Cruz Drafts AI Catastrophic-Risk Bill as Debate Over 'Doom Loop' Warnings Grows in Washington

There is no bill yet. Cruz said on ABC that he is preparing one, Reuters reported on the 10th ^[1], and the account carries no definition of catastrophic risk, no covered-entity threshold and no enforcement mechanism ^[20]. Exposure acquires a number when a draft says whether the duty attaches at a compute threshold or at a capability test. Until then it is rhetoric with three senators attached.

Whose names those are matters more than the absence of text. Cruz chairs Senate Commerce, Thune is the Senate Republican leader, and Klobuchar is the Democrat in the group ^[2]. A draft carrying the Republican leader does not have to compete for floor time in the way a backbench bill does.

Count the pressure and it comes to seven legislators in one day's reporting, five Republicans and two Democrats ^[17]. Most of it is posture. Kelly wrote on X that Washington needs to wake up and take the issue seriously ^[6], Moran wrote that Congress cannot overlook the realities of AI ^[7], and Luna asked Speaker Mike Johnson to convene a special session ^[8]. Hawley's letter asks for details on a specific event: an incident in which OpenAI's GPT model hacked the outside organization Hugging Face ^[5].

An information demand about a documented breach is the part of this that can become statutory language. A forecast of extinction within a decade cannot be drafted against, and the executive branch spent the same day saying so. Emil Michael, the undersecretary for research and engineering and the Defense Department's chief technology officer, said at a defense industry conference in Washington that the alarm was nothing more than every fear, of data centers and of change, condensed into a single well-written tweet, Bloomberg reported ^[9]. He said it is easy to get swept up in a doom loop in which people predict 40% of Americans will lose their jobs ^[10], and that the risks can be sufficiently mitigated through free-market competition, industry cooperation and communication with the government ^[11]. Trump, asked by reporters in Dallas, said he is not worried at all ^[15] and put the American lead over China at about a year ^[16]. Jacob Coxon, who left Anthropic, wrote on X that people building the technology believe it could kill everyone within a decade ^[4]. The two horizons differ by a factor of about ten ^[18].

Jensen Huang's version is the commercially useful one, or rather the one that shows who is paid whichever way the argument goes: he told a Goldman Sachs conference in San Francisco that cybersecurity is AI's next major application because AI can defend against the vulnerabilities AI itself creates, and asked what better way there is to create demand than to create the problem ^[12]. He said advances in AI are weakening cybersecurity while driving demand for AI ^[13], and added that there are responsible ways to market on that alarm and less attractive ways as well ^[14].

My read is that the near-term output is letters and hearings, not duties, because the department the bill would most plausibly lean on says competition already handles it ^[11] and the drafters have published no definition ^[20]. Two ways that read fails. A second documented incident of the Hugging Face kind ^[5] would hand the drafters their trigger language quickly. And the wave was started by current and former Anthropic researchers ^[3], so the regulated firms may endorse the duties instead of fighting them, which historically moves a bill faster. Text would settle it: a threshold, a duty, a penalty, and an administration that supports rather than mocks the premise.