There is a noodle shop in Mei Foo where I do most of my thinking, which is not the same as most of my writing but is often a prerequisite. The tables are small and the turnover is fast, and nobody cares if you sit there reading case law on your phone as long as you order something. I was reading the Seventh Circuit’s opinion in United States v. Anderegg on a Monday evening, and Ruth was across from me working through a bowl of dan dan mien, and at some point I must have made a noise because she looked up and said, “That bad?”
I showed her the headline. She read it, put her phone down, picked her chopsticks back up, and said, “Well, the law is only as smart as the people who wrote it.” Then she went back to her noodles.
She is rarely wrong about these things, which is inconvenient when you write about them for a living, because she can usually land in one sentence what takes me several thousand words.
But here are the several thousand words, because this one matters.
On August 25, 2026, Judge John Z. Lee of the U.S. Court of Appeals for the Seventh Circuit ruled that the First Amendment of the United States Constitution protects a person’s right to privately possess AI-generated child sexual abuse material in the home, provided the images do not depict a real, identifiable child. The case involved Steven Anderegg, a 42-year-old man from Holmen, Wisconsin, who used Stable Diffusion to generate roughly 13,000 hyper-realistic images of prepubescent children engaged in sexually explicit acts. He allegedly sent some of them to a 15-year-old boy over Instagram. He was charged with production, distribution, possession, and transfer. The district court dismissed the possession charge. The Seventh Circuit affirmed.
And then Judge Lee did something unusual. He wrote a separate concurrence, joined by Judge Joshua Kolar, in which he acknowledged that the technology has advanced so far beyond what the Supreme Court imagined in 2002 that the precedent his court was bound to follow no longer maps onto reality. He urged the Supreme Court to step in. A federal appellate judge affirmed something as constitutional and then publicly begged a higher court to make it unconstitutional. That sentence should not be possible. It is a description of the current state of American law on what ought to be the most unambiguous moral question a legal system can face.
The precedent trap works like this. In 1969, the Supreme Court decided Stanley v. Georgia and held that a person has a First Amendment right to possess obscene material in the privacy of his own home. In 2002, Ashcroft v. Free Speech Coalition struck down parts of the Child Pornography Prevention Act, ruling that sexually explicit images of children that do not depict actual children are not child pornography under the law and therefore receive First Amendment protection. The reasoning was that without a real child victim, the government’s interest in prohibition was insufficient to override free speech. Justice O’Connor, concurring in part, warned that the technology might eventually make this distinction meaningless. Justice Thomas said explicitly that the government could revisit the question if virtual images became indistinguishable from photographs of real abuse.
In 2002, “virtual” child pornography meant blocky CGI that nobody would confuse with a photograph. It was easy to be principled about speech categories when the speech in question looked like a bad video game. Stable Diffusion does not look like a bad video game. It generates photorealistic imagery that trained forensic analysts struggle to distinguish from documented abuse. The legal framework has not moved. The Supreme Court has not revisited Ashcroft. And so the Seventh Circuit did what lower courts do when bound by precedent that no longer fits the world: it followed the rule and begged someone else to fix it.
I read these things from Hong Kong, and the distance is clarifying.
Hong Kong’s Prevention of Child Pornography Ordinance, Cap. 579, criminalizes the production, possession, distribution, and publication of child pornography. It explicitly covers computer-generated images. There is no carve-out for material that does not depict an identifiable real child, because the legal framework treats the material itself as the harm, not merely the process by which it was created. Possession alone is an offence. Full stop. The law is not perfect, and enforcement is a different conversation, but the basic architecture does not require a judge to write an anguished concurrence explaining why he cannot protect children from the thing he just ruled is constitutional.
This is not unique to Hong Kong. South Korea criminalizes virtual child exploitation material. China’s regulatory framework for AI-generated content treats CSAM as categorically illegal regardless of method. European Union member states’ criminal codes broadly ban the material without the American distinction between real and virtual. Japan is a notable exception, its child pornography law explicitly excludes purely fictional or drawn depictions, a carve-out that has drawn criticism from international child protection organizations for years, though that exception developed around manga and anime, not photorealistic AI output designed to fool forensic examiners.
