Emergency Episode: Anderegg, Ashcroft & AI CSAM (with Debbie Garner)

Bonus EpisodeCurrent Events
Written By Coursing Justice

Content note: This episode includes discussion of child sexual abuse material (CSAM), AI-generated CSAM, child sexual exploitation, and grooming, including references to the increasing severity of this material. This episode goes further into disturbing specifics than our regular programming. Listener discretion is strongly advised. Subscribe on Spotify | View full transcript

Is AI-generated child sexual abuse material protected by the First Amendment? A Seventh Circuit panel just said yes, for private possession. In this emergency episode, breaking from Season One's MDT-role format, Robert talks with Debbie Garner about United States v. Anderegg, decided August 25, 2026, and what it means for the professionals investigating and prosecuting these cases.

They trace the ruling back through Ashcroft v. Free Speech Coalition and New York v. Ferber and break down what Anderegg actually holds. The decision is narrower than most headlines suggest. Robert and Debbie also explain why the "market substitute" theory behind it misreads offender psychology and cover new evidence that CSAM has turned up inside AI training data itself. It's denser and more legally technical than our usual format, but they make the case for why every MDT member needs to be tracking this issue right now.

About our guest

Debbie Garner is the Community Engagement Coordinator at Hexordia, where she connects digital forensic examiners and law enforcement with essential training and resources. With 30 years of experience from the Georgia Bureau of Investigation (GBI), she retired as Special Agent in Charge of the Child Exploitation and Computer Crimes Unit and Georgia ICAC Task Force Commander. In this role, she managed both sworn and non-sworn personnel and oversaw investigations that led to over 2,500 arrests.

A leading advocate for technology-driven innovation, Debbie also works with digital forensic tool developers to align the voices of law enforcement and technology providers. She is an instructor of wellness and resiliency skills for ICAC professionals through the Innocent Justice Foundation, and she is a data loss prevention investigator for a global company. A graduate of the FBI National Academy (Session 267), she is also a Founding Board Member of Raven.

Key moments

  • 3:00 — Meet Debbie Garner: former GBI Special Agent in Charge, Georgia ICAC Task Force commander, 2,500+ arrests

  • 8:00 — Ashcroft v. Free Speech Coalition strikes down the ban on "virtual" child pornography

  • 13:00 — The Anderegg decision: why this is about possession, not production

  • 15:00 — Debbie corrects the misleading headlines circulating about the ruling

  • 18:00 — Justice Thomas's 2002 warning: "that day may have arrived"

  • 22:00 — Taking apart the "market for CSAM" argument at the center of Ashcroft

  • 26:00 — Debbie: real CSAM "is becoming more violent, the victims are becoming younger" — and so is the AI-generated kind

  • 32:00 — CSAM found inside AI training datasets — and what that does to the "no real child" assumption

  • 43:00 — Thorn and Save the Children data on minor-on-minor AI-generated abuse

  • 57:00 — Robert's closing argument: "Liberty doesn't require naivete"

Mentioned in this episode

Want to learn more?

  • Explore Coursing Justice's artificial intelligence trainings, ranging from hour-long virtual webinars to full-day sessions, always co-presented with law enforcement.

  • Follow Hexordia at hexordia.com, where Robert and Debbie both instruct alongside Jessica Hyde.

  • Join Coursing Justice's free Justice in 30 webinar series.

Episode Transcript

*** Please note that this transcript is AI-generated and may contain errors.

[00:00:00] liberty doesn't require naivete. It doesn't require technological ignorance. The victims are becoming younger, and the content is becoming more violent and so is CSAM.

While we pontificate, there are very real harms being conducted here. It is some of the most violent CSAM that we see

Clearly AI CSAM can and does accompany sexual predation for all the obvious reasons.

They want to abuse a child.

Ashcroft needs to be reversed, full stop.

Welcome to Coursing Justice, the podcast, a program of Coursing Justice LLC. We're a consortium of current and former law enforcement, prosecutors, and CAC/MDT members serving the field through innovative training simulations and practical content for justice and child protection professionals. I'm your host, Robert Peters, former prosecutor, attorney at law, and founder of Coursing Justice.

And we are here for the first emergency podcast of Coursing [00:01:00] Justice's tenure, interrupting our regularly scheduled programming. I don't think anyone knows what that phrase means anymore, Debbie. But once upon a time, listener, if you're under 30, there was such a thing as regularly scheduled programming.

And now it's a meaningless term. But our regularly scheduled programming season one of the podcast, is the multidisciplinary team featuring individual MDT roles and doing a deep dive with folks in each of those.

Now, before we get into this discussion, just a quick note that I am barred in the great state of West Virginia, but none of this is legal advice. And today's episode is sponsored by the Robert J. Peters Law and Justice Firm PLLC, newly created in West Virginia, as I devote more time to the practice of law.

But I am not your attorney. I'm an attorney and so always talk with your local prosecutor or counsel for legally actionable information. None of this is or is intended to be legal advice. Another disclaimer I should note is that this is very much, as the title indicates, a [00:02:00] rapid reaction segment.

And so certainly there are nuances that we have not yet processed in our desire to have some reflections for what this case means out of the Seventh Circuit as we've seen quite a bit discussion. And so hopefully this helps bring some clarity to that. I should note that today's segment goes way more into the weeds on legal issues than we will likely do in other episodes.

And so this is very much an unusual episode. If you absolutely hate it, please come back for others, as this will not be our normal format. Unless, you know, we hear the opposite, that people love it, which I, I greatly doubt, as the more I think we get into the legal weeds, probably the less compelling it is.

But in any event, you know, feel free to skip over that if that is not your cup of tea, and we'll be back with our regularly scheduled programming with subsequent episodes focusing on specific MDT members in future episodes

but today we're gonna focus more on a decision that just came out of the Seventh Circuit Court of Appeals here in the United States the case of Anderegg.[00:03:00]

I'm assuming I'm pronouncing that correctly, but I don't know. I believe you are. All right. And with me to discuss that case is a former ICAC task force commander, Debbie Garner. Debbie is a technology advocate working in the private sector and nonprofit arena, mobilizing a movement and educating law enforcement officers and executives about solutions for modernizing their agencies.

