Lilith.
⌕
Editorial illustration: Helen Toner shifts AI governance from trusting CEOs to enforceable rules
Lilith illustration · editorial remix

In an interview with The Verge, Helen Toner rejects the idea that the safety of the most capable AI systems can be secured by the character of a few executives. The former OpenAI board member speaks as someone who voted to remove Sam Altman, then watched him return while the original oversight arrangement collapsed.

OpenAI showed that a board without leverage cannot hold the brake

OpenAI's nonprofit board was meant to protect the original mission from pressure for growth and profit. When it used its strongest intervention in 2023 and removed the CEO, the decision did not last. Toner says she would cast the same vote again, while acknowledging that the board could have built stronger relationships with company executives.

Her central argument is therefore institutional rather than a personal verdict on one manager. If the future depends on a single individual being trustworthy, the institution has already failed in its design. CSET, which she now leads, has about 50 people and has translated technical and security issues into material for US policymakers since 2019.

Safety thresholds should slow development when a lab cannot keep up

Toner describes labs that use AI across a growing share of their own research while lacking time to implement every known safety measure. She proposes a clear bar for risk assessment and risk management before another model is trained or released. A lab that meets the bar can proceed. One that does not should take more time.

For companies buying AI, the important shift is from promises to process. Accountability can be inspected only when mandatory tests, decision records and an authority outside the product team can stop a release.

Public markets could add scrutiny and a fresh growth pressure

Toner does not treat an AI lab going public solely as a threat. Public companies face more structured governance, reporting and external scrutiny than private labs. Quarterly investor expectations could also accelerate the race for capability. The concrete powers of independent committees will matter more than their names.

The interview remains the view of one former director, not an audit of today's OpenAI. Her experience nevertheless provides a rare stress test of a mechanism designed for exactly the moment when mission and management conflict.

The first real test will be an inconvenient decision

New rules will matter when a safety committee delays an expensive model, a regulator demands uncomfortable data or a board rejects management's plan. Until then, voluntary commitments remain promises. The signals to watch are published thresholds, the powers of independent oversight and records showing when that oversight actually changed a lab's decision.

Lilith's verdict

Toner offers a simple test: an emergency brake is defined by whether it stops the train against the driver's wishes, not by the label beside the handle.