Should AI legislation wait for demonstrated risks to emerge?

Amodei argues that AI models have become "tools of global and national strategic consequence," and that this changes the policy problem. The evidence he cites is recent: "in the last few months," the evidence of AI's power and risks "has become undeniable." Claude Mythos Preview is "perhaps the most emblematic example," and its cyber risks, with "the potential for disruption of the financial sector, critical infrastructure, and national security," are for him proof of the strategic point. He treats them as a first instance, not a limit: "biological risks may soon follow, and that serious AI autonomy risks may not be far behind."

The argument turns on timing. In 2023-2024, Anthropic could see possible harms, including biological weapons "that could threaten millions," but it was "less clear" what form the risks would take, how to test for them, or how they would play out. From that uncertainty he draws the warning that "legislation written ahead of time would end up being ineffective," creating "pointless or low-value compliance requirements while missing the most crucial sources of actual risk." The essay presents the present as different. Anthropic is releasing "a legislative proposal on frontier model testing" as one of "first steps to signal our seriousness." The excerpt does not state the link, but its sequence implies that the evidence now in hand is what makes a testing law concrete enough to write.

The nearest library notes approach the same governance question from other angles. Can regulation keep pace with AI's rapid evolution? also rejects fixed rules, but because models outpace them; Amodei's objection is that rules written before a risk takes shape miss it. Both reject static design and differ on what replaces it. How do we stop AI systems once they are already deployed? takes pre-release regulation as the frame and asks about stopping deployed systems. The essay addresses neither; its worry is rules written too early. On capability, its cyber evidence concerns a single model. Where do frontier AI models actually pose the greatest risk today? reports threshold zones across frontier models, with most still green for cyber offense. The two are compatible, because the essay's claim concerns strategic consequence rather than how many models cross a threshold.

The excerpt does not establish what Mythos Preview's evaluations found, how its cyber risk was measured, or what the proposed testing legislation would require. "Beyond doubt" is Amodei's characterization, and no test results appear. The job-displacement framework is described only by its existence and the "substantial financial backing" Anthropic intends to provide. The excerpt also omits the footnotes its superscripts point to. What it supports is narrower than its tone: Amodei judges that legislation tied to demonstrated risk is more useful than legislation written ahead of it, and that frontier AI now carries strategic consequence. It does not show that a testing law would catch the biological or autonomy risks he anticipates, which remain forecasts.