Having stalked your LinkedIn for far too long, Debbie I know you're a proponent of the, the so-called nerd cop. Absolutely ... which is a very cool kind of cop, I'll just say. I, that's what I think. I am, we in the office that I used to supervise the task force, we would embrace the nerd. I think that's the only way to be is to embrace our big brains that have chosen this particular profession.

I think it's worth celebrating Yes, absolutely. Yeah, it's a huge gift. There is no successful ICAC investigation or prosecution without the army of nerd cops, so- Exactly ... thank you for your service. As you just [00:04:00] mentioned or, or, or alluded to, Debbie Debbie served as the special agent in charge of the Georgia Bureau of Investigation Child Exploitation and Computer Crimes Unit, and was the Georgia ICAC Task Force commander.

And so she oversaw the state of Georgia's efforts to detect and investigate child sexual exploitation cases. And that included in her tenure, the arrest of over 2,500 offenders. So very, very good stuff. And Debbie, thanks so much for all you've done in this field, and for joining us for a pretty significant update.

The hope is, of course, that it leads to a much more significant update that would be a positive thing and lend some clarity in this field. And that is this Seventh Circuit case of Anderegg. Now I try to find ways to find a little bit of lightness without making light of the heaviness of what we talk about.

Good way to put it. It's important. It, it's a, it's a coping mechanism certainly. But obviously never wanna make a light of the material itself. But also gotta, gotta laugh or you'd cry at times. So in an effort to bring a little bit [00:05:00] of levity to our discussion I queried our AI overlords as to what Dr. Seuss would think about this situation. Oh, nice. And so Dr. Seuss says as follows. We could apply this to Anderegg, but I'm gonna shift it to Ashcroft 'cause I think Ashcroft is the real problem here.

But in any event Dr. Seuss says, "I do not like this Ashcroft case. I do not like its paper-thin base. I do not like the logic it bends. I do not like where that logic ends. I do not like the harm swept aside. I do not like the gaps open wide. I do not like the rights they embrace. I do not like it.

What a disgrace." Oh, wow. So thank you, Dr. Seuss. Nice. Okay, so I'm gonna use that for other things, yeah, it is a solid prompt. I like that Well, let's, let's go ahead and dive in.

I wanna talk a little bit about some of the background law that plays into this case that the Andre court was grappling with. And so let's rewind. Let's go back to, you know, a couple decades, 24 [00:06:00] years, give or take. I think 2002 was the year. As I get older, that doesn't seem as long ago It does not.

But there were far, far fewer Marvel movies back in the day. It was a very different environment. There's a case called New York v. Ferber, United States Supreme Court case. Ferber actually was before, Ashcroft was 2002. So we have to go even farther back.

But in New York v. Ferber the Supreme Court found that the state has a strong, or recognized that the state has a strong interest in protecting minors, which is good. Now, it introduced this phrase that may be jarring if, if you're not familiar just with the legal analysis around these issues at the federal level.

I should note at a high level here at Coursing Justice, we subscribe to the use of the term child sexual abuse material to adequately f- reflect the severity of the material. The term child pornography is used extensively in these decisions, so I may reference that where it's accurate to the decision.

But just note, child sexual abuse material is the term we should all be using. So use that [00:07:00] unless accuracy purposes. But in Ferber the court found that the First Amendment does not protect certain types of speech because of the inherent harm of that speech. And so it applied that specifically to what it termed child pornography or child sexual abuse material possession.

So a good case, from a prosecutor perspective, and I think just from a decent society perspective. We fast-forward to the crux of our current issue that, that many would say, myself included, sort of tied the Seventh Circuit's hands here. And that is the case of Ashcroft v. Free Speech Coalition.

And so that case was responding to a 1996 federal law. Again, I'm using the name of the law, the Child Pornography Prevention Act. And so that law added these categories of so-called morphed child pornography and virtual child pornography to a list of prohibited pornography depictions in, in addition to, quote-unquote, actual pornography.

And again, I just want to note, [00:08:00] these are all terrible terms, and I hate them, but they're the ones that are utilized and so I'm using them for accuracy. But what the act referred to as virtual child pornography is material that is not depicting a, quote-unquote, actual child, but some other representation.

So at the time, it was these really flimsy computer graphics cartoons. And that's- Anime. Yeah. Exactly. And so that was, you know, the, the material of the day in, back in 1996. Then you have this category of morphed, so-called morphed child pornography, which is some blend of the two. There's some use of an actual an image of a, of a, quote-unquote, actual child that's utilized as part of it So in Ashcroft v.

Free Speech Coalition, the U.S. Supreme Court strikes down that law, says it's too broad, and it sort of argues that even written depictions say, of Shakespeare's Romeo and Juliet would run afoul of this act because arguably it's some sort of [00:09:00] depiction. There's a, quote-unquote, "relationship," you know, between minors, and so that constitutes in, in their words, quote-unquote, "virtual child pornography."

So it strikes it down as overbroad. In response, Congress passes the PROTECT Act, and the PROTECT Act is directly implicated by what we're discussing here. Now there are a few shreds of hope within the Ashcroft decision specifically that quote-unquote, "morphed child pornography falls closer to real child pornography," again, their terms.

And just apply that to the rest of this episode. These are all terrible terms, and they're not Robert's. But because it harms the interest of a, quote-unquote, "real child." And so that's the posture that this has been in for, for a while. And interestingly, in the PROTECT Act that Congress passes right after Ashcroft there are a few legislative findings.

Some of those have not aged particularly well. One of them being quote, "There is no substantial evidence that any of the child pornography images being trafficked today were [00:10:00] made other than by the abuse of real children." So that doesn't work anymore. Secondly another finding was that, was, quote, "To the extent the technology exists to computer generate realistic images of child pornography, the cost in terms of time, money, and expertise is, and for the foreseeable future will remain, prohibitively expensive."

Mm. That also is no longer accurate. Right. There's a, a proliferation. So maybe, you know, that's a great point to hear from you, Debbie. Talk to us about why those things are no longer accurate from your perspective in law enforcement spaces. Well, you know, technology, we all know that, there have been, iterations of, say, computer generated CSAM.

We'll start with, what you were talking about. There were drawings, not necessarily computer generated. There was cartoons, there was anime, and those things started to be computer generated. Then it would be things like placing a child's face. It was very [00:11:00] crude, but you were placing a child's face on an adult's body in actual adult child, or in actual adult pornography.

