In March 2023, the Italian data protection authority suspended ChatGPT, citing the absence of any legal basis for collecting Italian users’ data at scale.[1] The suspension lasted one month. During that month, journalists, policymakers, and ordinary users confronted a question that had been obscured by the speed of adoption: on whose terms had this tool entered public life, and who was responsible for what it said?
That question is harder to answer than it seems. The companies behind large language models describe their systems as reading, writing, and reasoning.[2] These verbs import a set of assumptions — about intention, about understanding, about the possibility of being held to account. A system that ‘reasons’ is expected to explain itself. A system that merely compresses statistical patterns is not.
This matters most where the stakes are highest: automated benefit decisions,[3] medical triage tools, content moderation systems operating at national scale.[4] In each case, the vocabulary of machine intelligence serves to obscure the chain of human choices that produced the system and continues to govern its use.
The legal frameworks that do exist — the EU AI Act, national data protection regimes, sector-specific regulations — largely accept the vocabulary of their subjects. They regulate ‘AI systems’ and ‘automated decision-making’ as if these were stable, self-evident categories. The definitions matter enormously. Who controls them controls the scope of the law.[5]
There is a precedent worth examining. When commercial aviation became a mass technology in the mid-twentieth century, regulators did not ask whether aeroplanes ‘understood’ aerodynamics. They asked what failure modes existed, who was liable when they occurred, and what independent oversight was required.[6] The question was not philosophical. It was procedural. The result, however imperfect, was a safety culture that has made commercial aviation one of the most reliable systems in human history.
AI;DR covers these systems concretely: what they do, what they fail to do, how they are governed, and who bears the cost when they go wrong. We do not oppose AI as a category. We do oppose the idea that technical complexity excuses anyone from answering for its effects.