A federal appeals court struck down the US ban on possessing AI-generated child rape media inside the home on 25 August, holding that the First Amendment protects such material where it depicts no real child.

The Seventh Circuit affirmed the dismissal of one count against Steven Anderegg, 42, of Holmen, Wisconsin, whose prosecution the Justice Department brought in May 2024 as its first federal case over child rape media generated entirely by artificial intelligence.

Judge John Lee, writing for a panel that also included Doris Pryor and Joshua Kolar, held the possession provision of 18 U.S.C. § 1466A unconstitutional as applied to Anderegg across 27 pages. Lee, joined by Kolar, filed a separate concurrence asking the Supreme Court to revisit the 2002 decision that controls the question.

Section 1466A(b)(1) is the provision that reaches a person who obtains obscene synthetic child rape media without producing it or passing it on. Prosecutors in the circuit covering Wisconsin, Illinois and Indiana no longer have that count available where the images depict no identifiable child and stay inside the defendant's home.

How the case began

Meta reported an Instagram direct message to the CyberTipline run by the National Center for Missing and Exploited Children in October 2023, after Anderegg sent sexually explicit AI-generated images to an account belonging to a 15-year-old boy and told the boy in the same exchange how he had produced it.

Investigators executing a search warrant found Stable Diffusion installed on Anderegg's laptop, modified with add-on components that gave the model the ability to render human genitalia. The Justice Department put the total haul at roughly 13,000 images. Hundreds of them showed nude or semi-clothed prepubescent minors displaying or touching their genitals, alongside the curated text prompts and the negative prompts Anderegg used to keep adult bodies out of the output.

A grand jury indicted him on 15 May 2024 on four counts under § 1466A: production, distribution, transfer of obscene material to a minor under 16 and possession. Prosecutors put his maximum exposure at 70 years with a five-year mandatory minimum, and he was detained pending a hearing on 22 May 2024.

"CSAM generated by AI is still CSAM, and we will hold accountable those who exploit AI to create obscene, abusive, and increasingly photorealistic images of children," then-Deputy Attorney General Lisa Monaco said when the charges were announced.

The government conceded during the litigation that the images depicted no actual child and could not be traced to one. The concession took Ferber and Osborne out of the case and left Stanley and Free Speech Coalition to decide it.

Chief Judge James Peterson of the Western District of Wisconsin dismissed the possession count on 13 February 2025 and left the other three standing. Prosecutors took an interlocutory appeal. William Clayman and Ross Goldman of the Justice Department's criminal division argued it on 30 October 2025 against Joseph Bugni of Hurley Burish, who represents Anderegg. The panel took just under 10 months to rule.

A statute written for the last technology

Congress passed § 1466A in 2003 as part of the PROTECT Act, drafted in direct response to the Supreme Court striking down the Child Pornography Prevention Act the year before. Subsection (c) states that it is not a required element that the minor depicted actually exist. Subsection (f)(1) reaches any "digital image or picture, computer image or picture, or computer generated image".

Obscenity is the ingredient Congress added to keep the statute inside the constitutional line, and § 1466A(b)(1) bites only where the material meets that standard. Lee assumed for the sake of argument that Anderegg's images did, then held the statute still could not reach them in his house.

Two precedents, four rejected arguments

Stanley v. Georgia, decided in 1969, bars prosecution for private possession of obscene material at home, on the reasoning that government has "no business telling a man, sitting alone in his own house" what books he may read or what films he may watch. Lee wrote that the holding "was predicated on the location, the home", rather than on the character of what was found there.

Ashcroft v. Free Speech Coalition, decided in 2002, struck down federal bans on computer-generated abuse imagery involving no real child. New York v. Ferber and Osborne v. Ohio, which place abuse imagery of real children outside First Amendment protection altogether, rest on the permanent record of an actual child's abuse and the economic motive that distribution supplies to producers. No real child appears in Anderegg's files, so the panel held that neither case reaches his conduct.

