US lawmakers split on AI regulation amid safety fears

The energy in artificial intelligence moves from researcher to laboratory to product through a circuit of capital risk and iterative failure: a firm spends money it may lose, tests a model against a market that may reject it, and adjusts on the feedback of profit or loss. Congress, returning to Washington this Monday from recess, proposes to intervene somewhere in that circuit. It has not yet decided where, and this is the detail worth pausing on, because the location of the contact point matters more than the fact of contact itself.

The stated justification is that artificial intelligence might spin out of control - a phrase doing a great deal of unexamined work. Spin out of control toward what end, through what mechanism, at whose hands? A system does not become dangerous in the abstract; it becomes dangerous when a specific actor uses it for a specific purpose the rest of the circuit cannot correct fast enough. That is a problem of transmission speed and feedback, not a problem solved by a committee resolving, in the aggregate, that the whole apparatus needs a hand on the switch.

Congress is not agreed on how to intervene, and this disagreement is itself informative. It is not disagreement over facts - everyone in the chamber has read the same warnings - but disagreement over where in the circuit the government’s hand actually belongs: at the point of training the model, at the point of deploying it, at the point of the harm it eventually causes downstream. These are three different circuits with three different points of maximum leverage, and a body that cannot distinguish among them will, if it legislates at all, legislate at the point that is easiest to see rather than the point that is actually load-bearing. This is the perennial error: government intervenes where the wire is visible, not where the current is doing its work.

Suppose Congress licenses the laboratories - the visible wire. The immediate effect is a filter at the point of entry: fewer new firms, because the capital required to clear a licensing threshold is capital that must be raised before the product exists to test against a market. The downstream effect, invisible from the floor of the House, is that the circuit’s error-correction slows. A market with many competing firms fails fast, in public, in ways that generate data other firms use to avoid the same failure. A market with three licensed incumbents fails privately, in board memos, and the correction that would have happened in months happens in years, if at all. The warning Congress is responding to - that the technology might spin out of control - is answered by a policy that makes control-loss less visible, not less likely.

Suppose instead Congress imposes liability at the point of harm, downstream, on whoever deploys the model for the damaging use. This contacts the circuit closer to where the actual energy transfer occurs - a specific decision by a specific actor to put a tool to a specific purpose - and it leaves the upstream circuit, the actual research and iteration, undisturbed. It is a less satisfying policy to announce from a podium. It does not produce a hearing with a chief executive testifying before cameras about the dangers of his own product, an image Washington finds it very difficult to resist staging even when the hearing produces no legislation the following week. But it has the advantage of locating the government’s hand at the point where its grip can actually be felt by the party whose decision mattered, rather than at the entrance to the whole enterprise, where its grip is felt by everyone equally and correctly by no one in particular.

The public, and the economy, and whatever we mean by national security, are all served by a circuit that corrects itself quickly when it errs. The stakes described in this story are real; the fear that something might spin out of control is not misplaced. But the remedy under discussion in Washington this week is being chosen for its visibility, not its contact point, and a government that regulates the wire it can see will discover, a few cycles downstream, that the current went around it.