EU AI Act Article 50: what it means for teams running AI agents

On 2 August 2026, the transparency obligations in Article 50 of the EU AI Act (Regulation (EU) 2024/1689) become applicable. If your organisation deploys AI systems that interact with people or produce content — and AI agents do both — this is the deadline that matters, whether or not you are based in the EU: the Act applies wherever a system's output is used in the Union.

What Article 50 actually requires

The full text of Article 50 sets out four transparency duties. In practical terms:

• Disclosure of AI interaction. Providers must ensure people know they are interacting with an AI system, unless it is obvious from context.
• Marking of synthetic content. Providers of systems that generate text, audio, image or video must mark outputs as artificially generated, in a machine-readable form where technically feasible.
• Deployer disclosures. Organisations deploying emotion recognition, biometric categorisation, or deepfake-style generated content must inform the people affected.
• Published text disclosures. AI-generated text published to inform the public on matters of public interest must be disclosed as such, unless it has undergone human editorial review with a person taking responsibility.

Non-compliance carries penalties of up to €15 million or 3% of worldwide annual turnover, whichever is higher (Article 99(4)).

Why agents make this harder than chatbots

For a chat window, disclosure is a banner. For an agent — a system that plans, calls tools, edits files, sends messages and commits code — the transparency question is not "did we tell the user it's AI?" but "can we show what the system actually did?"

Regulators, auditors and customers will not take a log file's word for it. Application logs are written by the same system they describe, can be edited after the fact, and prove nothing to a third party. When an agent's action is challenged — a wrong transaction, a leaked file, a published article — "our logs say it behaved" is an assertion, not evidence.

From documentation to verification

The robust way to meet an obligation to demonstrate what an AI system did is to make the record tamper-evident and independently verifiable:

1. Sign the code. The publisher signs the agent skill with an ed25519 key over a content-addressed manifest — so it is provable which code ran.
2. Seal the execution. The operator runs it in a kernel-namespace sandbox and signs a record of inputs, outputs, timing and sandbox policy — as SHA-256 hashes, so the record proves integrity without exposing the data.
3. Let anyone verify. A third party checks the receipt with one command, offline, with no access to the operator's systems:

$ pip install traceseal-verify
$ traceseal-verify receipt.json
[OK] receipt.json — operator signature verified

Because the receipt is canonical JSON under a signature, any tampering — by the agent, the operator, or anyone downstream — breaks the seal. That converts an Article 50 conversation from a documentation exercise into a one-command check.

What to do before 2 August 2026

• Inventory your agents. Anything that interacts with people or generates content is in scope for at least one Article 50 duty.
• Decide your evidence standard. Screenshots and logs are trust-me evidence; signed receipts are check-it-yourself evidence.
• Instrument now, not after an incident. Receipts only cover executions that were sealed at the time. A record you start keeping after the dispute begins proves nothing about what came before.

The receipt format, verifier and transparency log are open — the spec, the verifier on PyPI, and the public log. You can adopt the format without adopting us.