Skip to content
Published on

EU Chat Control 1.0: what actually passed, and the client-side scanning question

🇰🇷KO🇺🇸EN🇯🇵JA
Split View필사 모드
Share
Authors

Introduction — a vote that says two contradictory things

On 9 July 2026 the European Parliament let a message-scanning regulation pass even though more of the MEPs in the room voted against it than for it. That sentence sounds broken, and the reaction online has been loud. But "Chat Control" is one of those topics where the headline and the mechanism disagree, and where the thing people fear (breaking end-to-end encryption) and the thing that actually passed (a voluntary derogation) are not the same thing.

This post is an attempt to separate the two, carefully, for people who build or depend on messaging and privacy tech. What the regulation mandates, what critics fear, and what supporters argue are three different registers, and I will keep them apart throughout.

What "Chat Control" actually refers to

"Chat Control" is not the name of a law. It is the label that critics — most prominently the former MEP Patrick Breyer and the digital-rights network EDRi — attached to a family of EU proposals to scan private communications for child sexual abuse material (CSAM). Two very different instruments hide under the one nickname:

  • Chat Control 1.0 is Regulation (EU) 2021/1232, a temporary derogation from the ePrivacy Directive. It permits providers of webmail and messaging services to voluntarily scan for CSAM; it compels no one, and by its own terms it does not reach end-to-end encrypted content. This is the thing that passed on 9 July.
  • Chat Control 2.0 is the proposed Child Sexual Abuse Regulation (CSAR), COM(2022) 209 final. It would introduce mandatory detection orders and a new EU Centre. It is still stuck in trilogue negotiations and has not been adopted.

Almost every misleading headline about this topic comes from collapsing 1.0 and 2.0 into one story. The mandatory, encryption-threatening version is 2.0, and 2.0 is not law.

What the 9 July vote actually did

The procedure is the story, so here is the sequence:

  • In late March 2026 Parliament narrowly rejected extending the 2021 derogation, and the interim rules lapsed in early April.
  • The Council then revived the file, adopting its position on 2 July and sending it back for a second reading.
  • Under the ordinary legislative procedure (Article 294 TFEU), at second reading Parliament can only reject the Council's text by an absolute majority of its component members — 361 of 720 seats.
  • On 9 July the motion to reject drew 314 votes in favour of rejection, 276 against, 17 abstentions. A plurality wanted it gone, but 314 is 47 short of 361. The motion failed, so the regulation is adopted and the voluntary-scanning derogation runs until 3 April 2028.

So the widely repeated line "the majority voted no and it passed anyway" is procedurally accurate: at second reading, silence and absence count as assent. Breyer and other critics call this a back-door manoeuvre by the largest group (the EPP); supporters call it the normal operation of the rules. Both descriptions fit the same facts.

Parliament also adopted amendments, including an explicit carve-out for end-to-end encrypted services. Breyer calls this exemption "symbolic," on the grounds that providers were not scanning encrypted content anyway — which is technically correct, and is the whole crux of the matter.

The technical tension: client-side scanning vs end-to-end encryption

Detection under these schemes uses three broad techniques, and they are not equally reliable:

  1. Perceptual hash matching against databases of known CSAM (PhotoDNA, and Apple's ill-fated NeuralHash). Mature and relatively accurate, but hashes are attackable: researchers reverse-engineered NeuralHash and produced deliberate collisions.
  2. AI classifiers for unknown imagery. The European Parliament's own research service concluded that no current technology reliably detects novel CSAM without unacceptable error rates.
  3. Text/NLP grooming detection, the least reliable of the three.

The reason this matters even when accuracy sounds high is the base-rate problem. A European Parliament study reported that in one Irish data set only about 20% of the CSAM reports forwarded to police were confirmed as actually illegal. Run any classifier across billions of daily messages and even a tiny false-positive rate produces an enormous absolute number of innocent people flagged, and their private messages exposed to human review.

Now the encryption piece. End-to-end encryption means only the sender and receiver hold the keys; the server literally cannot read the plaintext, so server-side scanning is impossible by construction. The only way to scan encrypted content is to inspect it on the device before it is encrypted — client-side scanning.

Server-side scan (non-E2EE):      plaintext ── scan ──▶ store/forward
Client-side scan (to reach E2EE): plaintext ── scan ──▶ encrypt ──▶ send

That second line is the entire controversy. Fourteen senior cryptographers — Abelson, Anderson, Rivest, Schneier, Troncoso and others — argued in "Bugs in Our Pockets" (2021) that a scanning hook on the endpoint is a general-purpose surveillance capability: once it exists it can be repurposed to new targets, and it becomes an attack surface in its own right. Apple shipped this argument the hard way, announcing on-device CSAM scanning in August 2021 and abandoning it by December 2022 after the backlash and the collision attacks. In February 2024 the European Court of Human Rights ruled that weakening encryption "cannot be regarded as necessary in a democratic society."

Closing — what this means if you build messaging or privacy tech

The honest summary is undramatic. What passed on 9 July is a voluntary legal basis for providers of non-encrypted services to keep scanning, extended to 2028. It compels no one, and it does not touch end-to-end encryption — the amendment even says so out loud. If you run an E2EE app, nothing in 1.0 forces a client-side scanner into your build.

The fight that actually threatens that architecture is 2.0, the CSAR, with its mandatory detection orders — and that one is still unresolved in trilogue, precisely because "detection order plus end-to-end encryption" has no technical solution that preserves the encryption guarantee. That is the file to watch.

Two things are worth holding at once. Supporters are not arguing in bad faith: providers' voluntary reports do feed real law-enforcement referrals, and letting the legal basis lapse removes a tool investigators use. Critics are also right that suspicionless scanning of everyone is, in EDRi's words, "textbook mass surveillance," and that client-side scanning cannot be bolted onto end-to-end encryption without redefining what that phrase means. For once, both sides are describing the same wall from different sides. If you build in this space, the design line is unchanged: content inspection on the endpoint and end-to-end encryption are mutually exclusive, and no vote changes that math.

References

Discuss on Twitter • View on GitHub