And then there were some crude Nudify-type apps and programs that were used. And so we started seeing this, this evolution of child sexual abuse material that sometimes was crude, but now is not. Often now it is indistinguishable from, real child sexual abuse material.

Some of the practical things that occur because now This is gonna be a crude way to describe this, and for lack of a better way to do it, you don't have to have a real child anymore to create something that appears to be CSAM.

The volume of, of CSAM that is that is being reported to NCMEC, then passed on to the ICAC task forces, well, it's already unmanageable. It just [00:12:00] increases that workload even more. , In the confusion between some of these cases and law enforcement trying to decide is this a prosecutable case, it adds another step, sometimes a pretty long step.

If we cannot distinguish between artificial generated CSAM and real CSAM, then do we even move on with the case? There are laws that that apply in most of the circumstances where we can prosecute those, but it, but we have to determine whether it's actual CSAM or, or generated, you know, artificial intelligence CSAM.

So the technology exists and, the things that they were arguing have real consequences in when we don't have enough resources to investigate what's occurring now. And then, you know, having some of this, , possibly not even be prosecutable, then that is very [00:13:00] concerning

but the technology exists. It, it absolutely exists now that brings us to the Andurag decision. And so the Andurag decision, this is a Seventh Circuit case, so that would be Illinois, Wisconsin, forget where else. I don't remember. There's a third state in there somewhere, maybe Indiana.

But the, it, there's two opinions here. One of them I don't really like and one of them I actually think is, is pretty defensible. Ironically, the same person wrote both opinions, Mm ... Judge Lee, which is interesting. And it's only a three, three-judge panel it looks like, so I'm curious what the discussion behind the scenes was.

But the... And he's joined by the by Judge Co- Kohler, I'm probably mispronouncing that, Kohler or Kohlar, in the concurrence. So it's two out of three judges in this concurrence. I think what's fascinating about this, it, so it reaches the decision that possession, key word, key limitation- Right

possession is [00:14:00] constitutionally protected. Possession of artificially intelli- artificial intelligence generated child sexual abuse material is protected by the First Amendment. Possession alone. Can you speak a little bit, Debbie, about, you know, we're already seeing some sensationalism about the breadth of this holding.

You know, what have, what have you seen out there? I know this has been a hot topic. So I have, you know, I've got Google Alerts set for a variety of different things, this was decided last week maybe, week before. I think so. Argued in October and decided August 25th.

And there are still articles coming out even today in my Google Alerts and the headlines are things like, "Federal court says AI-generated child sexual abuse materials are protected by the First Amendment." The next one is, "How AI-generated child sexual abuse material is protected under the First Amendment."

And those are misleading headlines based on really all Andurag is saying is that possession in your own home [00:15:00] of AI-generated CSAM is protected under the First Amendment. It's not saying that you know, manufacturing and, a- and, you know, he- in this particular case, he sent some of this to a minor. Right.

So there were other, there are other things that he did related to generating this artificial intelligence generated CSAM that are, that were not thrown out. It was only the possession in his own home that they said that they could not, that, that was thrown out I would suspect that the future is not bright for Mr.

Anderegg just based on the other charges. I agree. I, I think the other charges are gonna stick. It's interesting that these were not the only two articles, but there have been a lot of articles and even a lot of discussion. I'm very active on LinkedIn, and a lot of people are posting about this saying that this court indicated [00:16:00] that AI-generated sexual abuse material is not illegal, and it's not what it said.

Correct. Correct. Now I think the concurrence itself is fairly hard to argue with. I, I will maybe, time permitting, I will argue with some aspects of the opinion itself. I think the opinion itself, so for those who don't know, aren't aware, when you have a decision that's given by a court at this level, there's a majority opinion, right, that the majority of judges have signed on to, and then you can have dissenting opinions.

There's no dissenting opinion here. And then you have concurring opinions. And so concurring is I agree with it. I'm, I wanna emphasize something about that judgment, or maybe I wanna take issue or nuance some part of that judgment. So again, the fact that they're written by the same person is sort of interesting.

It is interesting. That does happen elsewhere, but it is interesting. So maybe, maybe the third judge wasn't willing to sign on to some elements of the concurring opinion. That's, that's the most likely probability. But in any event, [00:17:00] the, the concurring opinion I think is pretty hard to argue with, and it essentially points to our old friend Ashcroft and says, "Look, if we're going to apply Ashcroft to this, this forces our hand in terms of what the outcome is."

But in, in something that's kind of unique or at least, you know, it, it doesn't happen the majority of the time, they don't love the decision they're reaching here. At least that was my impression. Was that reciprocal? Yeah. They don't seem to love where this is going, and they acknowledge that it's a flawed result, but they're applying Ashcroft and they're pretty explicitly asking the Supreme Court to revisit this.

One thing I think is really important is they point to Justice Thomas's concurrence from Ashcroft v. Free Speech Coalition itself. 'Cause Justice Thomas Said all the way back in Ashcroft- Mm-hmm ... you know, decades ago, quote, "The technology may evolve to the point where it becomes impossible to enforce actual child [00:18:00] pornography laws because the government cannot prove that certain pornographic images are of real children."

And in Justice Thomas's view, quote, "If technological advances thwart prosecution of unlawful speech, the government may well have a compelling interest in barring or otherwise regulating some narrow category of lawful speech in order to enforce effectively laws against pornography made through the abuse of real children."

And I think that speaks, I think Justice Thomas's concurrence, Debbie, is directly on point to what you just described. Like, if this is the new reality where everyone does have a constitutional right, and let's even keep it with just possession. You know, to possession of AI-generated CSAM, how would that complicate, you know, your work and others in law enforcement?

What would that do to your ability to investigate these, these cases of of a- of, you know, known victims? Exactly. I mean, it, it does make it much more difficult. The, not just the volume, but the technical, like, I know, you know, we have both sort of [00:19:00] worked in digital forensics world or with digital forensic, you know, companies.

We're commonly in touch with training providers and those who are creating digital forensics tools and using them. And there are tools that are going to help tell us if something is AI-generated, modified, or, you know, an actual image that has not been modified. So tho- but those are not perfect tools yet.

