In recent conversations I've had with ICAC investigators, a troubling theme keeps coming up: they're encountering more and more AI-generated abuse imagery that looks and feels real, but isn't tied to a specific victim. No original photo. No traceable child. Just prompts and pixels used to generate content that’s nearly impossible to distinguish from real abuse.

And yet, the impact is no less devastating.

One investigator I spoke with put it bluntly: “How do we even begin to tackle this?”

That question has kept me up at night. Not just as someone who builds AI tools for child protection, but as a mother who knows that the same technology revolutionizing how we fight abuse is also being weaponized against the children we're trying to protect.

When Red Teams Need Legal Shields

Over the past few years, I've sat in conference rooms full of law enforcement officers discussing the latest surge in AI-generated child sexual abuse material. The statistics are always staggering. The realism always increasing. The volume continuing to overwhelm their systems.

But what strikes me most isn't the scope of the problem. It's watching seasoned investigators struggle with a frustrating paradox: the very companies that could help them fight this crisis are legally handicapped from doing their best work.

Here's the thing: the people building AI systems want to make them safer. They want to test rigorously for harmful outputs. But when it comes to child sexual abuse material, they're operating with one hand tied behind their backs while bad actors face no such constraints.

The Testing Trap We've Created

Through my work at Zero Abuse Project, I've seen firsthand how this legal limbo plays out. When we develop AI-powered training tools for law enforcement, we have to navigate an intricate web of regulations and partnerships just to ensure our systems won't inadvertently cause harm.

Recent research from Stanford highlights exactly what I've been witnessing: companies and researchers who want to "red team" their AI models for CSAM generation face potential prosecution for the very testing that would make their systems safer.

Red teaming means deliberately trying to break your system. It's like stress-testing a bridge by driving increasingly heavy trucks across it until you find the breaking point. For AI models, this means testing edge cases and adversarial prompts to see if you can trick the system into creating illegal content.

The problem? In the US, there are no federal protections for this kind of safety testing. While there are nuances, AI-generated CSAM is generally illegal, period. So the moment a researcher successfully demonstrates a vulnerability, they've potentially committed a crime.

Beyond Simple Solutions

Some people assume the solution is straightforward: just remove images of children and nude adults from training data. Problem solved, right?

Wrong. And this is where my background in UX design kicks in, because understanding user behavior, including adversarial behavior, is never simple.

New research suggests that even after removing 94% of training data for a particular concept, models can still generate composed images through something called "compositionality." Think of it like this: if an AI learns what "red" looks like and what "car" looks like separately, it can still imagine a red car even if it never saw that exact combination in training.

This means the threat isn't going away with basic content filtering. We need more sophisticated approaches, and that requires more sophisticated testing.

What Congress Could Actually Do

Here's where I diverge slightly from typical nonprofit advocacy. I'm not asking Congress for massive funding or sweeping new programs. I'm asking for something much simpler: legal clarity that enables the good actors to do better work.

The most effective solution would be tightly-scoped safe harbor legislation. Allow companies and researchers to red team their models for CSAM without fear of prosecution, but with clear boundaries about who qualifies and what activities are protected.

Think about how this worked for cybersecurity. In 2022, the Department of Justice created protections for "good-faith security research" under the Computer Fraud and Abuse Act. Before that, white-hat hackers were afraid to report vulnerabilities because they could be prosecuted for the very hacking that revealed the problem.

We need the same approach for AI safety research.

The Reality on the Ground

At Thorn, where I've seen some of the most thoughtful work on this issue, researchers acknowledge they would do more red teaming themselves if the legal landscape were clearer. Instead, they have to coordinate with outside partners authorized to handle CSAM, which slows down the iterative development process that makes software better.

David Rust-Smith from Thorn puts it perfectly: having to coordinate with outside partners instead of following typical development processes is "probably one of the most unique challenges to this work."

I've felt this friction in my own work developing simulation tools. Every safety check requires additional coordination, additional delays, additional uncertainty about what's legally permissible. Meanwhile, the people creating harmful content face no such constraints.

Simple Bills, Real Impact

As someone who's spent eight years watching nonprofit innovation get tangled in bureaucracy, I appreciate that Stanford's research points to a simple model: the REPORT Act.

Passed in May 2024, this bill did one thing well. It extended the time companies must preserve CSAM reports to law enforcement from 90 days to a full year. That's it. No sweeping reforms, no massive new bureaucracy. Just a targeted fix to a specific problem that was hampering investigations.

The bill passed the Senate unanimously and the House by voice vote. When policy is clear, focused, and addresses a real operational problem, it can move quickly and effectively.

The Path Forward

I'm optimistic, but realistic. The current political climate creates strange tensions between AI deregulation and child safety priorities. Some lawmakers see any support for AI companies as capitulation to Big Tech, even when that support enables better safety measures.

But protecting children isn't a partisan issue. And neither is ensuring that the people building these powerful systems can test them responsibly.

The solution isn't perfect. There are legitimate concerns about bad actors potentially exploiting safe harbor protections. But those risks can be managed through careful scoping of who qualifies for protection and what activities are covered.

What we can't manage is continuing to handicap the good actors while leaving the field wide open for those with malicious intent.

Every day we delay clear legal frameworks is another day that safety researchers work with unnecessary constraints while harmful content proliferates unchecked. Our children deserve better than this artificial limitation on the very research that could protect them.

The Personal Stakes

When I attended the AI4 Conference back in August in Las Vegas, I was surrounded by people excited about AI's potential. And I shared that excitement—I've seen how thoughtfully implemented AI can amplify mission-driven work in ways we never imagined.

But I also thought about the children whose safety depends on us getting this right. About the investigators trying to protect them with outdated tools and frameworks. About the mothers like me who want their kids to inherit a world where technology serves protection rather than exploitation.

Congressional action is a start, but it's only a start. The real work happens in how we implement these laws, how we build the systems to enforce them, and how we maintain our humanity while fighting an increasingly technological threat.

We can't let the perfect be the enemy of the good when children's safety hangs in the balance. But we also can't pretend that legislation alone will solve a problem this complex.

The question isn't just whether Congress should act—it's whether we can build the infrastructure to make that action meaningful.