What Congress Actually Learned From a Year of AI Testimony

The first round of artificial intelligence hearings in the United States produced some memorable exchanges, and left behind a durable impression that legislators were being asked to regulate something they did not fully understand.

That impression was not entirely unfair. It did obscure a more interesting process: the slow accumulation of a working consensus about what a statute would need to contain.

The four things the hearings settled

• That the industry itself would accept some form of regulation, which was not a foregone conclusion.
• That the sharpest disagreements were between companies rather than between parties.
• That compute thresholds, disclosure duties and incident reporting had become the vocabulary of the debate.
• That enforcement capacity, not statutory text, is the binding constraint.

Hearings ran monthly; bills did not

Legislative progress has been slower than the hearing schedule implied. Proposals have been narrowed, folded into other vehicles, or left pending while members negotiated over preemption, liability and the treatment of open-weight models.

The practical result is a patchwork: sectoral regulators applying existing authority to new uses, state legislatures moving faster than the federal government, and companies navigating several regimes at once.

Build for the strictest rule you can plausibly face

For anyone building on these systems, the useful conclusion is that no single rulebook is coming soon. The compliance environment will be layered and jurisdiction-specific, and it will be shaped as much by procurement requirements and insurance terms as by legislation. Designing for the strictest plausible standard is usually cheaper than retrofitting.

The safety engineers who were never called

The witnesses at the most consequential hearings were predominantly executives and academics. The people with direct operational knowledge — safety engineers, evaluation specialists, data workers, the officials who had already deployed systems in public services — appeared far less often.

That is a function of who is available and who is considered authoritative. It is also why the resulting legislation has tended to address the concerns of frontier labs rather than the concerns of the organisations that would have to use their products.

What would break the logjam

The clearest path to legislation is a widely publicised harm with identifiable victims, which is how most technology regulation has historically been enacted. The absence of such an event so far is one of the main reasons the debate remains abstract.

That is not an argument for waiting. It is an observation about why the current pace is what it is, and a prediction that the field’s regulatory environment will change suddenly rather than gradually.