Seventh Circuit Affirms Dismissal of One Count Against Wisconsin Man Over AI-Generated Child Sex Abuse Images
A three-judge panel ruled on August 25, 2026 that Supreme Court precedent bars prosecuting Steven Anderegg for possessing obscene AI-made images in his home; three other charges, including production and sending an image to a minor, still stand.
One Charge, Three Left Standing
Steven Anderegg has not been cleared of anything. On August 25, 2026, a federal appeals court threw out one criminal count against him — but three others, including making and sending the images in the first place, are still headed to trial[1][2].
The dismissed charge was for simply having AI-generated child sexual abuse images inside his own home. The Seventh Circuit Court of Appeals said it had no choice: two old Supreme Court rulings tied its hands[1][4]. What's left is a fight over whether those old rulings still make sense now that AI can produce images this realistic — and neither side disputes that the images themselves are repugnant[3][4].
The Two Rulings Doing All the Work
The case turns on two Supreme Court decisions decades old. Stanley v. Georgia, from 1969, says the government generally can't punish someone just for keeping obscene material private in their own home[3][4]. Ashcroft v. Free Speech Coalition, from 2002, says images that only look like child sexual abuse — but don't show a real child — count as protected speech[3][4].
Put together, those two rulings shield one narrow thing: having fake, AI-made abuse images at home, if no real child is shown. A three-judge panel — John Z. Lee, Doris Pryor and Joshua Kolar — said that shield applies to Anderegg's possession charge, so they upheld a lower court's decision to dismiss it[1][5].
That's it. The ruling doesn't touch the other three charges against him: producing the images, distributing them, and sending one to a boy under 16[1][5]. Prosecutors say Anderegg used Stable Diffusion, an AI image generator, to make hundreds of photorealistic images of prepubescent children. Some earlier charging documents put the number in the thousands[1][3][5]. The investigation started after Meta flagged a message he sent a minor on Instagram to the National Center for Missing and Exploited Children[3][7].
Why the Court Says Its Hands Were Tied
Judge Lee, who wrote the panel's opinion, didn't leave much room for reading between the lines. He said the government's best arguments — that fake images can groom kids, normalize abuse, or replace and expand demand for real abuse material — were already raised and rejected by the Supreme Court back in 2002[4][6]. A circuit court doesn't get to overrule the Supreme Court just because it thinks the facts have changed, he wrote, or "anarchy" would "prevail within the federal judicial system"[4][6].
That's the core tension in the whole case, and it's not really about AI at all. It's about who gets to update a legal rule once the technology underneath it stops matching the world the rule was written for. The panel's answer: not us.
But then something unusual happened. Lee and Judge Kolar wrote a separate opinion — agreeing with their own ruling, but arguing it might be wrong. Today's AI images, they wrote, are "virtually indistinguishable from CSAM created through the physical abuse of a child"[3][4]. They pointed out that two Supreme Court justices back in 2002, Sandra Day O'Connor and Clarence Thomas, had already flagged this exact possibility. They also raised something the 2002 Court never had to consider: some AI image generators may have been trained on real abuse material, which would undercut the entire "no real victim" premise the rule depends on[3][4].
A Rule Prosecutors Can't Prove, Against a Ban With No Edges
Here's the practical problem driving the Justice Department's argument. If guilt depends on proving a real child exists in the image, then any defendant caught with genuine abuse material can just claim a machine made it — and force prosecutors to prove otherwise, image by image[4][6]. For an agency trying to prosecute at scale, that's not a fringe worry. It's a hole in the entire system.
Free-speech defenders answer with a different worry: a rule with no edges. If the government can ban images that hurt no real, identifiable person just because they might encourage someone to commit a crime later, that same logic could reach violent novels or films. The Supreme Court rejected exactly that reasoning in 2002, and the Seventh Circuit says it's stuck applying that rejection[4][6].
Sitting underneath both arguments is a technical fact that complicates the "no real victim" idea for everyone. AI image generators learn from scraped data, and researchers have found real abuse material inside some training sets. NCMEC, the nonprofit that runs the national tipline for reporting this kind of material, logged more than 12,000 reports in 2025 of real abuse content turning up inside AI training data[8]. If a "synthetic" image was partly built from a real child's abuse, the clean line the ruling relies on starts to blur.
