Every year the Internet Watch Foundation publishes data on what its analysts encounter while hunting illegal imagery online. The 2026 report is different in scale and in character. It is not describing a trend that is building. It is describing a technological inflection, a point at which the barrier to creating child sexual abuse material has effectively collapsed. The question it raises is urgent: what should the people building AI systems, the governments regulating them, and the platforms that deploy them be doing differently?
What the IWF found in 2025
The Internet Watch Foundation is Europe's largest hotline dedicated to finding, reporting, and removing child sexual abuse imagery online. In 2025, its analysts assessed 8,029 AI-generated images and videos as depicting realistic child sexual abuse, a 14% overall increase on the previous year.
The video figure deserves particular attention. A 260-fold increase in a single year, from 13 videos to 3,443, is not a gradual escalation. It is the consequence of a specific technological development: realistic full-motion AI video generation becoming widely accessible in 2024. What required significant skill and resources in 2023 became a low-barrier activity in 2025.
"Advances in technology should never come at the expense of a child's safety and wellbeing. While AI can offer much in a positive sense, it is horrifying to consider that its power can be used to devastate a child's life. This material is dangerous."
. Kerry Smith, CEO, Internet Watch Foundation
The severity picture is more alarming than the volume
The raw increase in numbers is significant. The severity distribution is more disturbing. Under UK law, Category A is the most extreme classification for abuse material, encompassing the most serious forms of sexual abuse. In 2025, AI-generated content was more likely to be Category A than non-AI content.
- 65% of AI-generated videos were Category A
- 43% of non-AI videos were Category A (for comparison)
The IWF's interpretation of this disparity is that generative AI is not simply reproducing the distribution of harm that already exists. It is being specifically directed toward the most extreme content, in part because the tools allow users to specify in detail what they want generated, in part because the dark web communities that use these tools actively share techniques for creating more severe material.
An additional 82 items were classed as prohibited under UK law even though they were not photorealistic, including cartoons, illustrations, and animations. The law covers non-photorealistic AI-generated content. Severity and legal exposure are not limited to hyper-realistic imagery.
Three specific developments the IWF identified
1. LoRA fine-tuning, creating material targeting specific children
Low-Rank Adaptation (LoRA) is a fine-tuning technique that allows generative models to be customised toward specific subjects using as few as 20 existing images, in as little as 15 minutes. The IWF identified this being used to create material that appears to target real, identifiable children. The technology directly re-victimises survivors by using existing imagery of them to generate new abuse material.
2. AI-generated audio, synthetic child voices
Audio deepfakes, synthetically generated voices of children used in abusive scenarios, are an emerging category. The IWF identified a fully synthetic video involving both AI-generated imagery and AI-generated audio of a child's voice. As multimodal AI tools converge, the ability to generate coherent multi-sensory abuse material without any real child being involved is increasing.
3. The "liars' dividend", using AI as legal cover
The IWF identified a disturbing legal tactic: offenders claiming that genuine photographic evidence of real contact abuse was in fact AI-generated, and therefore does not depict a real child. The widespread public awareness of synthetic media creates a plausible deniability argument that is beginning to be tested in legal proceedings. AI-generated content is not a lesser harm, it is also being used to obscure greater ones.
What 82% of UK adults are calling for
The IWF published its report alongside Savanta polling of 2,204 UK adults, conducted in March 2026 and weighted to be representative by age, gender, region, and social grade. The results are unambiguous about public appetite for regulation.
- 82% say the government should introduce regulation to ensure AI systems are safe by design and future-proofed from causing harm
- 78% agree that AI companies should be required to test for AI-related harms before products are released to market
The IWF's specific policy asks align with this public position. They are calling for legislation requiring AI platforms to implement safety-by-design as standard, including pre-release testing to ensure models cannot be used to generate child sexual abuse imagery, content moderation policies, and use of trusted datasets to prevent CSAM from being included in AI training data. The IWF's own Hash List contains over 3 million verified CSAM hashes and is available to platforms as a training data filter.
The UK government has already moved in this direction. The Crime and Policing Bill introduced a new criminal offence for making, adapting, possessing, or supplying a "CSA image-generator." It also includes provisions enabling designated authorities, including the IWF, to test AI models for their ability to generate child sexual abuse imagery. The IWF has stated it stands ready to support this work.
What this means for AI developers and platform operators
The IWF report is careful to note that AI-generated content remains a comparatively small proportion of the total CSAM it processes each year. What has changed is the rate of increase, the severity distribution, and the direction of travel. A 260-fold increase in videos in a single year is not a steady-state problem. It is a problem that will compound every year that AI tools become more capable and more accessible.
For developers building AI systems, and for platform operators deploying them, the practical obligations this creates are now concrete. Pre-release testing for the ability to generate CSAM is being written into law. Training data hygiene, ensuring that abuse material has not been used to train generative models, is an expectation, not a best practice. Detection infrastructure that can classify synthetic child sexual abuse material before it is distributed is a necessity, not a product feature.
The IWF's framing matters here. The problem is not that bad actors exist. The problem is that safety safeguards were not consistently embedded into AI systems from the outset, and the cost of that omission is now being counted in 8,029 items assessed in a single year, with a trajectory pointing sharply upward.
---
Tuteliq. Synthetic content detection built for exactly this.
Tuteliq's synthetic content detection endpoints are designed to identify AI-generated child sexual abuse material across text, image, audio, and video, with classification into confirmed synthetic, suspected synthetic, or unknown, and severity scoring that surfaces the most urgent cases for human review.
Relevant detection categories: SYNTHETIC_CSAM · AI_GENERATED_IMAGE · AI_GENERATED_VIDEO · AI_GENERATED_AUDIO · AI_MANIPULATED_MEDIA · SYNTHETIC_IDENTITY
Every detection returns a classification field aligned with the IWF's own recommended triage categories, confirmed, suspected, or unknown, alongside an audit_id for your compliance documentation trail. Read the documentation or speak to the team.
---
If you need support or want to report:
- IWF Report Remove. Report child sexual abuse imagery online, iwf.org.uk
- CEOP. Report online child sexual exploitation to police, ceop.police.uk/safety-centre
- Stop It Now. Helpline for those concerned about their own or others' behaviour, 0808 1000 900
- NCMEC CyberTipline, US-based reporting for CSAM and child exploitation, missingkids.org
---
Sources: IWF. Harm without limits: AI CSAM report 2026, published 24 March 2026 · Savanta polling of 2,204 UK adults, March 2026, weighted by age, gender, region, and social grade · Fortune, April 2026 · IWF policy blog, March 2026. Kerry Smith quote drawn directly from the IWF press release.