Will the New Congress Ban Superintelligence? | Rob Wiblin

Election Outcomes and Congressional Composition

• Democrats secured a decisive victory, holding at least 250 seats in the House of Representatives and 56 seats in the Senate.
• Democratic gains include winning seats in Mississippi, South Carolina, and Kansas, as well as sweeping every office in Texas by 11 points.
• AI emerged as a critical fourth or fifth driver of voter backlash, despite not ranking in the top three issues of the election.

Public Opinion and Political Dynamics

• Public sentiment regarding AI is remarkably negative and bipartisan, with 80% of Americans favoring safety rules even if they slow development (including nearly 80% of Republicans).
• 67% of the public believes AI is advancing too quickly, while only 2% believe it is advancing too slowly.
• Despite the President's opposition to binding AI regulation, 60 Senate votes may be achievable due to bipartisan concern, led by Republican figures like Josh Hawley, John Kennedy, and Marsha Blackburn.
• Overcoming a presidential veto would require 67 Senate votes and 290 House votes, a threshold currently considered unlikely without significant policy watering-down.

Four Immediate Pathways for Congressional Action

• Popular Bipartisan Measures: Legislation regarding model weight theft by China, strengthened export controls, law enforcement against chip smuggling, whistleblower protections, mandatory incident reporting, and "kill switches" could secure 70–90 Senate votes.
• Must-Pass Legislation: Critical funding bills (e.g., defense bills) are veto-proof due to shutdown risks, providing a vehicle to negotiate AI safety funding, specifically for the Center for Advancing Innovation and Standards for Superintelligence (CASEY).
• Confirmation Powers: The Senate can block or delay confirmations of Cabinet secretaries and senior officials whose AI views are misaligned with congressional consensus.
• Investigative Hearings: Congress can compel AI company staff to testify under oath regarding specific incidents (e.g., the OpenAI AI swarm hack), the safety of recursive self-improvement, current model training parameters, and undisclosed failed safety tests.

Forward-Looking Risks and Critical Thresholds

• META's recent report suggests AI R&D could be fully automated as early as 2027, accelerating research progress to a point requiring immediate congressional intervention.
• A "warning shot" event (ranging from long-term power-seeking AI to infrastructure sabotage or AI-caused pandemics) is predicted to be the catalyst for a supermajority shift in policy.
• Following a major crisis, the political environment is expected to shift toward actively interfering with model training and deployment rather than minimizing disruption.
• The most likely public rallying cry post-crisis will be to "ban superintelligence" or "ban superhuman AI," which enjoys overwhelming public support.

Specific Legislative Proposals and Mechanisms

• Compute Allocation as a Speed Dial: Mandating that AI companies spend a minimum percentage of compute (e.g., 100% down to 70%) on inference for outside customers could effectively pause or slow frontier research, similar to the Federal Reserve setting interest rates.
• Throttling Recursive Self-Improvement: Prohibiting the use of models less than six months old for internal research would break the feedback loop for rapid self-improvement, allowing time for alignment testing.
• Chain-of-Thought Restrictions: Banning models capable of high-complexity tasks without visible chain-of-thought outputs to enhance monitorability and reduce scheming.
• Data Center Construction Bans: While anti-data center sentiment won 40 midterm ballot measures, experts note this must be paired with strict export controls and chip disconnection to effectively reduce risk.
• Sandboxed Environments: Requiring frontier training to occur in fully air-gapped facilities to prevent unauthorized access or misuse.

Institutional Capacity and Funding Gaps

• CASEY is critically underfunded ($15 million annually) and understaffed (dozens of employees) compared to the UK's AI Security Institute (100+ technical staff, ~8x budget).
• The UK has committed an additional £300 million over four years plus hundreds of millions for biosecurity and cyber agencies, putting US state capacity at a 50x disadvantage relative to economic size.
• Urgent action is required to double CASEY's budget and embed independent auditors within AI companies to ensure accurate risk assessment and enforcement.

Safeguards Against Executive Overreach

• Increased regulatory power shifts to the executive branch necessitates robust judicial review processes to challenge capricious decisions within two weeks.
• Companies must have the right to object to rulings that violate the First Amendment or lack credible safety justification, though preliminary injunctions should be restricted to prevent preemptive blocking.
• In worst-case scenarios where executive enforcement fails or becomes partisan, Congress may need to support state-level regulation (e.g., California, Texas) rather than pre-empting state laws.

Strategic Recommendations for Policy Preparation

• Technical experts and legislators must draft detailed legislation now, as the ideas available during a crisis dictate the outcome of the response.
• Key figures like Scott Weiner are expected to join Congress with the capacity to develop sophisticated proposals on AI training mechanics.
• The immediate two-year window is identified as the most critical period to avert an AI catastrophe, contingent on maintaining political will and technical readiness.