And the numbers behind this fight are not small. NCMEC's tipline took in 21.3 million reports in 2025 overall, more than 7,000 of them specifically about users generating or possessing AI-made abuse material[8]. The UK's Internet Watch Foundation, which tracks this kind of content globally, counted 3,440 AI-generated abuse videos in 2025 — up from just 13 the year before[9]. That's the surge child-safety groups point to when they say a rule built on proving a real victim can't keep pace.
What the Headlines Did With It
Coverage of the ruling split hard along predictable lines, and the gap between headlines was bigger than the gap in the reporting underneath them. Fox News called it an "unsettling" ruling but kept the court's own explanation front and center — that the Supreme Court "left it no choice"[1]. The Federalist went much further, running the headline "Judge: Child Sex Abuse Images Are OK If They're Made With AI," and Townhall called it "the most disgusting federal court ruling of the year." Neither mentioned that the ruling covers only possession, only in the home, and only one of four charges — or that the same judges who wrote it also asked the Supreme Court to overturn the rule they were bound by[1].
The Washington Post took the opposite tack, leading with the judges' warning that "law is being left behind by AI child sex abuse images" rather than with the outcome itself. That framing is accurate, but it also softens what actually happened: a defendant won a real ruling. Reason's Volokh Conspiracy blog produced what reads as the most precise headline of the bunch, describing it plainly as home possession of AI images "depicting fake children" remaining protected — though even that choice of words, emphasizing "fake," leans toward the framing favorable to the ruling. Legal trade outlets like Minnesota Lawyer covered it in flatter, more procedural terms aimed at readers who already understand the legal posture.
Notably missing from the coverage: any real reporting from outside the United States. Despite the UK-based Internet Watch Foundation supplying much of the global data everyone cites, this has so far been treated entirely as an American constitutional story.
What Happens Next
Anderegg still faces trial on the three surviving charges, each carrying real prison time[1][5]. The ruling only removes one narrow tool for prosecutors — and only within Illinois, Indiana and Wisconsin, the three states the Seventh Circuit covers.
What happens beyond that is now genuinely uncertain. Two sitting appellate judges have put it on the record that the 2002 precedent controlling this case may not fit a world with generative AI, and they've openly invited the Supreme Court to take the question back up[3][4]. Whether the Court does, and whether it moves before the surge in AI-generated material outpaces the law further, is the part nobody involved in this case gets to decide.
Summary
On August 25, 2026, a three-judge panel of the U.S. Court of Appeals for the Seventh Circuit ruled against the Justice Department in the case of Steven Anderegg, a Wisconsin man. The panel upheld a trial judge's decision to throw out one charge: knowingly possessing obscene, AI-generated images of child sexual abuse inside his own home[1][2]. The court said two Supreme Court decisions tied its hands. Stanley v. Georgia (1969) says the government generally cannot punish someone for keeping obscene material in the privacy of the home. Ashcroft v. Free Speech Coalition (2002) says images that only look like child sexual abuse, but involve no real child, are protected speech[3][4]. Three other charges against Anderegg — producing the images, distributing them, and sending one to a boy under 16 — were not part of this appeal and remain alive[1][5].
The ruling is narrower than many headlines suggest. It applies to one count, under one subsection of one statute, and only as applied to these facts: images the government does not claim depict a real, identifiable child, kept privately. The panel did not strike down the law[1][6]. Judge John Z. Lee, who wrote the opinion, said the government's arguments for banning possession — that such images groom children, normalize abuse, substitute for real abuse material, and make real cases harder to prosecute — were "expressly discussed and rejected" by the Supreme Court in 2002, and that a lower court ignoring binding precedent would invite "anarchy...to prevail within the federal judicial system"[4][6].