So, you know, we're, i- it's, we're, we're, we're progressing along this, this challenge, but it's, it's going to be a real challenge related to volume and not enough resources, and then also just the technical, mm, tools that are available to help us prove one way or another what something is. You know, it, it's interesting when you were, when you were reading the Justice Thomas's [00:20:00] comments you know, right, right under that in the concurrence, it says, "According to some experts, that day may have arrived."

Yes. I'm super hopeful that the justices who decided this case, they absolutely realize that this day has arrived and that we're going to need to revisit this issue. So I am hopeful in that regard that that this will actually make its way to the Supreme Court.

Yeah, it needs to. It needs to. And I'm, I'm hopeful as well If not, there will have to be a legislative, a legislative fix. Yeah. You know, if if if Congress will, will approve that. But I, I would imagine that there will be some efforts for a legislative fix if the Supreme Court doesn't agree to hear this.

Yeah. Yeah. Which I think they will. Don't you think they will? I would think they will. Don't you think they're gonna take this up? I would think they will. There's various ways this could go, of course. Different judges obviously have different [00:21:00] philosophies. And not all of them are necessarily Tailored to providing proactive guidance, shall we say, so it'll be interesting what, what happens with it. I think one piece that's going to be a fight at the Supreme Court level is this terrible, terrible argument, in my opinion from, it would've been Justice Kennedy, I believe from the, the Ashcroft case. And this came up in this came up in Anderegg as well, and it's one of the things that they- Yep

with in the main opinion. And so it's this discussion of the, of the market. And so I think that argument is problematic in a variety of ways. Again, they reject it pretty, pretty straightforward. If, if people are trying to follow along, it's on page 22 of the Anderegg decision. So they're pointing to the different arguments that the government made in terms of affirming this conviction for possession.

And so in Anderegg, the government argued that prohibiting the [00:22:00] possession of, of AI-generated CSAM would help reduce the commercial market for actual child sexual abuse material. So the reason why they're arguing that is that's pulled, that language is pulled directly from Ashcroft. That was a argument that Justice Kennedy had made in Ashcroft for striking down the CPPA i- in the '90s.

So let me back up for a second. The argument from Ashcroft is essentially as follows. The first premise is ... And there's so many holes in this, but bear with me. The first premise is that virtual child sexual abuse material, again meaning child sexual abuse material that's produced not using an actual child, and actual child sexual abu- abuse material, there's a, a, an assumption there that those satisfy the same consumer demand.

And so to attack the demand for one, you need to attack the demand for both, or so the argument goes. And then so the second premise of [00:23:00] Ashcroft is that producing actual child sexual abuse material entails greater criminal risk. Likely true. The third premise, and this is where the argument falls apart just from a logic perspective, rational suppliers prefer the lower risk substitute.

Now you'll see that right here in this is one of the problems with Anderegg that I take issue with, is it repeats that here. And they kinda have to 'cause it's precedent. Supreme Court's a higher court, so no offense, Justice Lee, I think. But, but they repeat it here where they say, you know, and, and it's referring back to Ashcroft that the Supreme Court squarely addressed and rejected that argument, finding it implausible because the Supreme Court said, oh, if the demand for virtual child sexual abuse material or, or AI-generated in our case here, were the same as for actual, few people would create actual child sexual abuse material and risk prosecution by abusing- Mm

real children. Fictional computer. You're- Mm ... listener, this will be audio, but Debbie's shaking her head right now. All right, [00:24:00] Debbie- i'm not choosing my words carefully, but the people that are creating child sexual abuse material I mean, they are, they are doing it for their own sexual gratification. They are molesting children. That is their fir- that is what they want to do first. They just happen to be filming it so they can watch it later for additional sexual gratification or trade it with somebody or send it to somebody , they want to abuse a child.

Yeah. That argument is flawed. It is. It is. It was flawed then and, and it's flawed now. And so the reality is in, in the wild, and I don't know, you know, to, to what extent the... I haven't thought through the legal ramifications of this reality, but the reality is fro- from what I'm hearing out there, and I'm curious of your take, Debbie, where you're seeing AI-generated CSAM, it seems like most of the time you're seeing actual CSAM.

Are you hearing that or, or not so much? No, I, I absolutely believe that's true. I, you know, as- I, I think there's a [00:25:00] lot of different scenarios here. There are, there have been people who have been arrested with only Gen AI CSAM. Now that doesn't mean they didn't have something elsewhere, but that, that's what they were charged with.

It was only Gen AI CSAM. But more often than not, they do have both. And one of the things that they're doing is they are, you know, if they can't, again, my words not going to be chosen carefully, if they can't get what they want to see in, you know, like by going on, going in chat rooms and, you know, joining groups and, you know, "Hey, do you have anything depicting this?"

Or they just create it themselves. And I will say while, while one of, one of the issues that I see that sort of relates to this is that real, you know, CSAM is using a- you know, with actual children is becoming more violent. The victims are becoming younger, and the content is becoming more violent, and I, and [00:26:00] so is generative Gen AI CSAM.

It is some of the most violent CSAM that we see, and I be- I believe that that's going to be a cyclical process. You know, one is going to feed off of the other, and that is a real worry. Definitely. Definitely. Yeah, I, you know, by this same logic, I would think that the argument is, You know, I, I think if you apply Justice Kennedy's reasoning, then the, then the use of 3D-printed ghost guns should have solved firearm issues.

Mm. I, I, I don't see a distinction between those arguments. You know, the, the psychology of this is just not taken into... I mean, it's making that assumption that, again, that rational supplier prefers a lower-risk substitute, and I think that fundamentally misunderstands the psychology of offenders. Yes.

Yes, that's, that's what I was trying to say, but yes, [00:27:00] you put it much more succinctly. It does not hold to the psychology of offenders. Because if we look at the progression of offenders, sometimes they will start out, you know, viewing CSAM and progress into actually abusing children, and other times they were already abusing children and now they're like, "Hey, I have this phone in my pocket.

I can actually create CSAM." Then they move to gen AI where, the psychology of it is very different. Yeah. Yeah. And based on some of what you were discussing these markets are absolutely linked, right? Both because- Yeah ... of just seeing out there in the wild, but also if you make it, and as Justice Thomas noted, you can make it impossible to actually investigate and prosecute legitimately.

