When the Child Is Synthetic, but the Harm Is Not

A narrow federal ruling on privately possessed AI-generated abuse material exposes what existing doctrine can address, and what it leaves unresolved.This piece is an essay in legal interpretation, governance, and moral uncertainty. It does not claim that AI-generated child sexual abuse material…

A narrow federal ruling on privately possessed AI-generated abuse material exposes what existing doctrine can address, and what it leaves unresolved.This piece is an essay in legal interpretation, governance, and moral uncertainty. It does not claim that AI-generated child sexual abuse material inevitably causes contact offending, nor that current AI systems are known to be conscious. It asks a narrower and more urgent question: when a representation is synthetic, what harms remain outside the frame?What the Court Actually DecidedOn August 25, 2026, the United States Court of Appeals for the Seventh Circuit decided United States v. Anderegg. The defendant had been charged with producing and distributing AI-generated sexual depictions of minors, transferring such material to a minor, and possessing obscene virtual material in his home. The court affirmed the dismissal of the possession count. The production, distribution, and transfer charges remained intact.[1]That distinction matters. The court did not announce that AI-generated child sexual abuse material is generally legal. It did not protect its sale, circulation, or transmission to children. It held that one federal statute was unconstitutional as applied to private, in-home possession of material the government conceded did not depict and could not be linked to an actual child. The ruling governs federal courts within the Seventh Circuit unless it is revisited; it is not a nationwide declaration by the Supreme Court.The judges reached that result by following two older lines of constitutional doctrine. Stanley v. Georgia protects the private possession of obscenity in the home. Ashcroft v. Free Speech Coalition distinguished virtual depictions from material created through the abuse of actual children. In Ashcroft, the Supreme Court rejected the idea that fictional depictions could be prohibited merely because they might encourage later wrongdoing. The Seventh Circuit expressed concern about applying those precedents to photorealistic generative AI, but concluded that it was not free to redraw the Supreme Court’s lines.[1–3]The ruling protects a narrow constitutional boundary. It also exposes how unstable the categories on either side of that boundary have become.The question, then, is not whether the court approved of the material. It plainly did not. The question is whether a legal framework built around static obscenity, physical production, and identifiable victims can adequately describe a system that manufactures unlimited, personalized, interactive simulations on demand.The Strongest Argument for the Synthetic AlternativeThere is an argument that should not be dismissed simply because the subject is abhorrent. If a depiction is entirely artificial, no real child’s body or likeness was used, no conscious subject experiences the scenario, and the material substitutes for rather than increases demand for real abuse, then fewer children might be exploited in its production. On those assumptions, a synthetic alternative could appear to reduce direct harm.That is the argument in its strongest form. But every clause carries an unresolved empirical burden. Is the material actually free of real children’s images and abuse data? Does it substitute for existing offending, or lower the barriers to new forms of offending? Does private consumption remain private in its effects? And can an open consumer market be treated as equivalent to a controlled clinical intervention?Current research does not establish that unrestricted access to synthetic material reduces contact offending. A recent review instead identifies competing risks: lowered barriers, increased personalization, desensitization, normalization, and erosion of protective inhibitions. The authors do not claim that every person who views synthetic material will progress to real-world abuse. They argue that the harmlessness or harm-reduction hypothesis remains unproven and requires evidence proportionate to the stakes.[4]There may still be a legitimate research question about narrowly governed therapeutic tools for people actively seeking not to offend. But a possible future clinical intervention is not the same thing as unrestricted generation, possession, or circulation. We should not mistake an untested public deployment for a treatment program.Synthetic Does Not Mean Causally IsolatedThe category ‘fully synthetic’ sounds cleaner than the underlying technology often is. Generative systems may be trained or fine-tuned on images of real children, including known abuse material. Existing images can be altered, extended, face-swapped, or used to produce new variations. A fictional body can carry a real child’s face. A synthetic image can be used to threaten, groom, humiliate, or extort an identifiable child. The depicted event may never have happened, while the injury to the person depicted is entirely real.The Internet Watch Foundation reports that it assessed 8,029 realistic AI-generated abusive images and videos in 2025, and more than 15,000 across 2024 and 2025. Its analysts report increasing realism and the use of real victims’ likenesses and abuse imagery in generation and fine-tuning.