Then Lee and Judge Joshua Kolar did something unusual. They wrote a separate concurrence — an opinion agreeing with the result but adding their own view — urging the Supreme Court to take the issue back up. AI images today, they wrote, are "virtually indistinguishable from CSAM created through the physical abuse of a child." They noted that Justices Sandra Day O'Connor and Clarence Thomas had flagged exactly this possibility in 2002. They also raised a point the 2002 Court never had to weigh: some AI image generators may have been trained on real abuse material, which would weaken the "no real victim" premise the whole rule rests on[3][4].
The genuine dispute is not really about whether these images are repugnant — no side argues otherwise. It is about whether harm that is indirect and statistical is enough to override a First Amendment rule. The Justice Department says realistic synthetic images cause real harm downstream. Free-speech defenders say the government must show a real child was hurt, or the rule has no stopping point. Both sides now largely agree the Supreme Court, not a circuit court, has to answer it[3][4].
The Event
On August 25, 2026, the U.S. Court of Appeals for the Seventh Circuit, in United States v. Anderegg, No. 25-1354, affirmed the dismissal of one count against Steven Anderegg of Wisconsin[1][2]. That count charged knowing possession of obscene visual representations of the sexual abuse of children under 18 U.S.C. § 1466A(b)(1); Chief Judge James D. Peterson of the Western District of Wisconsin had dismissed it on First Amendment grounds, and the appeals panel reviewed that ruling de novo[1][5]. The panel — Judges John Z. Lee, Doris Pryor and Joshua Kolar — held it was bound by Stanley v. Georgia and Ashcroft v. Free Speech Coalition[1][4]. Lee and Kolar filed a separate concurrence urging the Supreme Court to revisit Free Speech Coalition in light of generative AI[3][4].
Undisputed Facts
- The Seventh Circuit issued its decision in United States v. Anderegg, No. 25-1354, on August 25, 2026[1][2].
- The appeal concerned only one count: knowing possession of obscene virtual child sexual abuse material under 18 U.S.C. § 1466A(b)(1), and the panel affirmed the district court's dismissal of that count[1][5].
- Anderegg still faces the other charges in the indictment — producing the images, distributing them, and transferring obscene material to a minor under 16[1][5].
- Prosecutors allege Anderegg used Stable Diffusion, an open-source AI image generator, to create photorealistic sexual images of prepubescent children; the opinion describes hundreds of such images, while some news accounts and the original 2024 charging documents describe thousands[1][3][5].
- The investigation began after Meta flagged an Instagram message Anderegg sent to a minor to the National Center for Missing and Exploited Children, which led to a search warrant[3][7].
- In Stanley v. Georgia (1969) the Supreme Court held the government generally cannot criminalize private possession of obscene material in the home; in Ashcroft v. Free Speech Coalition (2002) it held that sexual images not depicting real children are protected speech[3][4].
- Judge Lee's opinion stated that the government's grooming, normalization, substitution and prosecution-difficulty rationales were "expressly discussed and rejected" by the Supreme Court in Free Speech Coalition[4][6].
- Judges Lee and Kolar wrote separately that today's AI images are "virtually indistinguishable from CSAM created through the physical abuse of a child" and called on the Supreme Court to revisit the issue[3][4].
- The National Center for Missing and Exploited Children reported that its CyberTipline received 21.3 million reports in 2025, including more than 7,000 reports of users generating or possessing AI-generated child sexual abuse material[8].
- The UK-based Internet Watch Foundation reported assessing 8,029 AI-generated images and videos in 2025 as showing realistic child sexual abuse, including 3,440 AI videos, up from 13 the year before[9].
The Pressure
Strip away the moralizing and blame. What structural realities persist regardless of which narrative wins?
- Lower courts cannot update precedent
- The single biggest driver of this outcome is not anyone's view of AI. It is that a federal appeals court must apply Supreme Court rulings even when it believes the facts underneath them have changed. Judge Lee said so directly, warning that the alternative invites "anarchy...to prevail within the federal judicial system"[4][6]. The panel's real message was addressed upward, not outward.
- Prosecutors need a rule they can prove
- The government's deepest interest is evidentiary, not moral. If guilt turns on whether a real child exists, then every realistic AI image becomes a reasonable-doubt argument, and every prosecution of real material invites the same defense. A categorical ban is enforceable at scale; a case-by-case victim-identification standard is not[4][8].