Prosecutors offered four reasons to override Stanley, and Lee traced each one to a passage in Free Speech Coalition rejecting it. On the government's case that offenders use such imagery to groom children, Lee quoted the Supreme Court's answer that government "may not prohibit speech because it increases the chance an unlawful act will be committed at some indefinite future time". On the case that synthetic material normalises abuse, the Supreme Court had demanded a significantly stronger and more direct connection to actual harm. On the case that virtual images substitute for the real market, the Supreme Court had called the theory implausible, since producers would switch to fictional images and escape prosecution altogether. On the case that AI output has become impossible to tell from photographs of real abuse, straining investigators, the Supreme Court had written that government "may not suppress lawful speech as the means to suppress unlawful speech".

Prosecutors also told the court that § 1466A(b)(1) survives because it reaches only material shipped or transported in interstate commerce, an element Stanley's statute lacked. Lee rejected the distinction on the ground that the eight-millimetre reels seized from Stanley's desk drawer had almost certainly crossed state lines as well, and treated the commerce element as jurisdictional rather than a description of anything wrongful.

A footnote set out federal sentencing data showing that nearly 44 per cent of possession offenders had engaged in aggravating conduct, alongside a caveat that the social science on whether viewing such imagery causes contact offending remains inconclusive. Another footnote flagged morphed images, which paste a real child's face onto another body, as sitting closer to Ferber. The panel did not decide that question.

Counts one through three remain live, and the opinion leaves untouched any prosecution for producing or sharing the same files.

The concurrence

Lee wrote separately, joined by Kolar, that the court would "benefit from additional guidance from the Supreme Court regarding the intersection between the First Amendment and virtual CSAM", and that in the years since Free Speech Coalition the technology has reached a point where AI output is virtually indistinguishable from images of physical abuse. Only in recent months, the concurrence said, have the tools become good enough to fool the human eye.

Justice Clarence Thomas anticipated the problem in 2002. Concurring in the judgment in Free Speech Coalition, he wrote that "if technological advances thwart prosecution of 'unlawful speech,' the Government may well have a compelling interest in barring or otherwise regulating some narrow category of 'lawful speech'". Justice Sandra Day O'Connor would have upheld the ban on virtual imagery indistinguishable from the real thing.

Lee cited testimony that the influx of synthetic material imposes growing burdens on investigators and lets offenders disguise pictures of real children by making the images look computer-generated, which slows the identification of actual victims among the files police recover.

Researchers at the Stanford Internet Observatory, led by chief technologist David Thiel, reported on 20 December 2023 that LAION-5B, the open dataset behind Stable Diffusion, contained 3,226 suspected instances of child sexual abuse material, of which 1,008 were externally validated. LAION withdrew the dataset and published a cleaned version, Re-LAION-5B, in August 2024. Demand for AI-generated material can therefore raise demand for the real abuse imagery the models train on, Lee wrote, which is the harm Osborne was decided to prevent.

21 million reports

NCMEC received 21.3 million CyberTipline reports in 2025. More than 400,000 of them carried what the organisation calls a generative AI nexus, and more than 182,000 involved offenders possessing, generating or attempting to generate AI abuse material. NCMEC has categorised 158,000 images and videos as AI-generated abuse material since 2023 and identified more than 275 victims of it.

Riana Pfefferkorn of Stanford, who predicted before the indictment that prosecutors would fall back on obscenity statutes for AI cases, wrote to NCMEC on 29 January 2026 to flag what those totals conceal. A Bloomberg investigation had found Amazon's 380,000 reports flagged as AI-related contained no AI-generated material at all, but known abuse imagery surfaced while scanning training data. Nearly 80 per cent of all generative AI CyberTipline reports in the first half of 2025 involved no AI-generated abuse material, Pfefferkorn wrote, and the single checkbox that produced the figure had misled lawmakers and researchers for six months. NCMEC itself notes that more than 10 per cent of industry reports arrive with inadequate information.

The Internet Watch Foundation logged 3,443 AI-generated abuse videos in 2025, against 13 the year before, a count drawn from material its analysts assessed directly rather than from industry self-reporting.

The law elsewhere

State legislatures moved faster than the federal courts. As of August 2026, 46 US states criminalise AI-generated or computer-edited abuse material, more than half of those laws passed in 2024 and 2025, according to the tracker maintained by Enough Abuse. Colorado, Massachusetts, Ohio, Vermont and the District of Columbia have no such statute. Oklahoma, Utah, Arkansas and Florida reach wholly fictional depictions; most of the rest require an identifiable minor or the manipulation of a real image, which leaves them closer to the morphed-image question the Seventh Circuit declined to answer.