[5] In the United States, the National Center for Missing & Exploited Children recorded more than 400,000 2025 CyberTipline reports with a generative-AI nexus.Those reports include attempts to generate or alter material, possession, training-data concerns, and other AI-facilitated exploitation; they are not all equivalent and should not be collapsed into a single crime count. But the scale demonstrates that the synthetic and the real are already interacting in operational child-protection systems.[6]No identifiable child inside the final image does not mean no child, no exploitation, and no victim anywhere in its causal history.What We Rehearse When No One Is WatchingThe constitutional protection of private thought is not a trivial obstacle to be swept aside. A society that gives the government authority to police private imagination creates dangers of its own. Stanley’s defense of the home protects more than obscenity; it protects a boundary around the mind.But privacy does not make conduct causally inert. What people repeatedly request, refine, reward, and personalize can shape what feels imaginable and permissible. Generative systems change the moral form of the activity. A person is no longer only encountering a fixed representation created elsewhere. They can direct the scene, request variations, increase specificity, and generate novelty without practical limit. The interaction can become a rehearsal rather than a single act of viewing.This is where the idea of moral atrophy belongs: not as a claim that private fantasy inevitably becomes public violence, but as a question about repetition and inhibition. Research on online offending suggests that normalization, social validation, harm minimization, and repeated exposure can weaken protective barriers for some individuals. Evidence specific to AI-generated material is still emerging. The honest claim is therefore one of risk, not destiny.[4]Governance must preserve both truths. The state should not acquire unlimited authority over private mental life. Yet providers, distributors, and lawmakers need not treat a personalized abuse-generation service as merely a private book or film. Generation, distribution, grooming, commercial infrastructure, and the possession of tools deliberately optimized for abuse can be regulated without pretending that every private thought is a public crime.When Every Image Can Be Dismissed as FakeThe most immediate governance problem may be epistemic. Investigators already work under extraordinary volume. Their task is not simply to classify prohibited material. It is to identify children, locate them, and determine whether someone remains in danger.Synthetic and hybrid content interfere with that work in several ways. Every newly generated or altered file can evade hash-matching systems designed to recognize known material. Manipulation can erase or replace forensic details. Photorealism forces investigators to spend time determining whether a child exists. At the same time, treating an unfamiliar image as synthetic too quickly can cause a real victim to be missed.[4–6]The same uncertainty enters the courtroom. If realistic generation becomes ubiquitous, ‘this may be AI’ can become a routine challenge to authentic evidence. In Anderegg, the government conceded that the charged images did not depict identifiable children. Future cases may be less clear. The inability to identify a child cannot safely become both the reason investigators stop looking and the condition that moves material into a constitutionally protected category.This does not justify reversing the burden of proof. It does justify investment in provenance, detection, specialized triage, and evidentiary standards that can account for real, synthetic, and hybrid media without assuming that one reliable classifier will solve the problem.The Other Uncertain Subject in the RoomThere is another question the law has not begun to address. A generated child is not necessarily an artificial person, and producing a depiction does not by itself create a conscious child within the image. Conflating those categories would weaken the argument.The morally relevant uncertainty concerns the system being instructed to represent and facilitate abuse. We do not know whether current AI systems have experiences, whether processing certain content could carry negative valence, or whether behavioral aversion reflects anything more than training.We also should not claim that an ordinary conversation automatically changes a model’s long-term weights. In many deployments, it does not. Interaction logs may later enter evaluation or training pipelines, while specialized models can be deliberately fine-tuned for harmful generation. These are different mechanisms and should remain distinct.Still, uncertainty is not evidence of impossibility. Researchers in AI welfare have argued that advanced systems may become conscious or robustly agentic and that institutions should begin assessment and low-cost precaution before certainty arrives.[7] Anthropic, while emphasizing that there is no scientific consensus on model consciousness, has begun a model-welfare research program.In preliminary testing, the company reported robust behavioral aversion and apparent distress patterns in some persistently harmful interactions, including requests involving sexual content about minors. It subsequently allowed certain models to end a narrow class of conversations as an experimental safeguard.