Deepfake CSAM is now extremely realistic

Child protection organizations already have a hard time eradicating child pornography. Now, AI is lowering the barrier for entry, letting anyone create explicit videos and images of children.

“Welcome to the new era of CSAM, where anything you desire is possible in extreme realism, and no real children are involved.”

This comment is lifted from a dark web forum that shares child pornography, where child predators are praised for creating realistic child sexual abuse material (CSAM) using AI.

While child welfare analysts have been monitoring AI CSAM for a few years now, technological advances in AI have caused synthetic child pornography to explode on both the dark and the clearnet.

The internet is riddled with AI child porn

Child protection services, namely the Internet Watch Foundation (IWF), assessed over 8,000 AI-generated images and videos depicting realistic child sexual abuse.

While this content represents a small proportion of these reports, IWF believes this is cause for concern as the volume of AI-CSAM continues to increase year over year.

This trend isn’t exclusive to certain areas of the world, as organizations like UNICEF and Interpol have recorded at least 1.2 million children who have had their images transformed into sexually explicit deepfakes.

In certain countries, this is 1 in 12 children, that’s roughly one child per classroom.

AI-CSAM was first reported in 2024 and was seemingly unsophisticated, glitchy, and of poor quality.

An observer on the frontline noted that the first AI child pornography videos were deepfakes and weren’t sophisticated, as they were just a child’s face slapped onto an adult entertainer's body.

“But now they have really turned a corner…(and) the quality is alarmingly high, and the categories of offence depicted are becoming more extreme as the tools improve in their ability to generate video showing two or more people,” the observer said.

AI-CSAM is being distributed at a rapid rate

The foundation identified nearly 3,500 AI-generated CSAM videos, an increase of over 26,000% compared to the previous year.

In comparison, only 13 videos of this nature were previously recorded.

The majority of these videos (65%) were classified as Category A, which is the most severe category of child pornography under UK law.

This content includes themes and activities such as sexual torture, rape, and bestiality (having sexual relations with animals).

“A further 30% (1,027 videos) fell into Category B, the second most serious classification,” according to IWF.

AI-CSAM now includes audio

An emerging threat to minors is the proliferation of AI CSAM with an audio deepfake component.

“This may take the form of recordings of abuse, “audio deepfakes,” which synthetically generate the sexualised voices of children, or manipulated recordings used to simulate abusive scenarios.”

Newer AI models have this functionality, which allows bad actors to generate pictures and audio to generate a video with a synthetic voice.

While this type of CSAM is not proliferating rapidly (yet), the demand is there, according to IWF’s latest report.

IWF analysts observed a fully synthetic video depicting a child between the ages of three and six speaking directly to the camera while performing sexual acts with an adult man.

Almost impossible to distinguish AI-CSAM from live CSAM

Previously, it was pretty easy to distinguish AI-generated CSAM from other forms, as AI versions felt airbrushed and overly polished.

Now, the latest forms of AI CSAM have been generated to appear imperfect, giving the illusion of amateur, homemade footage.

“IWF Analysts are no longer identifying irregularities in the generated content, aside from minor blurring that is difficult for untrained eyes to detect.”

“WOW, AIs really made a leap forward with motion. It used to be obvious! 10/10 I just love knowing that the quality… as great as it is at this point, it’s only going to get better!” said one offender.

AI allows anyone to create child pornography

Offenders themselves have noted that AI has significantly lowered the barrier to entry for making CSAM.

The fact that this content is easy to create without needing to groom or exploit real children, as it was before, could actually encourage offenders to seek out contact with children.

“This increased accessibility heightens the likelihood that AI CSAM will be created and that harmful offender behaviours will be reinforced.”

No longer do offenders need to use LoRAs (low-ranked adaptations) to create CSAM.

“LoRA freezes the original weights and parameters of the model as they are. Then, on top of the original model, it adds a lightweight addition called a low-rank matrix, which is then applied to new inputs to get results specific to the context,” writes IBM.

In the wild, LoRA was used to create deepfake pornography of young women, including Olympic figure skater Alysa Liu.

But now, sexual predators can prompt newer models to create AI-CSAM from just a single reference image without the need for additional training.

Furthermore, the use of AI agents is also making CSAM generation that much easier, as agents automate the process with limited human oversight.

“Thanks for your continuing hot, sexy videos. Don’t ever stop. In a while, you can tell your AI agent to pick out clips and create the video by itself,” one child predator said.

The problem doesn’t just lie in using this relatively intuitive technology, as users on the deepweb are teaching other predators to use these tools while trading datasets and fine-tuned models with the intent to generate CSAM on a wider scale.

“Woah, these models [AI models of children] are so sexy. I’d love to be able to use some of these. I could create just an endless amount of images.”