Section 163.1 of Canada's Criminal Code reaches depictions of imaginary children. The Supreme Court of Canada held in R v Sharpe in 2001 that "person" in the definition covers both actual and imaginary human beings, and upheld the possession offence subject to narrow carve-outs for private material a person creates and keeps for himself alone.

England and Wales criminalise pseudo-photographs under the Protection of Children Act 1978 and non-photographic prohibited images under section 62 of the Coroners and Justice Act 2009. The Crime and Policing Act 2026, which received royal assent on 29 April 2026, created an offence of adapting, possessing, supplying or offering to supply a child sexual abuse image generator, carrying five years. The same Act extended section 69 of the Serious Crime Act 2015, on paedophile manuals, to cover pseudo-photographs and prohibited images.

How the ruling was reported

The Federalist ran the headline "Judge: Child Sex Abuse Images Are OK If They're Made With AI" on 28 August, one of several treatments casting the ruling as a judge legalising synthetic child rape media. The decision reaches in-home possession alone, leaves production, distribution and transfer to a minor fully prosecutable and rests on Supreme Court precedent the panel said it had no authority to set aside. "We are not free to redraw them ourselves", Lee wrote of the lines the Supreme Court has drawn, adding that the court leaves to the Supreme Court the prerogative of overruling its own decisions.

Hans von Spakovsky of Advancing American Freedom said that "AI-generated images are now virtually indistinguishable from the real thing and some of these images may be created based on real child sexual abuse material".

The Justice Department can seek rehearing before the full Seventh Circuit or petition the Supreme Court, which under its Rule 13 allows 90 days from the entry of judgment to file for certiorari, putting the date at 23 November 2026 unless a rehearing petition resets the clock. No decision has been announced.

Anderegg still faces trial in the Western District of Wisconsin on the three surviving counts.

Where the ruling can still be challenged

The Office of the Solicitor General decides whether the United States petitions the Supreme Court, and it has announced nothing. Rule 13 allows 90 days from the entry of judgment, putting the date at 23 November 2026, and a timely petition for rehearing before the full Seventh Circuit would restart the clock from its denial. Both moves are discretionary calls inside the Justice Department that lost the appeal, and neither carries a public comment process.

S.3021, the ENFORCE Act, passed the Senate by unanimous consent on 16 December 2025 and was held at the House desk the following day, where it has sat for eight months without a scheduled vote. John Cornyn sponsored it with Richard Blumenthal, John Kennedy and Mike Lee. Constituents pressing for a floor vote have one named target each, their own House member.

The bill strips the statute of limitations from obscene depictions of child sexual abuse, requires sex offender registration, presumes pre-trial detention, mandates supervised release and tightens the knowledge element on interstate transport. None of that restores a possession charge for images depicting no real child. The obstacle in Anderegg is Stanley, not the drafting of § 1466A, and Congress cannot amend a Supreme Court precedent.

Colorado, Massachusetts, Ohio, Vermont and the District of Columbia have no statute reaching AI-generated or computer-edited abuse material, which leaves a live bill to write for residents of those five jurisdictions. Stanley was itself a Georgia prosecution, so a new state ban on possession inside the home meets the same barrier the Seventh Circuit applied. State law reaching production, distribution and possession outside the home is untouched by the ruling.

Nothing in US federal law matches the offence England and Wales created on 29 April 2026 of adapting, possessing, supplying or offering to supply a child sexual abuse image generator, carrying five years. An offence aimed at the tool rather than at the file in a person's house falls outside what Stanley covers, and no bill before Congress creates one.

LAION withdrew its dataset and published a cleaned version after the Stanford finding, on its own initiative. No US statute requires a model developer to audit training data for abuse material or to disclose what an audit found.

NCMEC's reporting form carries a single generative AI checkbox, which produced the 80 per cent miscount Pfefferkorn identified, and she has asked the organisation to replace it with granular options. Congress, the courts and the concurrence in this case all cite figures drawn from that form.

Overturning the holding itself takes the Supreme Court revisiting Free Speech Coalition and narrowing Stanley, which is what Lee and Kolar asked for in writing, or an amendment to the Constitution. No lower court in the Seventh Circuit can reach it