How have advances in tech (AI, hashing, platform detection) affected detection and conviction trends for CSAM?
Executive summary
Advances in hashing and AI have dramatically increased platforms’ ability to find and remove “known” CSAM at scale—industry tools processed billions of images and surfaced millions of matches in recent years (Safer: 112.3 billion processed; 4,162,256 potential files found) [1]. At the same time AI has produced new challenges: AI-generated CSAM reports surged in 2024–2025 and can outpace hash-based systems, while policy fights over client-side scanning and mandatory on-device detection show governments are unsettled about trade‑offs between privacy and detection (NCMEC/IWF counts and EU Council developments) [2] [3] [4].
1. Hashes: the scalable workhorse that only finds what’s known
Platforms and NGOs rely on perceptual and cryptographic hashing to detect re-uploads of previously identified CSAM; shared hash lists and industry hash‑sharing platforms have enabled rapid automated takedowns and reporting to bodies like NCMEC [5] [6]. Thorn, Google and others report large-scale throughput—Safer processed 112.3 billion images and videos in 2024 and found more than 4.1 million potential CSAM files for customers—while projects like the Video Hash Interoperability Project hashed hundreds of thousands of videos in 2025 [1] [7]. Hashing produces very low‑false positive, high‑recall matches for known content but cannot detect novel material that has never been catalogued [8] [9].
2. AI classifiers: finding novel content — powerful but imperfect
Organisations are deploying machine‑learning classifiers to surface previously unseen CSAM and prioritize investigative work. Thorn and others argue classifiers speed up victim identification and lift detection beyond hashes [10] [11]. Independent reviews find AI tools can achieve high true‑match rates for authentic CSAM in certain tests (between ~93.8% and 98.8% reported in a 2025 review), but classifiers carry limits around generalisability, evaluation quality, and privacy/legal constraints [12] [8]. Industry proposals therefore combine hashing (for known items) with predictive AI (for novel threats) to cover both gaps [13].
3. Videos and cross‑platform sharing: new technical frontiers
Video CSAM required bespoke solutions; Thorn’s Scene‑Sensitive Video Hashing (SSVH) and other video hashing efforts report high bench performance (>95% recall at 99.9% precision in target transformations) and have been integrated into cross‑platform workflows to reduce re‑circulation [14] [7]. Building bridges between hash systems—so different formats interoperate—has been a deliberate industry priority to broaden coverage and speed removal across services [7] [15].
4. AI-generated CSAM: a rising tide that can swamp existing systems
Multiple sources document a sharp rise in AI‑generated CSAM. The Internet Watch Foundation and US reporting show large volumes of AI‑produced images and rapid increases in reports (IWF and NCMEC figures cited by advocacy groups), with NCMEC reporting tens of thousands in 2024 and hundreds of thousands in early 2025—trends that strain responder capacity and complicate triage between real‑victim material and synthetic imagery [16] [2]. Law enforcement convictions for AI‑only creation have occurred (a DOJ arrest in 2024), but investigators report difficulty securing convictions in some AI‑linked cases and differentiating AI from authentic material for prosecution [17] [18].
5. Encryption, client‑side scanning and the policy tug‑of‑war
Policymakers have debated mandatory on‑device or client‑side scanning; recent EU positions have backed away from blanket scanning mandates but left open obligations to mitigate risks, creating ambiguity that could still press companies toward scanning iCloud or messages to comply [3] [19]. Scientists and privacy advocates warned that client‑side scanning undermines end‑to‑end encryption and raises false‑positive and circumvention concerns [4] [20]. This policy uncertainty affects what detection tools companies deploy and how law enforcement access to evidence evolves [3] [19].
6. Convictions and investigations: technology helps detection but legal outcomes lag in places
Technology has expanded detection and reporting volumes—industry reporting credits hash sharing and classifiers with huge increases in takedowns and referrals [1] [10]. Yet legal systems face new evidentiary challenges: AI‑generated material raises questions about victimhood, proof of contact or abuse, and whether possession of synthetic images meets existing statutes; some jurisdictions moved to criminalize AI‑generated CSAM expressly and prosecutions are beginning [21] [17]. Reporting shows low conviction rates in some regions for CSAM cases involving technology, underscoring that detection alone does not guarantee prosecution [18].
7. What reporting does not say — key gaps and limits
Available sources do not provide a global, comparable time series tying detection technology deployment directly to conviction rates across jurisdictions; they report detection throughput, hashes created, and spikes in AI‑CSAM reports, but not a consistent metric of prosecutions attributable to those technologies (not found in current reporting). Sources also vary by agenda—industry and NGOs emphasise effectiveness of tools (Safer, Thorn) while privacy advocates and academic scientists stress surveillance risks [1] [4]. Readers should weigh technical claims against governance and civil‑liberties critiques.
Conclusion: technology has reshaped detection—hashes and AI classifiers dramatically scale discovery and takedown of known CSAM and surface novel threats—yet AI‑generated imagery and policy disputes over scanning and encryption mean detection gains do not automatically translate into prosecutions or unambiguous legal remedies [1] [2] [3].