counts survive.
- Watch what the government concedes. A single admission at oral argument determined the constitutional outcome. Future concessions will be far more guarded.
- Treat the concurrence as a forecast. When two of three panel judges publicly request Supreme Court review and identify the untried argument, that materially raises the odds of en banc review or a cert petition — and of a different result down the road.
The Bottom Line
This is a narrow, somewhat hollow win. The dismissal of the possession count is solid and unlikely to be undone for this defendant. But the panel refused to redraw the Supreme Court’s lines, signaled discomfort throughout, and effectively invited the government upward.
The ground has been plowed before. The plow is rusting. And the government is already sharpening a new one.