- A rule with no stopping point
- The free-speech side's structural fear is that harm-by-influence has no natural limit. If images that hurt no identifiable person can be banned because they might encourage crime, the same reasoning applies to any disturbing fiction. The 2002 Court rejected that logic, and the Seventh Circuit applied that rejection[4][6].
- Training data undercuts the clean line
- The premise doing all the work — "no real child" — is factually contested at the technical level. Models are trained on scraped images, and real abuse material has been found in training sets; NCMEC logged more than 12,000 reports in 2025 of such material identified inside AI training data[8]. Two of the three judges raised this themselves[3][4].
Material realitySteven Anderegg has not been cleared. Three charges — production, distribution, and sending obscene material to a boy under 16 — survive and can still send him to prison[1][5]. What changed is narrower: in Illinois, Indiana and Wisconsin, federal prosecutors currently cannot charge private, at-home possession of obscene virtual images under 18 U.S.C. § 1466A(b)(1) when the government does not claim a real child is depicted[1][3]. Everything else about the material remains illegal: making it, sharing it, selling it, sending it to a child. Separately, the volume is not theoretical. NCMEC's CyberTipline took 21.3 million reports in 2025, of which more than 7,000 specifically involved users generating or possessing AI-made abuse material[8]. The IWF counted 3,440 AI-generated abuse videos in 2025 against 13 in 2024 — a change from a rounding error to a category[9]. Those numbers are what will drive legislatures and the Supreme Court, regardless of how this one count is described.
Narrative as a weaponThree groups are shaping how this is read. The panel itself is the most active: by pairing a defendant-friendly holding with a concurrence calling the underlying precedent obsolete, Lee and Kolar want you to believe the court's hands were tied and the Supreme Court should untie them. Right-leaning commentary — sharply at The Federalist and Townhall, more carefully at Fox News — wants you to believe a judge blessed AI-made child abuse imagery, which requires dropping the words "possession," "in the home," and "one of four counts." Child-safety organizations, which run the reporting pipelines the rule governs and depend on federal and industry funding, want you to focus on the training-data problem and the surge counts, because those are the facts that make a "no real victim" rule look like a fiction. Free-speech advocates want you to focus on the missing victim, because that is the fact that makes the ban look limitless. Both are arguing about the same images from different ends. Notably absent: any significant non-U.S. coverage, which leaves the framing entirely to American partisan and legal-trade outlets.
How Each Side Sees It
Each major actor’s view — how it frames things, its underlying incentive, and how it’s materially affected. Tap a side to read it.
Frames it asThe government's core argument is that the 2002 precedent was built on a factual assumption that no longer holds. In 2002, computer-generated images were crude and rare, so the Supreme Court could safely say fake images did not feed the market for real abuse. Prosecutors argue that is no longer true: when synthetic images are photorealistic, they can be used to groom children, they normalize abuse, they can substitute for and expand the market for real material, and — the practical point prosecutors press hardest — they make real cases nearly impossible to win, because a defendant holding genuine abuse images can simply claim a machine made them and force the government to prove a specific real child exists[4][6]. On that view, the home-possession shield from Stanley is not absolute; the Supreme Court has already carved out real-child material from it, and the reason for that carve-out is harm, not the location of the hard drive[3][4].
WhyProsecutors want a workable evidentiary rule. If the line is "real child or not," every AI-capable defendant gets a defense that is cheap to assert and expensive to disprove. DOJ also brought this as a test case in 2024, one of its first federal prosecutions over purely AI-generated material, precisely to establish that § 1466A reaches it[3][7].
Impact on themThe three surviving counts mean Anderegg can still be tried and, if convicted, imprisoned — so the department has not lost the case, only one count[1][5]. But the ruling makes the Seventh Circuit — Illinois, Indiana and Wisconsin — a jurisdiction where pure home possession of virtual material cannot be charged, which shapes where and how the department brings future cases[1][3].
