Schrödinger’s Data: Identifiability and Large Language Models

How’s the Cat? Ask Wigner’s Friend

• Alessandro El Khoury

Between 2024 and 2026 a data protection authority, the EDPB and the Court of Justice each answered some version of the question whether a trained language model contains personal data. And none of the answers agreed. This article argues that the question is malformed: A model is at once the object whose status is in question and the instrument by which identification is performed. Its status under Art. 4 para. 1 GDPR accordingly cannot be settled in advance of a processing operation. What can be regulated is the operation: the training run, the publication of the weights, the prompt. In this regard the quantum framework proposed by the author in 2018 is revisited and revised in this article.