So what, I mean, you're the prosecutor. What, what do we think about what the, what, what would the legal arguments they would make? Like, if, if it becomes impossible to [00:28:00] prosecute, there has to be some sort of fix for that. They don't want us to not be able to, I believe. I've been known to be naive, but I believe they don't want us to not be able to prosecute CSAM offenses.

I think the problem here is some lines were drawn in Ashcroft that sounded great at the time, I'm operating from a, a cynical posture currently 'cause I spent way too much time studying law over the summer for a variety of reasons.

But, but the, you know, as lawyers, my- myself being one of them, you're looking for these lines, right? You're looking for categories that you can sort of evenhandedly apply across cases, and sometimes those feel really good and, and conceptually sound. Right. But then when you apply real life to them, it just doesn't work, and I think that really is the case with Ashcroft.

Like, those lines, you know, felt accurate to Justice, Justice Kennedy quite a while ago in a very different [00:29:00] technological context, and they don't work anymore. And I think it's a mistake to stand by that decision for all the reasons we just mentioned. If literally anyone can just say, "It's AI-generated," and we can't tell- Right

In the interim, you know, that's- Maybe that's what I was getting at, yeah. They... I don't think they want us to not be able to prosecute CSAM cases. I would hope not. They've gotta fix it. Yeah. I would hope not, certainly. And I think, you know, if I were j- if I were Justice Counting, and I'm certainly not the most a- attentive Supreme Court observer, but if...

I think if I were Justice Counting, I think it's more likely or not that this gets reversed for the reasons that the different, you know, concurrence rolls through. Now I think one thing that I really have to give the concurrence a lot of credit for is it, it actually made some arguments that it doesn't appear that the, the government did in this case in terms of some of its reservations around how Ashcroft or Free Speech Coalition is being applied here.

And one of those [00:30:00] is this issue of, This issue of whether or not, quote, "a, a, a real child is used in the making of child sexual abuse material." And sometimes that's pretty straightforward. We are seeing quite a bit where, you know, you have fine-tuned AI models that are generating, You have fine-tuned AI models that are generating thousands of images of children.

They're, they're developed and created to produce images and videos of specific children, including of children whose images and videos were theoretically scrubbed from the internet, so prior victims of CSAM, and there's now whole new episodes that didn't happen, but it's of them. So you have sort of that category that's out there.

Then you have, you know, the, the, the nudification apps where, you know, many, many youth or a lot of peer-on-peer usage of that type of technology because of the, the problematic regulatory frameworks and sort of Wild West in, in, in some areas. And so there's a lot of different, you know, ways in which this is [00:31:00] utilized.

But then you also have, and this has been documented in at least a few instances, the fact that as the court in Anderag notes here in the concurrence, actual child sexual abuse material has found its way into training data sets. Mm. Now, it's hard for us to always know which ones, like which- Right

which specific data sets, which LLMs or which, you know, stable diffusion models, et cetera, are, are doing this. I think one thing for a listener to understand is there is no single AI that just pumps out stuff, right? Right. There's so many different models that are fine-tuned in so many different ways, and so the training data set that each of those is trained on is different.

And so- Right ... one of the challenges here is, and I, and I suspect this is gonna be an ongoing issue, is if the burden is on the government to show that there was actual CSAM in the training model for the AI, such that the CSAM that was produced did [00:32:00] use actual children because actual children were in the data set.

I think that's a sound argument. But as you've noted previously, it, the burden of proof on that is tough. It's gonna be super tough. I, you know, I don't know how we're gonna, how we're gonna prove that, but I, but it is, it is a real issue. It's something that I, I have, I I have no doubt that that's going to be something that we're gonna have to address because I do think somebody's gonna make the argument, right?

Some- somebody is going to make that argument in court. I, I would, I would think. But m- Debbie's opinion. I would think that somebody's gonna make that argument in court, and they're gonna have to have something to back it up. Yep. Well, Debbie, I know you are in demand in a variety of ways. Also, it's late on a Friday-

so thanks for jumping on an emergency pod at the end of the week Not a problem, not a problem. It is a very important case, and I know it it had been being watched by many, many people prior to the decision.

And I do believe that if the Supreme Court doesn't take it up, that there will be some [00:33:00] legislative fixes that will try to make their way through the congressional approval process. Yeah. It has to happen one way or the other, whether it's reversal through a reconsideration of Ashcroft, which would be my preference.

I think that's gonna happen. I, I predict. No, I, I think that that is actually what's gonna happen. I like your crystal ball. If I had to gamble, that's what I would predict. I think y- just even the Seventh Circuit's sort of half-hearted defense of the conclusion, like recognizing- Right ... the issues of it is, is its own encouragement, I think.

But yeah, we shall see. It's definitely something for us to watch, and it could have really far-reaching application. Debbie I've got some additional attacks I'd love to leverage on Ashcroft, but I won't subject you to them since you've been kind in sharing your time this afternoon.

We need to have you on once the regular programming comes back. We need to find a, a good topic. Maybe it will be when the Supreme Court reverses this. That would be a great one. Absolutely, I'd love to come back and say that we [00:34:00] were right. Yes. Fingers crossed. Let's- Yeah, exactly ... we're, we're manifesting it.

We're speaking it into existence. Exactly. Putting it out there. Thanks so much, Debbie. Thanks for all, all you've done and all you do- Absolutely ... work. Thanks for the great LinkedIn stuff. If you don't follow Debbie on LinkedIn, definitely check her out. And I would be remiss if I didn't note Debbie, that you and I are both instructors for a great company called Hexordius- Yes

where we're regularly bringing on issues like- Yes. Lo- love working for Jessica. Yes, yes. Jessica Hyde. Brilliant and very engaged on this issue as well. And I should also note, if you're listening to this and you are a federal prosecutor or a nonprofit addressing these issues I would certainly be happy to help however, whether that's amicus briefs this is an issue we absolutely have to get right, so certainly reach out if that would ever be of use.

Thank you, Debbie. Absolutely. Thanks, Rob.

Thanks. Let's recap and expand a bit about some of the issues we just touched on with Debbie regarding the Anderegg and by extension, Ashcroft v. Free [00:35:00] Speech Coalition decisions. For those who don't know, circuit courts of appeal are below the United States Supreme Courts. These are federal circuits. They're below the US Supreme Court, and therefore, they are bound to apply precedent, the decisions that the US Supreme Court has previously rendered.

