A Right to One's Own Method: Worker-paired AI memory and the recording paradox in Union law

Abstract

A class of AI system now entering professional workplaces maintains a persistent, individualised record of how a particular worker works, and adapts its outputs to that record. Existing and proposed Union instruments in this field — the GDPR’s automated-decision provisions, the Platform Work Directive, the AI Act’s high-risk regime, and the algorithmic-management measures now under consultation in the Commission’s proposed Quality Jobs Act — are built on a system’s supervisory or decisional function. A system that neither monitors nor decides, but assists, falls outside all of them. The Commission’s second-phase consultation documents of 20 July 2026 draw that line expressly, excluding artificial intelligence “more generally, in particular generative AI” from the algorithmic-management frame. The paper’s doctrinal claim is that this omission produces a paradox in the law of trade secrets. Article 1(3)(b) of Directive (EU) 2016/943 preserves employees’ use of experience and skills honestly acquired, and the definition of a trade secret is understood to exclude such experience and skills. That protection has operated for decades on a factual predicate rather than a legal test: method was tacit, and so could not be enumerated, copied or left behind. Systems whose central function is to convert tacit method into a retrievable record remove that predicate. In the same operation they arm the employer’s competing claim, because the recorded method begins to satisfy the trade-secret definition, the requisite reasonable steps being the ordinary access controls of the storage the employer administers. The material that was a worker’s inalienable experience on Monday is the employer’s protectable secret on Tuesday, and the only thing that changed is that a system wrote it down. Neither Union portability regime closes the gap, and the paper’s second finding is that they fail for the same reason: both Article 20 GDPR and Chapter II of the Data Act exclude derived and inferred data, in each case to protect the investment of the party that did the deriving. Applied to employment, where the additional investment consists of observing a worker doing their job, that allocation assigns the product of a person’s professional judgement to the party that bought the software. The remedy proposed is allocation rather than prohibition: a duty to maintain the episodic and procedural records separately, and to place the procedural record under the worker’s control. An indicative provision is set out in full.