Chat Control is back: EU Council pushes client-side scanning on a fast-track

What happened

The EU Council, under the Danish presidency, has put the so-called Child Sexual Abuse Regulation (CSAR) — better known to everyone outside Brussels as Chat Control — back on the agenda and is trying to push a general approach vote before the end of 2026. Heise reports the file has been moved onto a fast-track procedure that skips the deeper working-group review Chat Control failed to survive the last two times it was proposed.

The mechanism hasn't softened: providers of "interpersonal communications services" would be legally required to scan message content on-device, before end-to-end encryption is applied, against a database of known and unknown CSAM hashes maintained by an EU Centre. If a message trips the classifier, metadata plus the payload is forwarded to authorities. Voluntary detection — the current stopgap regime that lets Meta and Google scan unencrypted DMs — expires in April, which is the artificial cliff the presidency is using to justify the sprint.

The political shape has shifted since the last attempt collapsed in 2024. Germany, previously the swing vote that killed the file, has stopped signalling a hard no under the new coalition. France is in favour. Ireland, which hosts most of the affected companies, is publicly noncommittal but privately closer to yes than last round. Meredith Whittaker, president of Signal, has repeated her 2023 line: Signal will leave the EU market rather than ship a build with client-side scanning. WhatsApp's leadership has been quieter but has said the same thing internally, per reporting from Politico Europe last month.

Why it matters

The technical objection to Chat Control hasn't changed since Matthew Green, Bruce Schneier, and 300+ other cryptographers signed the 2023 open letter against it: you cannot bolt a content classifier onto a messaging client without breaking the security property the client existed to provide. End-to-end encryption is not a marketing label. It's a claim that only the endpoints see plaintext. Client-side scanning moves the adversary inside the endpoint. Once the scanning hook exists, the interesting question stops being *whether* it will be abused and becomes *by whom, first*.

The EU's own legal service told the Council in 2023 that the proposal likely violates Articles 7 and 8 of the Charter of Fundamental Rights — the private-life and data-protection clauses. That opinion has not been withdrawn. The fast-track procedure is, in part, a way to get to a vote before that opinion can be re-litigated in public. The European Data Protection Supervisor, Wojciech Wiewiórowski, called an earlier version "the point of no return for democratic surveillance in the free world." He has not walked that back either.

The classifier side of the proposal is where the engineering gets ugly. Known-CSAM hashing (PhotoDNA-style perceptual hashes) has a well-characterised false-positive rate — Apple's abandoned 2021 rollout leaked internal numbers around 1 in 1 trillion per account per year, which sounds fine until you multiply by 450 million EU users and daily message volume. Unknown-CSAM detection is worse: it requires an on-device ML classifier trained on material the developers legally cannot possess, evaluated against content the reviewers legally cannot look at. Every serious attempt to build this — Apple's NeuralHash, Google's Content Safety API, Thorn's Safer — has shipped with a manual human-review step in the loop. Scale that to every message on every phone in Europe and the review queue is the story.

The grooming-detection component, added in a 2024 compromise text and still present in the current draft, would use an LLM-style classifier to flag "suspicious conversational patterns" between adults and minors. There is no published benchmark for this class of model that survives adversarial review. The false-positive surface is every teenager talking to a coach, tutor, or older cousin.

What this means for your stack

If you ship an encrypted messaging feature into the EU — and "messaging feature" here includes Slack-alike team chat, in-game DMs, dating-app inboxes, customer-support widgets with file upload, and anything using Matrix, XMPP, or MLS — start scoping the compliance ask now. The draft's definition of "interpersonal communications service" is broad enough to catch products that don't think of themselves as messengers, including federated protocols where there is no single provider to serve the order on. Matrix.org has published a position paper arguing the regulation is technically incompatible with federation; nobody in the Council has engaged with the argument.

For WebRTC and SFU-based products, the current text exempts real-time audio and video but not the signalling channel or any chat sidebar. If your app has a text field next to the video call, that field is in scope. For open-source projects, the obligation attaches to whoever "provides the service in the Union," which likely means anyone running a public homeserver, relay, or hosted instance — not the upstream maintainers, but close enough that maintainers will get the compliance requests anyway.

The practical near-term move for teams shipping in the EU: audit which of your products fall under the interpersonal-communications definition, model the two failure modes (comply and lose Signal-tier users, or refuse and lose EU market access), and start the conversation with legal now rather than in Q2 when the trilogue moves. If you sell to enterprises with a "we use E2EE" claim in the contract, that claim becomes a live liability the moment a general approach vote passes.

Looking ahead

The file has failed twice. It could fail a third time — the German position is genuinely unresolved, and the Parliament, which has to co-decide, killed the last version 537-133. But the fast-track exists precisely to change the outcome, and the presidency wouldn't be spending its political capital if the whip count didn't look better than it did in 2024. If you're building on the assumption that end-to-end encryption in the EU is a settled architectural choice, the next six months are when that assumption gets tested.