⚡ Quick Summary
Published by GovAI, this policy brief examines whether Article 50(2) of the EU AI Act requires providers to label actions produced by AI agents, as well as synthetic audio, image, video and text content. It focuses on online actions, including web requests, browser actions, purchases and social-media interactions, which may be mistaken for genuine human activity and thereby influence beliefs or behaviour. The brief states that Article 50 is expected to be in force from 2 August 2026 and identifies an upcoming Code of Practice as an opportunity to clarify compliance measures for parts of the Article.
Alan Chan argues that agent actions likely count as AI-system outputs under Article 50 because Article 3(1) includes decisions that can influence physical or virtual environments, its examples of outputs are non-exhaustive, and many virtual actions are represented as text. The proposed approach is to attach metadata to web requests and browser actions, make labels verifiable for authenticity and integrity, and standardise their implementation. An appendix supplies proposed Code text covering outputs, watermarks and metadata, verification, transmission to external services, and cooperation among signatories.
🧩 What’s Covered
The brief develops its proposed interpretation and implementation approach in the following order:
- Problem and scope: The introduction explains how AI-agent activity such as liking, sharing or commenting on political content could be mistaken for human opinion. It frames labels as a way to make users more conscious of attempted influence and to help platforms address influence campaigns.
- Article 50(2) and outputs: The text reproduces the requirement that providers of systems generating synthetic audio, image, video or text mark outputs in a machine-readable and detectable format. It then asks whether actions, including purchases and social-media interactions, are outputs that must also be marked.
- Actions as outputs: Three strands of reasoning are offered: actions can be decisions affecting physical or virtual environments; the Article 3(1) examples are non-exhaustive; and an output can be understood as what a system infers how to generate from its inputs. Examples include online purchases, social-media posts and requests to weather services.
- Transparency purposes: The brief connects action marking to misinformation and manipulation risks identified in Recital 133. It describes how labels could support user awareness, platform tagging or filtering, measurement of AI-generated misinformation, enforcement of platform policies, and provider accountability through provider-specific identifiers and pseudonymous user tags.
- Metadata and interoperability: Web requests are described as structured text data that could carry a field such as “AI_generated: True”, while browser actions could include an indicator detectable by web pages. The brief recommends standardising and publicly documenting metadata methods, while leaving providers flexibility for other action types where feasibility or usefulness is uncertain.
- Verifiability and limitations: The brief argues that effective, interoperable, robust and reliable marking requires verification of the provider that created a mark and of whether it has been altered. It discusses digital signatures, then identifies removable metadata, failures to pass metadata onward, the constraints of text watermarking, user misunderstanding and the limits of provenance as challenges.
- Open questions and draft commitment: The conclusion identifies enforcement, certificate authorities and additional provenance requirements as unresolved. The appendix proposes Code language on watermarks, metadata, verification, onward transmission, standardisation and a specified criminal-law exception.
💡 Why it matters?
For developers and providers of AI agents, the brief translates a disputed legal question into concrete system-design considerations: which outgoing actions may need marking, where metadata can be attached, and what recipients need in order to verify it. It is especially relevant where agents interact with web services, browsers, social platforms or communications tools.
For platforms and governance teams, the proposal links provenance information to practical uses including identifying AI-generated activity, enforcing existing platform policies and investigating misuse. It also makes clear that a basic synthetic-content label does not itself establish credibility or accuracy, so implementation and user-facing handling of labels remain important.
❓ What’s Missing
The brief does not settle the legal interpretation of Article 50; its central conclusion is that action marking is “likely” required, and it expressly states that the work is not legal advice. It leaves enforcement mechanisms, the appropriate certificate authorities for digital signatures and the circumstances in which AI-user identities should accompany outputs as open questions. The proposed Code text covers web requests, browser actions and other outputs “as appropriate”, but does not provide detailed technical specifications for every type of agent action. It also acknowledges that metadata can be removed or not passed to end users, that watermarking actions is unclear, and that users may ignore or misinterpret provenance signals. The document notes that the brief has not undergone an official peer-review process.
👥 Best For
This brief is best suited to AI-agent providers considering Article 50(2) compliance, technical teams designing outgoing web-request or browser-action metadata, and platform governance teams deciding how to detect, tag or filter AI-generated activity. It is also useful to participants developing or assessing voluntary commitments for the forthcoming Code of Practice.
📄 Source Details
Labeling of AI Agent Activity in Article 50 of the EU AI Act is an English-language GovAI policy brief by Alan Chan, dated November 2025. The supplied PDF is 12 pages, with the substantive brief numbered pages 1–10, followed by a cover and closing page. It states that GovAI policy briefs are short, accessible pieces that have not undergone official peer review.
