US lawmakers split on AI regulation amid safety fears
The plan - and there is not yet a plan, only the demand for one, which is itself instructive - requires that the practice of governing something by working alongside it be replaced by the practice of governing something by declaration issued in advance of the knowing. Congress returned to Washington this week, after its recess, to warnings that artificial intelligence might slip its controls, and found itself, as of Monday, without agreement on what an intervention would even consist of. This is not a failure of will, whatever the newspapers say. It is a symptom of the kind of thing artificial intelligence is, meeting the kind of thing legislation is, and the two not fitting.
Consider what is being asked. A statute is a piece of technical knowledge: a rule, stateable in advance, applicable by anyone who can read it. It works tolerably well for activities that are themselves rule-governed - activities where the practitioners already possess, in their bones, a working sense of what the rule is trying to capture, so that the rule merely formalises an understanding already shared. Banking regulation, for all its faults, rests on centuries of accumulated practice about what a bank is and what can go wrong with one. The regulators and the regulated inhabit, more or less, the same practical world.
Artificial intelligence, as an object of legislation, presents no such shared world. The engineers who build the large models, the researchers who worry about their behaviour, and the members of Congress who must decide whether Washington ought to act at all are not participants in a common practice that a statute could merely codify. They are, at present, three different kinds of knower, and the gap between them is precisely what the warnings gesture at without naming. When a member of a subcommittee asks a witness whether a model might one day act against its operators’ intentions, and the witness answers with genuine uncertainty rather than evasion, this is not a failure of testimony. It is the accurate report of an activity that has not yet accumulated the practical wisdom - the tacit sense of what tends to go wrong and how - that any rule would need to draw on to be more than words.
This is why the contested question is not really “how should Washington intervene” but “whether Washington should,” and why that prior question resists resolution by argument alone. The Rationalist temperament, impatient with this, wants a bill: a text specifying permitted architectures, mandated audits, licensing thresholds. Such a text can certainly be written - texts can always be written - but writing it before the practice exists to inform it is rather like drafting navigation rules for a strait no ship has yet sailed, on the assurance of engineers that the strait is dangerous and mapmakers who have not agreed on where the rocks are. The rules would not be wrong so much as premature; they would freeze into statute a set of guesses about a technology whose behaviour is still being discovered by the people building it, day to day, in the laboratories where the actual practical knowledge is being formed.
None of this counsels inaction, and it would misread me badly to suppose it does. The stakes named - for security, for the economy, for the public who will use and be used by these systems - are real, and a tradition that has no politics adequate to a new practice is not thereby entitled to wait indefinitely. What it suggests is a different tempo of response: legislation that follows the practitioners rather than anticipates them, that asks first what the people actually running these systems have learned about their failure modes, and defers the comprehensive statute until there is comprehensive practice to draw on. Congress’s paralysis this week may look like drift. It may also be the beginning of an appropriate humility - the recognition, however inarticulate, that a rule written now would be a rule about a thing not yet understood, and that this is worse than no rule at all.