Federal Appeals Court Protects Narrow Category of AI-Generated Abuse Material

Google News carried a stark claim: a federal judge protected AI-generated child sexual abuse material under the First Amendment. The actual decision is narrower. On August 25, 2026, a three-judge federal appeals panel protected only private, in-home possession of obscene virtual imagery that depicts no identifiable real child.

The ruling did not legalize creating, receiving, transporting, or distributing such material. It did not dismiss allegations that defendant Steven Anderegg transferred AI-generated sexual images to a 15-year-old through Instagram. Those production, distribution, and transfer charges remain pending.

That distinction matters because the case is not simply about whether judges approve of synthetic abuse imagery. They clearly do not. It concerns whether lower courts can disregard two binding Supreme Court precedents because generative AI has changed the factual landscape beneath them.

The Seventh Circuit concluded that it could not redraw those constitutional boundaries. At the same time, two judges urged the Supreme Court to revisit them. The result exposes a growing conflict between speech doctrine developed decades ago and image generators capable of producing photorealistic synthetic content.

What the Google News Headline Leaves Out

The Seventh Circuit decided one possession question under a specific set of facts, not the legality of AI-generated abuse material as a whole.

The case is United States v. Steven Anderegg. According to the government, Meta detected suspected abusive material in an Instagram direct message sent to a minor in October 2023. Meta reported it to the National Center for Missing and Exploited Children, commonly known as NCMEC.

Investigators connected the Instagram account to Anderegg, a Wisconsin software engineer. A search of his devices allegedly found Stable Diffusion and hundreds of generated images. The government acknowledged that the charged images did not depict, or link back to, an actual child.

Federal prosecutors charged Anderegg with producing and distributing obscene visual depictions of minors. They also charged him with transferring such material to someone under 16 and possessing it at home. The allegations remain allegations unless proven in court.

In February 2025, Chief U.S. District Judge James Peterson dismissed only the possession count. Peterson held that applying 18 U.S.C. Section 1466A(b)(1) to private possession violated the First Amendment under existing Supreme Court precedent.

The government appealed that dismissal. The Seventh Circuit heard arguments on October 30, 2025, and affirmed the lower court on August 25, 2026.

Judge John Z. Lee wrote the court’s appellate opinion. Judges Michael Kolar and Doris Pryor joined the result. Lee and Kolar also filed a concurrence asking the Supreme Court for new guidance.

The operative phrase is “as applied.” The panel did not erase the federal statute. It held that one provision was unconstitutional when prosecutors used it against possession inside the home under these particular allegations.

The location also matters. The opinion repeatedly separates possession in the home from acquiring, transporting, producing, or sharing material. Constitutional protection for one act does not automatically protect the conduct needed to obtain or create the files.

That boundary explains why the ruling left three counts intact. Someone who makes synthetic sexual imagery, downloads it from another person, sends it online, or carries it outside the home faces different legal questions.

The Justice Department’s original charging announcement said Anderegg allegedly created thousands of images and sent some to a minor. The appeals court did not determine whether those claims were true.

A more precise Google News headline would say that a federal appeals court protected private possession of wholly synthetic, obscene imagery under existing precedent. That wording is less dramatic, but it reflects the holding.

Two Supreme Court Cases Controlled the Result

The panel’s decision joined a 1969 privacy ruling with a 2002 virtual-imagery ruling, then applied both to modern generative AI.

The first precedent was Stanley v. Georgia. Police searching Robert Stanley’s home for evidence of bookmaking found and viewed several reels of film. Georgia prosecuted him for possessing obscene material.

In 1969, the Supreme Court overturned the conviction. Its Stanley ruling established a narrow constitutional protection for possessing obscenity in the privacy of the home.

Stanley did not create a general right to buy, import, transport, or distribute obscene material. Later cases limited its reach. However, its protection for possession inside the home remained important in Anderegg.

Actual child sexual abuse material falls outside that rule. The Supreme Court has allowed governments to prohibit its possession because making it necessarily harms a real child. Continued circulation also records and compounds that child’s abuse.

The second controlling case, Ashcroft v. Free Speech Coalition, addressed virtual depictions. Congress had expanded federal law to cover images that appeared to show minors even when no real child participated in their production.

The Supreme Court struck down parts of that law in 2002. Its virtual-imagery decision distinguished synthetic depictions from records created through the abuse of actual children.

The government offered several reasons to regulate virtual material. It argued that such imagery could facilitate grooming, encourage later crimes, sustain demand, and make authentic abuse material harder to identify.

The Supreme Court found those connections too indirect under the statute before it. It also rejected suppressing protected expression merely because it resembled unlawful material.