The American system’s inability to close this gap is not an accident. It is the logical endpoint of a legal philosophy that treats speech as so categorically sacred that the framework cannot adapt when the category of “speech” expands to include photorealistic depictions of child sexual abuse generated by a text prompt. But the sacredness is selective, and everybody knows it. The United States charged 1,575 people in connection with storming the Capitol on January 6, 2021, over a thousand of whom pleaded guilty or were convicted, before the incoming president pardoned nearly all of them on his first day in office. A PBS News/NPR/Marist poll in October 2025 found that 30 percent of Americans, nearly one in three, believe people may have to resort to violence to get the country back on track. This is a nation where a meaningful share of the population has made peace with political violence as a tool, where the legal system proved entirely capable of prosecuting political expression when it wanted to, and where the same constitutional framework now says it cannot prohibit a man from generating 13,000 photorealistic images of children being sexually abused on his laptop because that would be a speech violation.
Nobody sat down and decided this should be the outcome. It is the residue of a constitutional structure that was not designed for a world where a laptop and an open-source model can produce, in an afternoon, a volume of material that would have required an international criminal network and years of real-world abuse to create a decade ago. The question is not whether the First Amendment is valuable. Of course it is. The question is whether a legal system that can mobilize to prosecute a riot but not to prohibit synthetic child abuse material still has a coherent theory of what it is protecting, and from whom.
The production charges against Anderegg will proceed. The distribution charges will proceed. The charge for allegedly sending material to a minor will proceed. He may go to prison for a long time. The Seventh Circuit’s ruling is narrow, and the sky has not fallen, and all the usual qualifications apply. But the precedent is now written. Defense attorneys in every federal circuit now have a Seventh Circuit opinion holding that Stanley and Ashcroft together create a constitutional right to possess AI-generated CSAM in the home. The Department of Justice must decide whether to seek rehearing en banc or petition the Supreme Court for certiorari. If it does, the Court will have to confront the question Justices O’Connor and Thomas warned about 24 years ago.
Can the state prohibit the possession of material that is, for all practical purposes, identical to documented child sexual abuse, on the grounds that it fuels demand for real abuse, degrades the ability of law enforcement to identify real victims, and normalizes the sexualization of children, even if no specific child was harmed in its creation?
Most other advanced legal systems have answered yes. The United States has not. And a federal judge is on the record saying he wishes it would.
Forty-six states have tried to legislate around this. Congress passed the TAKE IT DOWN Act in May 2025, which criminalizes the publication of nonconsensual intimate imagery, including AI-generated material, and secured its first conviction in April 2026 when an Ohio man pleaded guilty to creating and distributing deepfake intimate images of women he knew. But the TAKE IT DOWN Act targets publication, not possession. And state laws criminalizing AI-generated CSAM may now face constitutional challenges under the logic the Seventh Circuit just laid out. The patchwork is the problem. Fifty states trying to build walls around a gap that the Supreme Court could close in a single opinion, if it chose to hear the case, if it chose to revisit Ashcroft, if it chose to acknowledge that the world Justice Kennedy described in 2002 no longer exists.
I live in a city where the rule of law is a subject of active, sometimes painful, conversation. Hong Kong has its own fractures, its own arguments about what the law protects and what it permits and where those two things diverge. I am not holding it up as a model of anything except this: when the question is whether photorealistic AI-generated images of children being sexually abused should be legal to possess, Hong Kong’s answer is no, and it does not require a judge to apologize for giving it.
The gap between what a society knows is wrong and what its legal system can do about it is not a sign of principled restraint. It is a sign that the system has fallen behind the world it governs. And from a noodle shop in Mei Foo, where my wife has already finished her dan dan mien and is checking her phone and has no interest in reading another 500 words about constitutional originalism, the decay is visible from a very long way off.
Ruth would say it shorter. She usually does.