Frames it asTheir strongest argument is not that the images are harmless. It is that the government is asking courts to punish thoughts and private tastes when it cannot point to a victim. Child pornography is one of the very few categories of speech the Supreme Court has left entirely unprotected, and the Court's stated reason was specific: making it requires abusing an actual child, so banning it dries up demand for that abuse. Strip out the real child and you strip out the reason[3][4]. If "it might lead someone to commit a crime later" is enough, they argue, the same logic reaches violent novels and films — an argument the Supreme Court rejected in 2002 in exactly those terms. They also stress the structural point Judge Lee made: a circuit court does not get to overrule the Supreme Court because the facts have moved, or the whole system of precedent stops working[4][6].
WhyDefense counsel's immediate goal is narrowing the exposure of one client. The broader free-speech interest — represented in commentary by outlets like Reason's Volokh Conspiracy — is preventing a "harmful ideas" exception that could later extend to other synthetic media[6].
Impact on themAnderegg loses one charge from the indictment but still faces trial on three others carrying substantial prison exposure[1][5]. For free-speech litigators, the win is precedential and jurisdiction-wide, but fragile: two of the three judges publicly invited the Supreme Court to take it away[3][4].
Frames it asTheir case rests on volume and on the training-data problem. They argue the "no real child" premise is an illusion at scale: image generators learn from scraped data, and researchers have found real abuse material inside training sets — so a "synthetic" image can carry a real victim's likeness or be built partly from their abuse[3][8]. Judges Lee and Kolar flagged this same point[3][4]. They also argue the flood is real and measurable, not speculative: NCMEC logged more than 7,000 reports in 2025 of users generating or possessing AI-generated abuse material, plus over 145,000 reports of users altering existing files with AI tools[8]. The Internet Watch Foundation counted 3,440 AI-made abuse videos in 2025, against 13 the year before[9]. Their conclusion: investigators cannot triage 21.3 million CyberTipline reports a year if every file requires proving a specific child is real[8].
WhyThese groups want a bright-line rule they can enforce at scale. A per-image, prove-the-child standard is unworkable for a tipline handling tens of millions of reports; a categorical ban is. NCMEC is a congressionally chartered nonprofit funded substantially by federal grants, and the IWF is a UK industry-funded body — both operate the reporting pipelines that this rule directly governs[8][9].
Impact on themThe ruling does not change their referral work, since the surviving charges show production and distribution remain prosecutable. But it removes one enforcement tool in three states and strengthens the argument they have been making to Congress and state legislatures for statutes that do not depend on identifying a real victim[3][8].
Frames it asThe panel's position is deliberately split in two, and reading only one half distorts it. As a court, its argument is about hierarchy: lower courts follow Supreme Court precedent even when they think it has gone stale, or, in Lee's phrase, "anarchy" would "prevail within the federal judicial system"[4][6]. As individual judges, Lee and Kolar wrote that the 2002 rule may no longer fit reality, that AI output is now "virtually indistinguishable" from images of real abuse, and that Justices O'Connor and Thomas anticipated this in 2002 by suggesting the government might have a stronger case once the technology caught up[3][4]. In effect, the panel ruled one way and signaled the opposite result may be warranted at a higher level.
WhyInstitutional. The panel is protecting the rule that lower courts do not update Supreme Court doctrine on their own, while using a concurrence — the one tool available — to flag the case for review.
Impact on themThe opinion functions as an invitation. It gives the Justice Department a clean vehicle for a petition to the Supreme Court, with two sitting appellate judges already on record that the precedent needs a second look[3][4].
Like this article?