Congress responded with the PROTECT Act in 2003. The law prohibited obscene visual depictions of minors and expressly covered computer-generated images, even when the depicted minor did not exist.

That revision addressed one problem with the earlier statute. It connected the prohibition to the constitutional test for obscenity. Yet obscenity’s legal status still changes when possession occurs entirely inside a home.

The government therefore faced a difficult combination in Anderegg. Stanley protected home possession of obscenity. Free Speech Coalition rejected several rationales for treating wholly virtual depictions like material involving actual children.

Prosecutors argued that current AI systems made the old distinctions less workable. The panel acknowledged that concern, but found it legally insufficient. Lower federal courts must follow Supreme Court holdings until the Supreme Court changes them.

This constraint is the ruling’s central tension. The judges did not declare the material harmless. They concluded that technological change alone did not authorize them to revise constitutional doctrine.

The court also stressed what the government conceded. For the possession count, prosecutors relied on material kept at Anderegg’s home. They were not asking the court to decide whether Stanley protected transporting or receiving it.

The government further conceded that the relevant images could not be linked to real children. That concession removed the most established justification for treating the material as categorically unprotected.

The panel did not extend protection to manipulated images of identifiable minors. When a real child’s photograph is altered into a sexual image, that child can suffer direct and continuing harm. Existing precedent treats those cases differently.

Google News summaries can compress this layered analysis into a misleading proposition. The constitutional line does not run between “good” and “bad” images. It turns on real victims, obscenity, location, and the precise conduct being prosecuted.

The Real Conflict Is Old Doctrine Versus New Evidence

Generative AI has weakened the factual assumptions behind the precedent, but the government did not give the Seventh Circuit a new legal bridge.

In 2002, computer-generated imagery was far less accessible and convincing. Today, downloadable image models can run on personal hardware. Users can modify them, add specialized components, and generate large batches without relying on a hosted service.

According to prosecutors, Anderegg used Stable Diffusion locally. They alleged that he installed additional components and refined text prompts to exclude adults from the outputs. The appeals court accepted the indictment’s allegations as true for reviewing the dismissed count.

Local generation changes enforcement conditions. A platform may never see the prompt, output, or model modifications. Cloud moderation systems cannot inspect activity that never reaches a provider’s servers.

It also challenges the premise that wholly synthetic imagery has no connection to harmed children. Modern models learn statistical patterns from large training datasets. Researchers have found suspected abusive material within datasets used by image-generation systems.

That does not establish that every generated output depicts a specific victim. Nor does it prove that a particular output reproduces a particular training image. It does complicate the clean division between virtual content and real-world exploitation.

Judge Lee’s concurrence focused on that gap. He wrote that the Supreme Court would provide useful guidance on the intersection between the First Amendment and virtual abuse material in an appropriate case.

The concurrence also noted that synthetic images can become virtually indistinguishable from authentic evidence. That creates a different harm from the one considered in older cases.

Investigators traditionally use visual details, metadata, known-image hashes, and surrounding communications to identify victims. A hash is a digital fingerprint that helps platforms recognize files already confirmed as abusive.

Fresh AI outputs do not necessarily match known hashes. Their realism can force investigators to spend time determining whether a depicted child exists. Every convincing synthetic file can become another lead requiring review.

NCMEC says the scale has expanded quickly. Its latest AI data states that staff categorized more than 158,000 submitted images and videos as AI-generated between January 2023 and December 2025.

Electronic service providers marked only slightly more than 11,000 files as AI-generated during that period. That difference suggests platforms often fail to identify or properly label suspected synthetic material before submitting reports.

NCMEC also reported more than 182,000 CyberTipline reports during 2025 involving attempts to create such material or possession of AI-generated material. A report is not proof of a crime, but the volume shows the operational burden.

These numbers give the government a stronger factual story than it possessed in 2002. Synthetic content can consume investigative resources, obscure authentic evidence, and complicate victim identification.

However, stronger evidence does not automatically produce a broader criminal prohibition. First Amendment restrictions normally require a close fit between the regulated expression and the demonstrated harm.

The Seventh Circuit said prosecutors largely repeated arguments already rejected in Free Speech Coalition. The government asserted that virtual material could support grooming, normalize abuse, and sustain markets involving real victims.

The panel found the submitted link between private possession and actual abuse inconclusive. It noted that the Supreme Court had left room for a significantly stronger and more direct connection, but concluded that the government had not established one here.

This is where the ruling becomes more consequential than a single criminal appeal. Research, reporting systems, and forensic evidence must now do more than demonstrate widespread concern. They must help courts identify a constitutionally sufficient causal connection.

That creates pressure for prosecutors and child-safety researchers. They need evidence about how synthetic material affects offending behavior, grooming, trafficking markets, and the discovery of actual victims.