Now, Ashcroft has some serious problems. I've hinted at that a few times. I think it's wrongly, wrongly decided in a variety of ways, and so it's certainly understandable that because it is binding precedent on the Seventh Circuit Courts of ap- Court of Appeals, it's understandable that they are in a bind here.

However, I would argue that there are some contingent assumptions from the year two thousand and two that, AKA the time that Ashcroft w- was decided, that are being applied to the modern context as if they're ironclad constitutional holdings, when in reality, there's a temporal nature to some of those findings.

And some of those assumptions I think now are, are baseless or at least without adequate foundation because of how the [00:36:00] technology has developed. We talked a little bit ago about Justice Thomas' concurrence from Ashcroft, and that shows how at that time, at the time Ashcroft was decided, at least some on the Supreme Court contemplated the developments in tech complicate the analysis and likely change the outcome.

Anderegg did explore that in the concurrence but didn't embrace that result in the opinion itself. Again, I understand why. I should also note some of Anderegg's conclusions are too simplistic. It characterizes some of the aspects as the argument being flatly rejected by Ashcroft, when in reality it was a specific factual circumstance or nuance, and there's more nuance between the two situations in terms of both time and technology that are relevant to the outcome, and Anderegg didn't fully explore some

Now, I can't help but take another swing at this discussion of the market, the, quote-unquote, "market for child sexual abuse material," or [00:37:00] AI-generated child sexual abuse material. I have just always found that to be a particularly weak argument. It's a weak argument in Ashcroft, in Justice Kennedy's opinion, and I think it's an even weaker argument in Anderegg because of the new factual context.

So for those that aren't avid consumers of constitutional law, the Seventh Circuit in Anderegg said that Ashcroft, quote, "squarely addressed and rejected the argument that prohibiting so-called virtual child sexual abuse material could diminish the market for actual child sexual abuse material." In other words, the government's arguing that, you know, we need to prohibit, you know, AI-generated CSAM because if we do that, we can help diminish or reduce the market for actual child sexual abuse material.

And so the Seventh Circuit is saying Ashcroft already considered that argument and rejected that argument. And in doing so, the court in Anderegg repeats Justice Kennedy from Ashcroft, who said that that [00:38:00] theory is, is, quote, "somewhat implausible." Justice Kennedy said that if synthetic material and actual material were actually interchangeable it's not a direct quote, it's a paraphrase, why would producers of child sexual abuse material risk abusing real children when fictional images would satisfy demand?

That's Justice Kennedy's argument. I can't help but kind of crack up saying it because I think it's ridiculous. You heard Debbie's rejection of that argument. Again, I think it's absurd on a variety of levels. If you work in this space, and particularly in investigation or prosecution, you're probably, probably already tracking with me.

If not, maybe not. But I think to pinpoint, you know, some of this logically, I think Kennedy's reasoning assumes several facts without establishing them. So first, he's assuming that consumers of child sexual abuse material consider actual and synthetic child sexual abuse material as substitutes, or they regard them as sufficiently close [00:39:00] substitutes, that there is an indifference to authenticity in that synthetic material.

I think this is the big assumption, actually. The synthetic material will not be complementary with actual material. So he's making that big assumption, that synthetic material will not be complementary with actual material. That's just not true. That's not what's happened. As we discussed earlier, you know, the cases in the wild that I'm aware of at this point, these things are existing concurrently.

We see that even in Anderegg, not so much the, the AI generated alongside the, the original child sexual, sexual abuse material. But what we do see is the sexual predation alongside the AI-generated CSAM, and then we're arguing over whether those, those-- that's relevant, which we'll get into in a second.

So that's one assumption that Kennedy's making. Another assumption Kennedy is making is that actual child sexual abuse material will not become an input into the production of [00:40:00] synthetic child sexual abuse material. Now, I don't fault him for that. The year was two thousand and two. The tech is much different, but clearly that is happening, which we'll get into in a second.

Again, we-- I don't know that we can always prove that that's the case, but that's a, an assumption that he made that's just not accurate. At a high level, Justice Kennedy w-- in Ashcroft was assuming rational actor and risk aversion. He's basically doing economic analysis. I think that's absurd. It doesn't have a corollary when it comes to basic offender psychology.

It's not rational to abuse a child. It's not rational, and you're not doing that in this, you know, this sort of economic way. That's not how the psychology works. So, I should note on that, that final assumption that actual child sexual abuse material won't become an input. Again, that's particularly problematic in the year twenty twenty-six, because in Anderegg, the court [00:41:00] points out that researchers have found child sexual abuse material in major AI training data.

And so it's relevant that AI-generated demand could affect demand for actual child sexual abuse material used to train models. Let's do, since Justice Kennedy is all fired up about the economic stuff, let's do some math. It'll be short and simple because I'm an attorney and I am not good at math.

But let's assume for a second, bear with me we have the symbol M. Again, I, I did not love algebra, but let's assume M for market. That's the total market or consumer population for CSAM-like material. And then let's assume the letter S is the percentage of that demand that's directed toward actual child sexual abuse material, some smaller percentage, 5%, 10%, whatever we wanna say it is.

And then so our, our actual demand for child sexual abuse material becomes M times S. Now again, this is absurd, but you know, we're, we're now [00:42:00] in this living hell of economics according to the U.S. Supreme Court, so let's, let's go down the rabbit trail with them for a second Logically, even if AI causes S to fall, that percentage of the demand, even if the percentage of the demand towards actual CSAM falls because of the introduction of virtual child sexual abuse material, it doesn't follow that the actual CSAM demand falls if artificial intelligence is drastically expanding the market.

If it's drastically expanding M, even if the percentage S is lower, the actual CSAM demand M times S, where that material is made cheaper, customizable, private, instantaneous, accessible, it actually increases. So again, to prove this out, and again, I just... Listener, I find this all very distasteful, but let's, let's stay in the twisted world of, of the Supreme Court for a second.

If we have a, a market of one [00:43:00] hundred consumers and that entire market, so that's our M, and then our percentage of the market that's directed towards actual, actual child sexual abuse material is a hundred percent. So we have one hundred units total. Now let's introduce AI, drastic expansion of consumers.

