Article 50 enforcement hands the scoping call to the deployer

The trigger in Article 50 is a capability rather than a technology. Providers have to disclose whether their systems can generate or manipulate content, including text, audio and video ^[5]. That phrasing moves the compliance question inside the product: the answer turns on what each system does to material a person will eventually see, and on whether anyone has written that down in a form a third party could check. When a law names an effect and declines to name the class of systems that produce it, the defensible position is an inventory of your own making, with dates on it.

The tradeoff is between two failure modes with different payment schedules. Garraghan's illustration is a software provider drawing on 1,000 different libraries, and his question is whether each one needs its own disclosure ^[8]. Disclose everything and the duty is discharged, at the price of a document that has to be regenerated whenever a dependency changes. Disclose on a capability test you wrote yourself and the maintenance is tractable, but you are holding an interpretation rather than a rule. The first cost recurs at every release; the second arrives once, in a conversation where you have to show your reasoning.

The age of the term is part of why it makes such a poor scope key. Artificial intelligence dates from a workshop held at Dartmouth in 1956, and the artificial neural network from which modern models derive was conceived in the 1940s ^[4][10]. Enforcement therefore began 70 years after the field acquired its name ^[1]. A category that has absorbed seven decades of unrelated engineering will not sort a product catalogue, which is why the workable internal question is what a system can do rather than which decade's vocabulary it belongs to.

Garraghan's sharpest claim is that agents may already sit outside the text, because they are not designed to generate content but to perform and facilitate actions, and he believes there is a strong argument that this exempts them ^[6]. A skeptic would point out that the reading comes from the founder and chief science officer of Mindgard, writing in a vendor council column ^[12], and that authorities tend to read transparency duties for purpose rather than for grammar. That is fair, and it leaves the deployer roughly where it started, because nothing in the material tells us how any authority will read the provision. We do not know yet, and a firm that never recorded which reading it adopted has nothing to defend later.

This is where the quarter's decision sets up the next one. A scope call made once, when a product is first classified, is inherited by every release that follows, and agent features are exactly the kind of change that alters what a system does without altering how it was described. If the internal test is whether a system produces or alters content a user sees, action-taking features fall out of scope quietly unless someone re-runs the test at release. The choice worth making this quarter is who owns that re-test and where its output is filed, because that record is what a later inquiry can actually read.