It also places pressure on AI developers. Better provenance records, safety filters, dataset audits, and reporting practices can help distinguish synthetic outputs from authentic evidence. Yet locally modified open models can operate outside those controls.

The court’s ruling does not resolve this technical problem. It reveals that the legal system needs better evidence and more precise statutory tools to address it.

Protection for Possession Is Not Permission to Produce

The practical scope of the decision remains narrow because almost every path into possession involves separate conduct that the ruling did not protect.

Riana Pfefferkorn, a policy fellow at Stanford’s Institute for Human-Centered Artificial Intelligence, identified this limit in reporting after the decision. A person must generally produce, receive, or transport material before privately possessing it.

Those surrounding actions can support different charges. Downloading files from a third-party service involves receipt and potentially interstate transmission. Sharing them involves distribution. Generating them can implicate federal or state production statutes.

Taking a device containing obscene material outside the home can also change the analysis. Stanley protects possession within the home, not a general right to move obscenity through public or commercial channels.

The distinction already affected another federal prosecution. An Ohio court rejected a similar constitutional challenge where prosecutors alleged that material had been downloaded and stored on a phone that left the defendant’s home.

Anderegg’s remaining counts illustrate the same boundary. Prosecutors may continue pursuing allegations involving production, distribution, and transfer to a minor. The Seventh Circuit expressed no view on whether the government will prove them.

The decision also excludes synthetic images based on identifiable children. So-called morphed material uses an innocent photograph of a real person and digitally changes it into sexual content.

A real child in that situation experiences reputational, psychological, and continuing distribution harms. Courts have recognized that connection as materially different from an entirely fictional depiction.

The ruling does not affect prosecutions involving authentic abuse material either. When an image documents a real child’s exploitation, its possession and circulation remain criminal under established federal law.

Nor does the decision create immunity for grooming. Communications intended to entice, solicit, or exploit a minor remain subject to separate laws. Sending sexual material directly to a child is not equivalent to keeping a file inside a private residence.

Earlier enforcement reporting documented how prosecutors were already separating wholly synthetic content from altered images of identifiable children. The legal tools differ because the evidence of harm differs.

State laws add another layer. Legislatures have enacted measures covering AI-generated or digitally altered sexual images of minors. Their language, required proof, and constitutional exposure vary.

A state cannot avoid the First Amendment merely by labeling content harmful. Broad laws covering fictional depictions can still face challenges under Free Speech Coalition.

More targeted statutes have stronger foundations. They can focus on identifiable victims, nonconsensual alteration, distribution, solicitation, deceptive presentation, or conduct directed at minors.

Congress can also revise federal law, but it faces the same constitutional boundaries. A new statute needs evidence and narrow drafting. Repeating a prohibition already rejected by the Supreme Court would invite another challenge.

This is why the headline “AI-generated CSAM is protected” overstates the result. The panel protected one form of possession under one combination of precedents and conceded facts.

The decision offers no safe harbor for online platforms or model developers. Companies still face reporting duties, contractual obligations, app-store policies, and potential exposure under other laws.

It also provides no practical assurance to users. A person cannot reliably know whether an image is wholly synthetic, derived from a real child, transported across legal boundaries, or retained outside the home.

The legal category depends on evidence that investigators and defendants may dispute. Photorealism makes that classification harder, not safer.

AI Platforms Face a Detection and Provenance Test

The ruling increases the value of evidence showing where an image came from, how it was created, and whether a real person contributed to it.

Meta’s report began the investigation in Anderegg. The platform allegedly detected material sent through Instagram and supplied information to NCMEC. That process demonstrates what centralized services can do when activity crosses their systems.

Hosted AI services can inspect prompts, block prohibited requests, monitor repeated abuse patterns, and retain relevant records. They can also report suspected exploitation when federal law requires it.

Local models present a different challenge. Once model weights run on a personal computer, the original developer may have no visibility into later modifications or outputs. A user can add components that remove safety restrictions.

The court’s account says investigators examined prompts, installed software, generated files, messages, and devices. That surrounding evidence allowed prosecutors to distinguish production from possession.

Future cases will depend increasingly on that forensic trail. A generated image alone may not reveal whether it is wholly synthetic, morphed from a real child, or produced from abusive reference material.

Provenance systems can attach information about an image’s origin and editing history. Watermarks, cryptographic credentials, and signed metadata can help, although none offers a complete solution.

Metadata can be stripped. Screenshots can break provenance chains. Open models can omit signing tools. Determined offenders can also use systems that never implemented these standards.

AI detection models introduce another uncertainty. They can estimate whether an image was generated, but false positives and false negatives limit their use as conclusive evidence.

A false synthetic classification could cause investigators to overlook a real victim. A false authentic classification could divert resources toward identifying a person who never existed.

