All posts

The AI Act clause that is already live

Article 50 transparency and the Commission’s GPAI enforcement powers are in force; the Annex III high-risk regime is not, and its date has moved. What that asymmetry means for how you build.

The part of the EU AI Act that touches most people building with language models is already in force, and it is not the part the headlines rehearsed. Since 2 August 2026, the Commission’s enforcement powers over general-purpose AI apply (the Article 101 fining regime), alongside the Article 50 transparency obligations. The high-risk regime — the one most coverage treats as the whole Act — is not yet operative.

What Article 50 asks for

Article 50 is a disclosure duty. Systems that interact with people have to make it apparent that the person is dealing with an AI system, unless it is obvious from context. Synthetic audio, image, video, and text carry a machine-readable marking duty. The through-line is provenance: a user, and increasingly a downstream system, should be able to tell where a piece of content came from and whether a model produced it.

The high-risk timeline moved

A common planning error right now is to treat Annex III high-risk obligations as live. They are not. The Digital Omnibus package (Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026 and in force from 27 July 2026) moved the Annex III high-risk application date to 2 December 2027, with the Annex I embedded-product track following on 2 August 2028. Building a compliance plan around a date that has shifted is its own kind of risk.

Why this favours things you can inspect

Transparency and provenance obligations reward architectures whose behaviour can be shown, not just asserted. A control that produces an auditable record of what it saw and what it decided is easier to evidence than one whose reasoning lives inside a model’s weights. That is a quiet advantage of deterministic controls over probabilistic ones when the question an auditor asks is show me, not convince me.

It is the same reason a pattern-and-structure guardrail like BonkLM leaves a tamper-evident trace on every verdict: an inspectable decision is worth more, under a transparency regime, than a confident one you cannot reconstruct.

This is a reading of public regulatory text, not legal advice. Dates and article numbers are those in force at time of writing (August 2026); confirm against the current Official Journal for your own compliance work.