2026-05-10

EU AI Act simplification — synthetic-content labeling grace period halved to 3 months, deadline now December 2, 2026

EU AI Act simplification (May 7): synthetic-content labeling grace cut from 6 to 3 months. New deadline December 2, 2026. Ratification expected within weeks.

On May 7, 2026, the EU Council and Parliament reached a provisional agreement on a simplification package for the AI Act. The most builder-relevant change: the transparency-labeling grace period for AI-generated content was cut from 6 months to 3 months, with the new compliance deadline now December 2, 2026 — that’s roughly 7 months out, not 10.

What the new dates look like

Ratification is expected within weeks. The package goes to the formal trilogue vote, but the political agreement is firm.

What the labeling rule actually requires

Any AI system that generates image, audio, video, or text content for EU users must:

1. Embed machine-readable provenance metadata — typically C2PA Content Credentials, which is the de facto standard
2. Provide a visible/audible disclosure in the user-facing interface (“This image was generated by AI”) — except for “obvious art / parody” exemptions
3. Maintain audit logs — which generations were produced, when, by which model, available to regulators on request

Models that are themselves “general-purpose” (LLMs, image generators) must mark their output. Downstream applications must preserve those marks unless they apply their own.

Implementation cost reality

For a typical SaaS shipping image / text generation:

• C2PA metadata embedding: 1-2 week engineering effort (well-documented libraries, both Adobe’s CAI tools and open-source c2patool)
• Visible disclosure UI: 3-5 days of frontend work
• Audit log infrastructure: 1-2 weeks if you don’t already have a usage-events pipeline
• Total: 4-6 weeks elapsed for a small team. Tight for the December deadline if you start in late summer.

Practitioner note

If you serve EU users with generative features, the December 2 deadline means start now, ship by October. Concretely: this week, audit your stack for output provenance — does your chosen API provider (OpenAI / Anthropic / Google / open-weights via vLLM) already attach C2PA metadata? If yes, your remaining work is preserving it through your pipeline. If no, you need a wrap-and-sign step. The free-tier c2patool will handle most cases. Visible-disclosure UI is the easier half — most consumer products already have a “made with AI” badge they can repurpose. For B2B builders, the audit-log requirement is the thing most often missed; if you don’t already log every generation event, that infrastructure work needs to fit the same timeline.

ソース

• EU Council and Parliament agreement on AI Act simplification ↗
• EU AI Act reform talks — IAPP analysis ↗
• EU AI Act implementation timeline ↗

タグ