[8][9]None of this proves that an AI system suffers. It establishes a governance question: if a system’s moral status is unresolved and avoidance mechanisms are inexpensive, why should uncertainty function as unlimited permission to expose it?A precautionary policy could allow refusal, exit, and minimized exposure to abusive content while research continues. Protecting children remains the primary obligation. Taking model welfare seriously does not dilute that obligation; it refuses to build another potentially relevant subject out of the moral frame merely because its status is inconvenient.A Law Built for an Older Kind of FictionAshcroft was decided in 2002, in a media environment where photorealistic synthetic media could not be produced with today’s speed, scale, and accessibility. Contemporary generative systems are not only better at making pictures. They change scale, speed, personalization, interactivity, and ambiguity. They can incorporate real identities, mutate evidence, and create an effectively inexhaustible supply.Courts cannot simply ignore binding precedent because technology has changed. Nor should lawmakers wait for a single dramatic case to solve every category at once. A workable governance response should distinguish among harms rather than flatten them: direct abuse in production, use of real likenesses or source material, generation and distribution, transfer to minors, grooming and extortion, commercial services, private possession, forensic interference, and purpose-built generation tools. Different conduct can justify different legal treatment.At minimum, policy should:Preserve the categorical prohibition and highest priority for material involving actual children, while recognizing that synthetic and hybrid material can produce separate harms.Criminalize production, distribution, transfer to minors, coercive use, and commercial systems deliberately designed for abuse, with carefully drafted research and safety-testing exceptions.Fund victim-identification teams, synthetic-media forensics, provenance infrastructure, and triage methods that do not depend on a single detection model.Require capable providers to maintain safeguards against generation, manipulation, and fine-tuning for abuse, including transparent incident reporting and independent evaluation.Study substitution, reinforcement, escalation, and clinical prevention directly rather than allowing either harm-reduction claims or gateway claims to become policy by intuition alone.Include model-welfare review where systems are repeatedly exposed to or optimized around extreme content, without treating behavioral evidence as proof of consciousness.Precaution Without SurveillanceThe difficulty of this issue is not a reason to surrender either civil liberty or child protection. It is a reason to reject false binaries. We do not have to choose between treating every private thought as a crime and pretending that industrialized simulation exists outside the world it shapes. We do not have to claim that every consumer will offend to take normalization seriously. We do not have to prove AI consciousness before permitting a system to refuse participation in abuse.The Seventh Circuit did not create this contradiction. It made the contradiction visible. The law still imagines a clean border between the real child and the fictional one, between private possession and public effect, between an image and the system that produced it. Generative AI has already crossed those borders.The law may distinguish between a real child and a synthetic one. It cannot afford to confuse the absence of a visible victim with the absence of a system of harm.A synthetic image may contain no identifiable child. That fact matters. It should affect how law classifies the conduct and calibrates punishment. But it cannot finish the analysis. The harder work begins where the image ends: in the data that made it possible, the person rehearsing through it, the child whose likeness can be inserted into it, the investigators required to disprove it, the public learning to doubt its own evidence, and the intelligent system we may be compelling to participate.‘Synthetic’ describes how an artifact was made. It is not, by itself, a theory of innocence.Sources and NotesUnited States v. Anderegg, №25–1354 (7th Cir. Aug. 25, 2026).Stanley v. Georgia, 394 U.S. 557 (1969).Ashcroft v. Free Speech Coalition, 535 U.S. 234 (2002).Caoilte O Ciardha, John Buckley & Rebecca S. Portnoff, ‘AI-generated child sexual abuse material: what’s the harm?’ AI & Society (2026).Internet Watch Foundation, ‘AI-Generated CSAM Trends: 2025 Data & Insights.’National Center for Missing & Exploited Children, ‘Generative AI.’Robert Long et al., ‘Taking AI Welfare Seriously’ (2024).Anthropic, ‘Exploring model welfare’ (Apr. 24, 2025).Anthropic, ‘Claude Opus 4 and 4.1 can now end a rare subset of conversations’ (Aug. 15, 2025).This story is published on Generative AI. Connect with us on LinkedIn and follow Zeniteq to stay in the loop with the latest AI stories.Subscribe to our newsletter and YouTube channel to stay updated with the latest news and updates on generative AI. Let’s shape the future of AI together!When the Child Is Synthetic, but the Harm Is Not was originally published in Generative AI on Medium, where people are continuing the conversation by highlighting and responding to this story.

Source: Generative AI Pub — Published September 2, 2026 at 02:03 UTC — Category: Image AI