EU AI Act Transparency Provisions Take Effect as Tech Giants Launch Compliance Efforts
Nashnova编辑部
The EU AI Act's transparency provisions took effect on August 2, requiring chatbots and AI-generated content to disclose their origin; Anthropic, Google, Meta, and OpenAI have all begun watermarking and disclosure compliance — and these rules are fast becoming a de facto global benchmark for AI transparency.
What exactly kicked in on August 2?
The AI Act's transparency and disclosure provisions are now enforceable — chatbots must tell users they are talking to AI, and AI-generated text or images must be labeled as machine-made.
In plain terms = from this date, if European users encounter AI-made content, the platform is legally required to say so.
The EU AI Office has not yet opened enforcement against any company, but it now holds the power to demand information and access models — This means → the enforcement toolkit is loaded; it just hasn't been fired.
How are the tech giants responding?
Anthropic announced that future Claude models will embed watermarks — invisible markers a machine can detect but a human eye cannot — in generated text, and plans to release a "watermark detection API" for outside verification.
Google and Meta both said they will adopt and develop transparency and watermarking tools; Meta added it will work with the EU AI Office to ensure compliance is "practical, interoperable, and genuinely useful."
OpenAI published a training-data summary and embedded "provenance signals"; Microsoft focused on internal governance and risk-management adjustments — This reflects different entry points: some companies watermark at the product layer, others restructure governance first.
Why might watermarking become a global standard?
Amy Worley, managing director at Berkeley Research Group, noted that watermarking does more than satisfy EU rules — it doubles as a litigation defense tool, proving "this content was AI-generated" in copyright or misinformation disputes.
She stated: "We need a standard. This technological moment is moving fast, and the EU AI Act at least gives us one."
This means → even if no other country copies the Act clause by clause, once companies treat it as the default transparency benchmark, Brussels has effectively shaped global AI rules beyond its own jurisdiction.
How long until the Act fully lands?
The August transparency rules are only step one. Provisions covering high-risk AI — education, biometric identification, immigration — do not take effect until December 2027; rules for AI in physical products follow even later, in August 2028.
In plain terms = what has launched so far is the "label it properly" layer; the hard rules deciding which AI applications are permitted and which are not are still two to three years away.
Patrick Van Eecke, a partner at Brussels tech law firm Cooley, called it "a messy piece of legislation" — This reflects real industry doubt over whether Brussels can convert a complex legal text into a framework companies can actually follow.
市场有风险，内容仅供研究参考，不构成投资建议。