Sanders’s ASI proposal: what the released document says

Sayer Ji’s article prompted this pass. It raises a worthwhile question about who funds and staffs the evaluation ecosystem. Its stronger legal conclusions need a different evidentiary footing.

The official summary has two alternative capability definitions. Broad human-level-or-better cognition can qualify without additionally proving an ability to overthrow human control. Ease of modification matters in connection with the specified capabilities.

The wording deserves scrutiny. It contains no named Anthropic or closed-model exemption. The proposed agency enforces; the board advises. Development and deployment are the stated targets; a blanket criminal offence for merely downloading an open model is not established.

The sponsor announcement calls the legislation forthcoming. Targeted searches did not recover introduced statutory text or a bill number; direct Congress access was restricted. Appointment rules, review procedures and exact offences remain consequential gaps.

The important question survives the correction: if evaluation becomes a condition of lawful development, whose tests count, who can afford them and how can an adverse result be challenged? The evaluator network and the separate ownership proposals show different routes through which decision rights might shift.

ControlAI’s own account says it consulted on the US framework and prefers a narrower approach than the announced advanced-AI pause. That establishes a policy contribution and a disagreement over scope. It does not identify the authors of every clause or a donor’s drafting instructions.