We'll say a thousand consumers. But let's say this introduction of AI-generated CSAM means that the, the percentage factor is only twenty percent now, twenty percent. Let's double the total demand, even assuming twenty percent. That's two hundred units. A thousand times twenty percent equals two hundred units.

And one thing that I, I, I think too few are reckoning with so far, one aspect where the market for child sexual abuse material has absolutely increased is minor-on-minor incidents. This was either twenty twenty-four or twenty twenty-three. There was a study by Thorn where one in ten minors reported that they knew of friends or classmates who used artificial intelligence tools to generate nudes of other kids.

In July of [00:44:00] twenty twenty-five, one in five young people in Spain reported being victims of AI deepfakes, with almost all reporting sexual violence online. That's a Save the Children study. That's a huge percentage, a huge percentage of youth that are now aware of this, not just aware of it, but being, you know, victimized by it or being exposed to others' victimization by it.

AI has drastically increased the market for artificial intelligence generated child sexual abuse material. But also of child sexual abuse material itself. That's the reality that our kids see all too often, and we can pontificate about economics and formulas and is it affecting the market, is it not affecting the market?

While we pontificate, there are very real harms being conducted here. I think that's something that, to its credit, the Anderegg- Anderegg decision's concurrence notes it's critical for this to go the right way, you know, at the next level. So we're seeing these cases in the wild, all these things going together, right?

Anderegg sent AI-generated CSAM [00:45:00] to a minor. Clearly AI CSAM can and does accompany sexual predation for all the obvious reasons. We expect to see that. That is just extremely logical and, and not something that we should have to hash out. There's a symbiotic relationship which w- between, or there can be, a symbiotic relationship between child sexual abuse material and AI-generated child sexual abuse material.

Those seeking to abuse children can customize their abuse and their fantasies through artificial intelligence. They can disguise and seek essentially immunity for abuse. They can fine-tune AI models to generate new sex acts a- and abuse based on prior images acquired of victims. These things all go together.

Now, it should be noted that different offenders have different preferences as it relates to AI-generated child sexual abuse material. There was some really interesting information published by the Internet Watch [00:46:00] Foundation in, back in twenty twenty-four, so relatively early on in sort of the proliferation of, of child sexual abuse material.

And so there were different perspectives on that

One of these perspectives, it should be noted, is that there's a category of offenders that are very much against the use of artificial intelligence. It is not of interest to them. So in that same Internet Watch Foundation, they were pulling these quotes from dark web forums where offenders gather both to share illicit material.

It should be noted both child sexual abuse material and AI-generated child sexual abuse material, yet another hit against Justice Kennedy's interesting opinions about markets. And as part of that, we get some insight into how they feel about the use of artificial intelligence in CSAM generation. And so you have these quotes [00:47:00] from offenders that are talking about, you know, being annoyed that they zoomed into something that they thought was CSAM and realized it was AI and talking about how they only wanted, quote, the real stuff.

So that certainly is a category of offender, those that are simply not interested. I suspect the same as adults. There may be some adults who are interested in AI-generated adult pornography and may and others who are not interested at all in AI-generated anything. Obv-obviously different context, but likely similar distinction.

You also have pro-artificial intelligence perpetrators, those who talk about how exciting it would be to live in a world where their favorite child actors, for example, and this is a direct quote, "Never aged beyond their most cutefulness." And they share strategies for finely tuning models and finely tuned models and the ability of that to generate custom child sexual abuse material You also have the discussion [00:48:00] amongst perpetrators of how to collaborate, talking about different methods for doing so safely if in terms of avoiding detection and how to develop child sexual abuse material.

And you see them growing in terms of their skill set as they're ch- trading strategies. And so you do certainly see perpetrators there very much pro artificial intelligence. There was one quote in IWF's research from one perpetrator who said, quote, "If the future will involve us being dominated by artificial intelligence, I say bring it on if it can create stuff like this."

And so very much embracing that technology. So again, the, in these forums, the AI-generated child sexual abuse material is coexisting alongside child sexual abuse material. It's the same market Another problem to address is this quote, "Does not depict actual child language from Ashcroft." And Justice Kennedy very much leaned into that [00:49:00] in sh-- in stating that the First Amendment, you know, protected so-called, quote, unquote, "virtual child pornography."

And so in the Ashcroft opinion, Justice Kennedy wrote, quote, "These images do not involve, let alone harm, any children in the production process, but Congress decided the materials threaten children in other less direct ways." So that's another piece of Ashcroft that just no longer works in the modern technological context.

Anderegg, in the opinion itself acknowledges that there is evidence of child sexual abuse material and at least some significant training data sets. And we talked with Debbie about the burden of proof issue but that certainly is not the case anymore. So under these an-another note, another quote from Ashcroft that's relevant is where Justice Kennedy states, "Under these rationales," referring to the government argument in Ashcroft, "harm flows from the content of the images, not from the means of their production."

Fast-forward [00:50:00] to twenty twenty-six, there are harms to be had from the means of production itself in a variety of ways. I would also quibble with this quote from Anderegg, where it, it, it tries to, it, it, it tries to make two things equivalent. In, in Anderegg it stated that the argument the government advances today, meaning here in, in twenty twenty-six or twenty twenty-five, whenever the oral arguments were or the briefings were, that it is increasingly difficult to distinguish between virtual CSAM and CSAM produced with actual children.

The Anderegg court said that argument is still, quote, "identical to the one the Supreme Court rejected in Ashcroft." I don't think that's accurate that the arguments are identical. Again, I wasn't at the oral argument but it certainly is a very different context. Modern AI image generation takes a ton of forms.

It could be completely synthetic output from standard training [00:51:00] data, potentially. Again, I am not the, the resident technologist, though we have those in our network. The meaning, the training data, maybe it has no CSAM whatsoever. Again, I, I, I think there's a variety of things to be explored here and, and to not make assumptions, but theoretically, that's a possibility that you can make completely AI-generated out-- completely AI-generated output from training data that has no CSAM.

And then you have models that could be trained on ordinary images of children, actual identifiable children, as part of this model, and then that model can then generate child sexual abuse material. And the tech that enables it to do that is pulling from, as reference images with or within the training data set, actual images of children.

And so the children themselves, even in this, this quote unquote clean training data where there is no CSAM, the images themselves that the models are trained on are images of actual children. And if [00:52:00] that's the case, it's simply not true, the assumption that Kennedy had in Ashcroft that no actual children are involved.

