Two transparency regimes, one week apart

Brussels made its rules enforceable and published every word. Washington finalised its framework and classified it. Both are called transparency policy. Only one can be checked.

Within seventy-two hours, EU Article 50 transparency obligations became enforceable and the White House finalised its frontier-model framework and declined to publish it. The contrast is not about strictness. It is about whether anyone outside can tell what the rule says.

What Brussels actually requires

Tell users they are talking to an AI. Make synthetic audio, image, video and text machine-readable and detectable as generated. Label deepfakes. Penalties to €15 million or 3% of worldwide turnover, enforceable by national authorities, with the AI Office simultaneously gaining investigation powers over general-purpose model providers.

The marking requirement is the one with engineering weight. A disclosure banner is a string. Making every generated output detectable in a machine-readable format that survives re-encoding, cropping and platform processing is a provenance pipeline. The regulator evidently knew — providers already on the market have until 2 December for marking and detection specifically, while everything else is live now. That staggering is an admission that retrofitting provenance is a different order of task.

What Washington will not say

Under Executive Order 14409, developers can give federal agencies secure access to covered frontier models for up to thirty days pre-release. The framework defining how was reviewed with Meta, Nvidia, Microsoft, OpenAI, Anthropic and smaller companies on 4 August. The contents will not be released.

A voluntary framework runs on reputation. Companies participate because participation is visible and abstention is visible. Withhold the contents and that engine stops: no outside party can judge whether the bar is demanding or cosmetic, and no participant can be held to a standard nobody can read.

The definitional hole

The executive order does not specify what makes a model a covered frontier model. The criteria that would fill that gap are what is being withheld. So the scope of the regime and the substance of the regime are both non-public, and the question of who is covered is answerable only by the people doing the covering.

There is a real argument for some opacity — an exact capability threshold tells a bad actor precisely how to sit beneath it. That argument covers specific numbers. It does not cover the process, the participant list, or the compliance criteria. Running the whole thing dark is a choice, and a narrower redaction was available.

Meanwhile the two leading labs are drafting the capability threshold their rivals must clear, which will feed into this framework. Private threshold, secret framework, voluntary participation, undisclosed participants. Every individual link is defensible. The chain is not.

European Commission — Safer and more transparent AI → · Fortune — White House won't publicly release AI model evaluation framework → · EU Artificial Intelligence Act — The EU AI Act's transparency rules: a practical guide to Article 50 →