The Bias Ledger average rating 5.2
The same story, as framed by outlets across the spectrum, ordered least to most biased. The bias score (1 = straight, 10 = heavily spun) is an AI assessment of that framing — click an outlet to see its track record. The tell is the word choice or omission that reveals the angle.
| Outlet | Vantage | Bias | How they frame it | The tell |
|---|---|---|---|---|
| Minnesota Lawyer | U.S. legal trade press (BridgeTower Media) | 2 | "Judge: Law is being left behind by AI child sex abuse images." | Trade-press framing centers the doctrinal gap and the procedural posture, with almost no adjectives. Reaches a professional audience, so it assumes readers know what an as-applied challenge is and does not spell out what survives for a general reader. |
| Reason (Opinion) | U.S. libertarian; the Volokh Conspiracy is a signed law-professor blog hosted by Reason | 3 | "Home Possession of AI-Generated Porn Depicting Fake Children Remains Constitutionally Protected, Seventh Circuit Rules." | The most legally precise headline in the set — "home possession," "fake children," "remains" all carry accurate load. The word choices "porn" and "fake children" are also the framing most favorable to the free-speech reading, minimizing the realism the concurring judges emphasized. |
| Fox News | U.S. right | 4 | "Appeals court delivers unsettling ruling — and says Supreme Court left it no choice," reporting that AI-generated CSAM possession is protected by the First Amendment. | "Unsettling" is the outlet's own characterization, not a party's, placed in the headline. But the second clause does real reporting work: it moves responsibility to the 2002 precedent rather than the panel, and the body keeps the no-real-person and in-the-home conditions. |
| The Washington Post | U.S. left | 4 | "Federal judge warns law is being left behind by AI child sex abuse images." | The headline leads with the concurrence's warning rather than the holding, which softens the outcome — the ruling went for the defendant, and the headline reads as a judge sounding an alarm. The body is precise, stating both conditions (not a real person, stays in the home) in the first sentence. |
| The Federalist | U.S. right | 9 | "Judge: Child Sex Abuse Images Are OK If They're Made With AI." | The headline puts a flat endorsement in a judge's mouth that appears nowhere in the opinion, and deletes every limit that defines the holding: possession only, in the home only, one of four counts. It also omits that the same judge asked the Supreme Court to overturn the rule he applied. |
| Townhall | U.S. right | 9 | "This Is the Most Disgusting Federal Court Ruling of the Year." | A superlative moral verdict replaces the ruling's content entirely; a reader cannot learn from the headline what count was dismissed, under what statute, or that three charges survive. |
References
- USA v Anderegg, No. 25-1354 (7th Cir. 2026) — Justia · Legal database republishing the primary court opinion; no editorial orientation
- In the United States Court of Appeals For the Seventh Circuit — Opinion, No. 25-1354 — U.S. Court of Appeals for the Seventh Circuit · Primary source; the court's own published opinion
- Judge ruled certain AI-generated child sex abuse material is protected by First Amendment. Here's context — Snopes · U.S. fact-checking site, ad- and subscription-funded; generally center to center-left in topic selection
- Judges rule AI child abuse images are protected in your own home — Techlicious · U.S. consumer-technology site, affiliate- and ad-funded; not politically aligned
- United States v. Steven Anderegg, 3:24-cr-00050 — docket — CourtListener · Free Law Project, a nonprofit court-records archive; publishes primary filings without commentary
- Home Possession of AI-Generated Porn Depicting Fake Children Remains Constitutionally Protected, Seventh Circuit Rules — Reason · U.S. libertarian magazine published by the Reason Foundation; this post is the signed Volokh Conspiracy law blog, not newsroom reporting
- Possession of AI-generated child sexual abuse imagery may be protected by First Amendment in some cases, judge rules — NBC News · U.S. broadcast network news, owned by NBCUniversal/Comcast; generally center-left
- The Work Never Stops: A First Look at NCMEC's 2025 Data — National Center for Missing & Exploited Children · Congressionally chartered U.S. nonprofit funded substantially by federal grants; operates the CyberTipline and advocates for stronger CSAM enforcement
- AI CSAM Report 2026: Harm Without Limits — Internet Watch Foundation · UK charity funded largely by internet and technology companies; advocates for takedown and criminalization of AI-generated CSAM
- Federal judge warns law is being left behind by AI child sex abuse images — The Washington Post · U.S. national daily owned by Jeff Bezos; news pages generally center-left
- Judge: Law is being left behind by AI child sex abuse images — Minnesota Lawyer · U.S. legal trade publication owned by BridgeTower Media; professional audience, minimal political framing