On 2 November 2026, Thorn published “What Is AI-Generated CSAM?” on its Safer website, examining the growing challenges of AI-generated child sexual abuse material (AIG-CSAM). The article raises legitimate concerns about how generative AI is changing the investigative landscape, particularly the increasing volume of synthetic material confronting victim identification teams.
However, one statement deserves closer examination:
“Because AIG-CSAM could depict a real child suffering sexual abuse, it’s not enough for investigators to simply detect if content is AI-generated and move on.”
While the premise has validity, it overlooks an important distinction that has significant consequences for investigative prioritisation.
The possibility that a child may have been abused is not equivalent to evidence that a child has been abused.
This distinction matters because investigators operate with finite resources. Every hour spent examining one piece of evidence is an hour unavailable for another. Decisions about investigative priority must therefore consider not only the possibility of harm, but the strength of the evidence indicating that an actual child may be at risk.
The emergence of generative AI makes this distinction more important than ever.
Generative AI has changed the evidential significance of novelty
Historically, newly discovered CSAM has carried considerable investigative significance. A previously unseen photograph may represent an unidentified victim, an unknown offender or an abusive event that law enforcement has not previously encountered.
Generative AI fundamentally changes that relationship.
An individual can now produce thousands of entirely synthetic depictions without photographing a single abusive event. Every image may be technically novel, yet none necessarily documents actual victimisation.
This creates a problem for investigative systems that have traditionally treated novelty as an indicator of potential new victims.
Consider two previously unseen images depicting apparent child sexual abuse. One exhibits evidence consistent with physical camera capture. The other exhibits evidence consistent with synthetic generation.
Both may warrant investigation, but they do not carry equivalent evidential implications.
The first provides a more direct basis for investigating whether an actual abusive event was recorded. The second requires additional evidence to establish whether a real child was involved.
That difference should influence investigative priority.
Ignoring it risks allowing the sheer volume of synthetic material to consume resources that could otherwise be directed towards identifying children experiencing actual abuse.
Voluntary compliance methods are overlooking the forensic evidence already in the image
Much of the current discussion surrounding AI-generated imagery concentrates on detecting the characteristics of generative AI. We develop classifiers, examine metadata, implement provenance standards and embed watermarks.
These technologies serve legitimate purposes, but they overlook another forensic question. Rather than simply asking whether an image looks AI-generated, why aren’t we examining the evidence that a physical camera captured it?
Digital cameras leave measurable characteristics in photographs arising from their physical image sensors and the process of capturing light. Generative AI produces imagery through a fundamentally different computational process.
That distinction provides an opportunity to examine image origin using forensic evidence rather than visual appearance or embedded declarations.
At RHEM Labs, we have been developing SPOT (Sensor Pattern Origin Testing), a forensic capability arising from my doctoral research into sensor pattern noise.
SPOT examines characteristics of image formation to evaluate the evidence supporting physical camera capture or synthetic generation. Rather than simply assigning a binary classification, the approach evaluates the relative strength of evidence supporting competing explanations.
Our published 2026 research reported 98.5% classification accuracy when distinguishing camera-origin photographs from DALL·E-generated images under controlled experimental conditions.
The significance extends beyond detecting AI-generated content. It demonstrates how information about an image’s origin can become useful forensic intelligence.
This is particularly relevant to CSAM investigations, where distinguishing evidence consistent with a photographed event from an entirely synthetic depiction can directly inform investigative decisions.
The objective is to rescue victims, not investigate images equally
Thorn correctly identifies that the growing volume of synthetic material threatens to overwhelm existing victim identification systems. But this raises a question that receives insufficient attention. What happens when synthetic material begins to dominate investigative queues?
If every synthetic depiction receives the same priority as material presenting stronger evidence of an actual photographed event, investigators risk spending increasing amounts of time examining images that may have no corresponding victim.
This does not mean synthetic CSAM should be ignored. An image incorporating a real child’s likeness, genuine abuse imagery or evidence of ongoing exploitation may warrant immediate attention regardless of how it was generated.
The point is that these decisions should be informed by evidence.
Forensic origin assessment using sensor pattern origin testing provides another piece of that evidence. Combined with victim identification indicators and contextual intelligence, it can help investigators distinguish between material requiring immediate intervention and material warranting further examination at a different priority.
SPOT was designed to address precisely this missing component. It offers a way to evaluate image-origin evidence before scarce investigative resources are committed and does so assuming no novel hardware or computational resources are needed beyond the abandoned workstation sitting under an officer’s desk.
Our current focus is working with operational partners to evaluate how this capability can be integrated into existing investigative workflows. The opportunity is not simply to detect more content, but to make better use of the evidence already available in the scenarios that matter most.
The distinction matters
Thorn is correct that identifying synthetic content should not automatically terminate an investigation. However, that observation should not obscure the equally important consequence of failing to distinguish synthetic material from evidence of actual photographed abuse.
Generative AI has fundamentally changed the relationship between an image, its apparent content and the likelihood that it documents an actual event.
Investigative practices must evolve accordingly.
Forensic technologies such as SPOT demonstrate that we can examine more than what an image depicts. We can investigate how it came into existence and use that evidence to inform the decisions that follow.
The critical distinction is between the possibility of victimisation and the evidential support for concluding that victimisation occurred.
Treating those propositions as equivalent risks overwhelming investigators with synthetic material while children experiencing actual abuse remain unidentified.
The objective has never been to investigate the greatest number of images.
It is to identify and protect the greatest number of children.
Achieving that requires us to prioritise evidence, not merely detect content.