Courts just allowed AI child sex abuse material because it’s ‘virtual.’ Who do you think it’s trained on?
The recent 7th Circuit opinion permitting in-home possession of AI-generated child sexual abuse material raises many questions regarding the relationship between protected speech and generative AI.
While most people would say CSAM of any child — real or virtual — should be illegal, the decision relies on precedent from the 2002 Ashcroft v. Free Speech Coalition decision, holding that computer-generated CSAM, possessed in one’s own home, does not violate the First Amendment. But that decision did not contemplate the present realities and may not be applicable to today’s AI-generated CSAM.
Here’s why: On one hand, the judiciary holds in high regard First Amendment jurisprudence and views cases — including those involving new technology — through the lens of preserving what the courts have historically said is “free speech.” On the other hand, new unregulated technology, including GenAI, is forcing a reckoning on the boundaries of what we agree the First Amendment protects. Amid the increasingly narrow hair-splitting reasoning courts are using, it’s clear that the question courts are addressing today regarding AI-generated CSAM is not the same one they were addressing 25 years ago.
Unrestrained AI technology has brought us to a dark place: the evils of child sexual abuse material and the horrors of unregulated AI have merged. CSAM and generated CSAM are no longer wholly distinct categories, as some AI models have been trained on real sexual abuse images of children. The AI model that Steven Anderegg — the defendant in the 7th Circuit case — used to create sexual images of prepubescent minors was trained on a dataset that contains billions of real images from hundreds of different domains. Virtual CSAM is no longer a cartoon or depiction that no one could reasonably mistake for a child. The 7th Circuit Court got this right, recognizing that, “GenAI models can render images depicting the abuse of virtual children that are virtually indistinguishable from those depicting the abuse of actual children.”
This is happening with other artificial intelligence image generators, too. In a case filed this week, a victim of both CSAM and AI-generated CSAM sued xAI, alleging in part that she believes, “Grok has been trained on both the initial set of images that have haunted her for more than 20 years and the more recent AI-generated ones.”
Where CSAM ends and AI-generated CSAM begins is no longer a bright line, and courts are going to have to wrestle with what this means for the future of free speech. The 7th Circuit Court acknowledges this tension, noting the “relentless advancement in artificial intelligence models” requires lines to be drawn, but simultaneously tries to distance lower courts from the battlefront, saying, “we are not free to redraw them ourselves.”
But courts can contribute to a solution by acknowledging that modern technology changes the very battlefield on which lines are drawn. The existence of new technology means the questions posed in Ashcroft about CSAM possession are not the same questions being posed today.
Because “the Supreme Court left open the door for the government to make a ‘significantly stronger, more direct connection’ between virtual CSAM and child abuse,” as the 7th Circuit notes, there is a path forward to develop the argument that AI-generated CSAM relies on real CSAM and non-abuse images of real children. This argument should yield a different result from the arguments asserted in Ashcroft, where the court found generated CSAM was protected because the “materials at issue did not involve actual children.” This suggests that the analysis changes when actual children are involved. The question now becomes: Does First Amendment protection extend to AI-generated images created by training on real CSAM of real children?
Perhaps the government needs to clearly advance a factual theory that demonstrates how AI-generated CSAM is not wholly distinct from CSAM depicting real children to give the courts a new argument to latch onto. And courts should respond accordingly, consistent with case law that holds in high regard the state’s compelling interest in “protecting the physical and psychological well-being of minors and preventing the sexual exploitation and abuse of children,” while preserving free speech. (U.S. v. Anderegg, at 8.)
The National Center on Sexual Exploitation notes the very real harms to children from AI-generated CSAM, noting that it “can be used as a grooming tool by predators, and NCMEC reports that children and their families have been sextorted through predators using AI-generated CSAM.”
CHILD SAFETY PAYS: META PAID $17 BILLION FOR EXPLOITING CHILDREN — AND MADE $130 BILLION DOING IT
And the horrific reality is that AI-generated CSAM will increase the demand for more of this content, putting all children at risk.
The Supreme Court must weigh in on this issue and find that AI-generated CSAM is not protected by the First Amendment.
Tori Countryman is legal counsel at the National Center on Sexual Exploitation, the leading national nonprofit organization exposing the links between all forms of sexual exploitation, such as child sexual abuse, prostitution, sex trafficking, and the public health harms of pornography. www.EndSexualExploitation.org
There needs to be a Congressional Inquiry into whether AI generators were actually trained on real CSAM. If so, serious charges need to be laid against those involved with real prison time as a consequence.
=====================================================================================
