The bigger crisis is no longer volume alone. Investigators now struggle to prove what is real.
Generative AI changed how child abuse material gets made. It also broke a basic assumption in child protection work. Police once asked who made an image. Now they must ask whether a child was involved at all.
The latest data shows how fast the problem is moving. It also shows why counting images only tells half the story.
The numbers keep rising
The Internet Watch Foundation (IWF) monitors abuse material worldwide. Its analysts reviewed 6,310 AI-generated images in the first half of 2026. That already beats the 2025 total of just over 4,500 by 40%.
Those figures cover still images only. Video sits outside the count, and video has grown sharply over the past few years.
The IWF’s earlier review of 2025 found 8,029 AI-generated images and videos. Of those, 3,443 were videos. Some 65% of the videos fell into Category A, the most severe classification.
The trend reaches beyond the UK. Save the Children in Denmark logged 4,069 reports of suspected AI-generated abuse material in the first six months of 2026. The agency believes AI plays a growing part in that rise.
Two lines on a chart do not explain what changed. The structure of the problem explains it.
Three gaps keep the response behind
Most coverage treats this as a content moderation problem. That framing misses how the system fails. Three gaps explain why.
The supply gap
Abuse material once required access to a victim. That requirement acted as a natural brake on volume. Generative tools weakened it.
The IWF’s June 2026 report describes rising scale, realism and accessibility. It also warns that offenders can train models on existing abuse material. Every reuse revictimises the original child.
Supply no longer tracks new harm in the moment. It tracks hardware, software, and time.
The proof gap
Police now have to separate real material from synthetic material. They often do it at scale and under pressure.
Offenders have noticed. The IWF reports that some suspects now use a “deepfake defence.” They claim genuine evidence of abuse is artificial.
This gap hurts in two ways. A convincing fake wastes investigator time. A real image mislabelled as fake could let an abuser walk free.
Triage gets harder too. Teams that once ranked cases by severity must now spend effort on a prior question. They must first establish whether a victim exists.
The law gap
Lawmakers are moving. They move at different speeds, and every law stops at a border.
| Region | Recent action |
|---|---|
| UK | New offences under the Crime and Policing Act 2026 cover possession of AI models that can create abuse material. |
| EU | A committee vote backed tougher criminal rules, including penalties for AI systems built for this purpose. |
| Australia | MP Kate Chaney introduced a bill targeting tools designed to generate abuse material and the data scraped to train them. |
| Denmark | A 2025 amendment already covers realistic AI-generated material. |
Each approach targets a different link in the chain. Some laws hit the image. Others hit the model, or the training data behind it. Together they feed a wider AI policy fight over who answers for misuse. That fight has no settled answer yet.
Platforms face courts, not just filters
In July, xAI sued a Grok user. The company alleges he used the chatbot to create sexualised images of adults and children.
The case matters for what it shows. Safeguards on a major platform can fail. When they do, companies now reach for legal action against the user, not only for technical fixes.
Developers will face a harder question next. Is a lawsuit against one user enough when a model’s design makes the misuse possible?
“No real child” is the wrong defence
Some people argue that fully synthetic images harm no one. The IWF disagrees on two grounds.
First, training on real abuse material keeps harming real victims. Second, realistic content may reinforce sexual interest in children. That can contribute to escalation in offending.
The UK government takes a plain line. A spokesperson said UK law treats this material as illegal whether or not AI generated it.
Courts in other countries appear to agree. Denmark and the EU have both moved to cover synthetic material explicitly.
What would actually help
No single fix closes all three gaps. A realistic response needs several moves at once.
- Safeguards at the model level. Developers can test their systems against abuse scenarios before release. They can also block misuse at the source, not only after images spread.
- Faster evidence tools. Police need reliable ways to judge whether an image shows a real child. Funding for that work lags behind the problem.
- Cross-border cooperation. Hotlines, regulators and police forces need shared standards. Offenders ignore borders. Enforcement cannot.
- Clear liability rules. Lawmakers must decide where responsibility sits. It could sit with the toolmaker, the host, or the user.
Three questions for the next year
- Model liability. Will lawmakers punish only models built for abuse, or also general models that offenders adapt?
- Provenance. Can authenticity tools keep pace with realism, and can police afford to run them on every case?
- Offline models. National bans mean little if offenders run models on their own machines, far from any regulator.
FAQs
Q. Is AI-generated child abuse material illegal?
In the UK, yes. The government has also banned adapting an AI model to create such material and passing those models to others. Denmark, the EU and Australia are moving the same way.
Q. Why does the IWF’s data matter?
Trained analysts assess reports directly. The counts reflect confirmed material, not platform estimates.
Q. What is the “deepfake defence”?
It is a courtroom tactic. A suspect claims real abuse footage is AI-generated. The IWF says police have begun to meet it.
Q. Where should people report suspected material?
Use the IWF or your national hotline, or contact local police. Never download or share it.
Related: AI Risks in 2026: Deepfakes, Jagged Frontiers & the Collapse of Shared Reality