You might not know who, you might, might not know which ones, but those children, those actual children are involved in the creation of CSAM. Another permutation of this might be models, AI models, where they just scrape the web, huge amounts of collection of information from the Internet, and within those data sets, there is child sexual abuse material.

And so the these models are training on child sexual abuse material And in fact, this is a thing that actually happened. There is a paper called Identifying and Eliminating CSAM in Generative Machine Learning Training Data and Models. The author was David Thiel. It was published by the Stanford Digital Repository, and that study found that having a particular dataset, I'm gonna mispronounce it, the L-A-I-O-N, LAION-5B, maybe that's how you say it, dataset populated even in late twenty twenty-three, [00:53:00] so it's been a minute in AI, in the AI timeline, it's been a minute.

The study found that even having possession of that data, dataset in late twenty twenty-three, quote, "Implies the possession of thousands of illegal images," referring to child sexual abuse material, not including all the intimate imagery published and gathered non-consensually. So within that training dataset, which it should be noted, this wasn't some custom dataset to generate CSAM.

This was a major significant dataset for generating AI material, presumably not intended to generate child sexual abuse material. And even there, just from the broader scraping of the web or however that dataset was created, child sexual abuse material was within the training information that these models were trained on.

It should also be noted that that study also stated that the material detected, referring to the CSAM, the material detected during this process, this study, is, quote, "Inherently a significant undercount." [00:54:00] Undercount. So certainly that happens, models with datasets that unintentionally contain child sexual abuse material.

Then you have AI models that very intentionally are trained on specific child sexual abuse material. They're designed to generate images and videos either of specific children or specific categories of children or just children more broadly. So that's a whole other category that is different today than it was when Ashcroft was decided Then yet again, you have AI manipulation of existing child sexual abuse material.

So a whole separate form of modern AI image generation. AI manipulation of existing child sexual abuse material. So you could have, quote-unquote, "real CSAM" that's manipulated in some way and made to look AI, for, for example, or altered in some way, and so our tools flag, oh, that's synthetic. And so there's not a actual victim there.

Well, no, actually there was. It just AI was manipulated the other way. And if we're gonna create this hellscape [00:55:00] where we can't... where virtual child sexual abuse material, you know, has the, has First Amendment protections from a, a, a possession perspective, again, not to overstate Anderegg's impact, then these are the type of nightmares that we're forcing investigators and forensic examiners and prosecutors to deal with, and it's fundamentally problematic in a variety of ways.

So we had wholly synthetic output from clean training data. We had models trained on ordinary images of children. We had models who had massive data sets, oftentimes scraped from the internet, that unintentionally contained child sexual abuse material. We had fine-tuned models that are intentionally trained on, on CSAM.

Then we have AI manipulation of existing CSAM, and then we also have deepfakes of identifiable real children, and tho- those are prolific. So the idea that those things can be collapsed, that you can call those categories, quote, "virtual child sexual abuse material," as Anderegg does where it says, oh, you, you're making this argument that it's increasingly difficult to [00:56:00] distinguish between virtual CSAM and CSAM, that's the same thing the Supreme Court rejected in Ashcroft.

No, Anderegg. The fact that you collapsed those into virtual CSAM is a huge category error. And then treating virtual as synonymous with victimless in production, that's applying a two thousand and two technology, a two thousand and two understanding onto a twenty twenty-six reality, and I think that's fundamentally flawed.

The concurrence understands that a bit and is a bit more nuanced than the main opinion, but the main opinion, I think really is problematic if it's trying to e- if it's trying to equate those things, a- and it is facially, I think that's a problem So ultimately, we come down, I think, to a moral issue. Is this the world we wanna live in?

Certainly, if we care about liberty, if we care about freedom, which we should we should hesitate before criminalizing what someone private-- privately views. But liberty doesn't require naivete. It doesn't require [00:57:00] technological ignorance. It doesn't require us to pretend that a photorealistic simulation of child sexual abuse material exists in a vacuum when real victim imagery feeds AI, AI models, real children are turned into deepfakes, synthetic imagery is used to groom and extort them, and that actually happened in Anderegg itself.

Recall, he sent this imagery to a victim, and investigators have to divert their time from identifying children being abused because they can't reliably tell what is what, what is real. And this broader legal framework dating back to Ashcroft creates some major, major issues when it comes to child safety.

So communicate with your elected representatives. Support your law enforcement partners. This is an issue that needs to be pushed. Ashcroft needs to be reversed, full stop. We've talked quite a bit. Let me turn to some resources. Obviously, at Coursing Justice, we have [00:58:00] training on artificial intelligence ranging from shorter hour-long virtual webinars to full-day trainings on artificial intelligence.

I learned long ago that if you ever wanna be taken seriously in a law enforcement audience, you do have to co-present with law enforcement officers. So we've got a big roster of those. Though certainly some of that draws from my days in prosecution. And then also we mentioned Hexordia which you can access at hexordia.com.

Hopefully, one of your takeaways listening to this, even if you didn't understand any of it, admittedly a bit hard to follow is we collectively need to upskill in this, in this new era of technology. So Hexordia is a great, great place to do that. I've loved working with them. As I mentioned, both Debbie and I are instructors at Hexordia and partner with Jessica Hyde and others.

And so it's a great group very dedicated to the mission so encourage you to check that out[00:59:00]

I'm going to close what has been for most probably a dense and possibly disturbing or hopefully disturbing discussion or epis- or podcast episode with a quote that has nothing to do with artificial intelligence, this is a quote that I was reminded of earlier in this week, and it comes from Lord of the Rings, no artificial intelligence in sight, fear not, from J.R.R. Tolkien in The Return of the King. The quote is, "It is not our part to master all the tides of the world, but to do what is in us for the succor of those years wherein we are set, uprooting the evil in the fields that we know, so that those who live after may have clean earth to till.

What weather they shall have is not ours to rule." Let's focus on what we can, do all we can to uproot evil in the fields that we know. There's no shortage of that, certainly. Thank you for your commitment to that work. And thank you for listening to Coursing Justice. Sign up for our newsletter at coursingjustice.com and join us as we continue charting a [01:00:00] way forward for justice in the cases that matter most.

Coursing Justice
Next

#2: The Victim Advocate (with Hannah Fordice)