The Information Machine
Updated today·Day 2·first covered 22 Sep 2026·4 sources

OpenAI's proposal for international AI oversight

The gist

Sanders-Casar Bill Would Ban Superintelligence, Create Federal AI Dept

The Sanders-Casar bill represents a concrete legislative attempt to define and prohibit superintelligence at the federal level, including destruction requirements for confirmed systems. The RAND paper and OpenAI framework reflect ongoing efforts from research and industry sides to shape how the US government approaches frontier AI governance.

The full picture

Senator Bernie Sanders and Representative Greg Casar have introduced legislation that would immediately pause training and deployment of AI systems trained with at least 10^25 operations until new federal safety rules are established. The bill would prohibit systems with 'superintelligence precursor' capabilities, defined as including the ability to automate AI research, resist shutdown, self-modify, or assist with weapons of mass destruction. A new Department of Artificial Intelligence would have authority to force flagged models offline, and confirmed superintelligence must be destroyed. Advanced AI developers would require federal charters and must grant government access to systems, staff, and physical infrastructure. Violations carry asset and IP forfeiture, described in the bill as a 'corporate death penalty,' and certain individuals could face up to 20 years in prison.

A RAND Corporation paper separately recommends the US adopt a 'Freedom of Action' strategy rather than commit to any single approach, arguing the optimal path to managing superintelligence is unknown and preserving options is prudent. RAND groups seven archetypal strategies into three families: coexistence (dominance, co-development, preparedness), denial (moratorium, deterrence, continuity of society), and acceleration. The paper characterizes the current US posture as closest to acceleration, describing it as analogous to 'sitting in a car and spending all your resources on making the car go faster and upgrading the engine, and nothing on proactive safety measures like seatbelts or headlights or brakes.' Five key uncertainties RAND identifies as governing which strategy is optimal are danger proximity, coexistence feasibility, restraint feasibility, decisive strategic advantage, and suppression feasibility. The stated purpose of the recommended strategy is 'to build and preserve the ability to secure U.S. geopolitical advantage and ensure humanity's survival with agency through the transition to Superintelligence.'

OpenAI published a framework proposing four priority areas for third-party safety assessments of frontier models: safety cases spanning the model lifecycle, critical safeguards evaluation, capability and alignment evaluations, and independent investigation of misalignment incidents. The framework calls for assessors to have 'proportionate access,' including grey-box access for adversarial testing, and distinguishes between a 'safety claim' (a specific, assessable assertion about model behavior) and a 'safety case' (a structured argument connecting claims to evidence about why risks are adequately managed). OpenAI states it is already in conversation with multiple third parties and aims to help grow a diverse independent assessor ecosystem.

Researchers published in Nature transplanted human stem-cell-derived cortical organoids into mice with deliberately depleted neocortex and hippocampus, producing chimeric xenocortical mice. The human cortical neurons integrated electrically and functionally with the mouse nervous system, showing organized activity resembling developing circuits. Behavioral analysis placed these mice between normal controls and brain-depleted mice, and they performed above chance in a maze memory test similarly to normal controls. The primary stated application is a better platform for testing human brain therapies in animal models, though the research raises ethical questions about chimeric intelligence.

How it developed
23 September 2026

Sanders and Casar unveil legislation to ban superintelligence and create a Department of Artificial Intelligence

RAND published a paper September 22 recommending the US adopt a 'Freedom of Action' strategy for the transition to superintelligence, grouping seven archetypal approaches into coexistence, denial, and acceleration families and framing the current US posture as closest to acceleration. OpenAI published a separate framework the same day for third-party safety assessments of frontier models, defining a 'safety case' as a structured argument connecting claims to evidence and calling for grey-box adversarial access for evaluators.

First citedOpenAI Blog
22 September 2026

OpenAI publishes framework for third-party safety assessments of frontier AI models

OpenAI published a proposal September 22 for a US-led global framework covering AI capability evaluations, risk assessment, and incident reporting, with thresholds triggering human review as AI autonomy in research increases, concurrent with US-China talks on an AI incident notification mechanism. A multi-institution paper from Harvard, Cambridge, and Wharton published the same day argued AI pacing will occur and called for principled frameworks, while former Google CEO Eric Schmidt said a pause is unrealistic and Dean Ball called for 30-day pacing of frontier AI development starting November 19.

21 September 2026

Import AI newsletter covers RAND 'Freedom of Action' paper and Nature xenocortical mice research

Sources
The daily email

Want this in your inbox?

I send one email each morning with the stories that moved. If you would rather just read here, that works too.

Subscribe free