Can industry self-regulation slow AI without government enforcement?

Karpf argues that the pacing plan in Dario Amodei's essay "We Must Pace the Frontier" is self-regulation that serves the company proposing it, and that its first safeguard, embedded third-party evaluators, fails without state power behind it. Karpf grants the premise, quoting Amodei that "A race to the bottom, spurred by commercial incentives, can make [the risks of AI doing serious harm] more acute," and calls the proposals appealing "at a high enough level of abstraction." The argument turns on the details. Karpf notes that Anthropic is "preparing for its own IPO" and is "currently winning that AI race," and concludes that the three proposals "all benefit Anthropic" and "are all subject to revision if-and-when they get in Anthropic's way."

The argument has two supports. The first is an inference from motive: a company with IPO plans and "two consecutive quarters of profitability" is read as having commercial stakes that outrank its stated concern, since it is "not worried enough to delay going public." The second is an analogy. Amodei cites banking's embedded regulatory "supervisors" as precedent for the evaluator proposal. Karpf answers that those supervisors exist "because the government requires them," and that without "the force of government oversight (and the looming threat of draconian fines), this sort of monitoring doesn't work." Karpf ties this to Lina Khan's point that the Federal Trade Commission could enforce far more aggressively, and closes that "It is a mistake to let the AI industry shape the contours of its own regulatory system."

The essay sits near Can slowing AI development resolve who stops deployed systems?, which separates measures that act on the pace of advance from the open questions about who intervenes once a system is deployed. Karpf adds a third "who": who designs and enforces the pacing measures in the first place. Where that note leaves the deployed-system "who" open, Karpf answers the pacing "who" with government rather than the lab. Karpf's point that monitoring needs state force also bears on the enforcement gap in How do we stop AI systems once they are already deployed?, though Karpf addresses monitoring, not halting a running system. The reading is a counterweight to the self-governance posture of the evaluator proposal, not a rejection of the evaluators: the claim is that they need force behind them to work.

The excerpt does not establish most of what the full essay contains. It quotes Amodei at length only on the first of three proposals, and gives one clause each to the self-regulatory regime and to bilateral coordination with China, so the claim that the whole scheme favors Anthropic is tested against one proposal in the excerpt. The motive argument rests on public commercial facts and an inference about what a delayed IPO would show; the excerpt offers nothing on Amodei's private reasoning. The banking analogy is asserted, not studied: no evidence on supervisor effectiveness or on AI monitoring appears. At the strength the evidence allows, this is a pointed position on who should hold the pacing lever, built on analogy and an incentive inference. It leaves open whether third-party evaluators could be credible under a voluntary regime, which is a design question the banking precedent does not settle on its own.