The most useful approach combines signals. Investigators can examine model artifacts, device records, prompt histories, communications, known-image matches, and evidence identifying a real person.

Platforms need clearer reporting fields as well. NCMEC’s figures show a large gap between files its staff classified as AI-generated and those labeled that way by submitting companies.

Better labels can help triage reports. They should indicate whether a file appears wholly synthetic, depicts an identifiable child, alters an existing photograph, or has an uncertain origin.

Uncertainty cannot become an excuse to suppress a report. A real child’s safety must take priority when evidence remains ambiguous. However, structured reporting can direct scarce investigative attention more effectively.

Training data creates an additional provenance problem. If a model learned from unlawful material, synthetic outputs may connect indirectly to real exploitation even when no output reproduces one victim.

The Seventh Circuit did not decide that issue. The concurrence mentioned research about suspected abusive images in a major training dataset, but the government did not build its appeal around a proven training-data connection.

Future litigation can test whether that connection changes the constitutional analysis. It would require specific evidence about the model, dataset, output, and harm involved.

Developers therefore face pressure from both safety policy and litigation. Dataset auditing, documented removal procedures, access controls, and abuse-resistant model design can affect the evidence available to courts.

The ruling also challenges broad claims that filters alone solve the problem. Restrictions in a current hosted product do little to govern older weights already circulating on personal computers.

Stability AI told reporters it opposes misuse and has invested in safeguards. Those company statements do not independently establish how effective the measures are across model versions and local modifications.

The harder policy question concerns distribution architecture. Openness supports research, customization, and independent development. It can also reduce a developer’s control after release.

That tradeoff does not make open models uniquely responsible for abuse. Hosted systems face misuse as well. It does mean that safety commitments require different verification methods across deployment models.

Google News readers should see the decision as a provenance case as much as a speech case. The ability to connect an image to a real child, training source, transmission, or production act determines which legal rules apply.

What Happens After the First Amendment Ruling

Three developments will determine whether Anderegg remains a narrow exception or becomes the case that forces new national rules.

The first signal is a request for Supreme Court review. The Justice Department can ask the justices to take the case after losing its appeal.

Review is not automatic. The Supreme Court accepts a small share of petitions and often waits for disagreement among federal appeals courts. However, the concurrence openly invited guidance about modern AI capabilities.

If the Court accepts the case, the central question will be whether technological change has strengthened the connection between virtual imagery and harm to actual children. Acceptance would elevate Anderegg from a regional appellate decision to a national constitutional test.

A denial would not endorse every part of the Seventh Circuit’s reasoning. It would leave the ruling in place within the circuit and preserve uncertainty elsewhere.

The second signal is the government’s handling of the remaining charges. Production, distribution, and transfer allegations continue in the district court.

A conviction on those counts would reinforce the ruling’s narrow practical effect. It would show that prosecutors can target harmful conduct without relying on possession alone.

An acquittal, dismissal, or major evidentiary problem would expose a larger enforcement gap. That outcome would increase pressure on Congress to revise the statute and on investigators to improve provenance evidence.

The third signal is whether lawmakers develop targeted rules supported by stronger research. Effective legislation must distinguish wholly synthetic files from images tied to identifiable children.

It must also separate possession from creation, receipt, distribution, grooming, and direct communication with minors. Each act presents different harms and constitutional considerations.

Evidence will matter as much as wording. Legislators need reliable findings about investigative burden, offender behavior, victim impact, training data, and market connections.

The NCMEC reporting gap offers one measurable benchmark. Improved platform labeling should produce more actionable referrals and reduce time spent classifying uncertain files.

Researchers should also examine whether synthetic material displaces authentic material, increases demand, supports grooming, or escalates offending. Courts will scrutinize causal claims rather than accept intuition alone.

The issue deserves precision because sensational summaries distort both the danger and the ruling. Saying that judges legalized AI-generated abuse material obscures the surviving criminal charges and the decision’s strict location requirement.

Saying that the case presents no enforcement problem is equally misleading. AI has made synthetic imagery more realistic, easier to generate privately, and harder to distinguish from evidence involving real victims.

The judges recognized that conflict. Their conclusion was institutional: Supreme Court precedent bound them, and only the Supreme Court could change the controlling constitutional rule.

For developers, safety teams, and online platforms, waiting for that change is not a strategy. They can improve dataset controls, retain appropriate forensic records, strengthen reporting quality, and design protections around both hosted and downloadable systems.

For readers following the story through Google News, the best next step is to watch the case docket rather than the headline. Look for a Supreme Court petition, the outcome of the remaining charges, and specific legislative proposals. Those events will show whether the law is adapting to generative AI or merely documenting how far behind it has fallen.