'We're Pretty Screwed' if AI Rests on One Trustworthy CEO, Says Helen Toner, the Ex-OpenAI Board Member Who Voted to Fire Altman
On The Verge's Decoder, the Georgetown CSET director said she would vote the same way again, called alignment 'not on track to be solved,' and argued for a higher safety bar before labs train new models, not a blanket pause. She also sees Anthropic's IPO as a possible governance upgrade.
Published
Last updated / Fact-checked on

Illustrative photo of studio microphones. It does not show the Decoder recording.
Three years after she helped remove Sam Altman as chief executive of OpenAI, only to watch him return within days, Helen Toner has not changed her mind about the vote. Asked on The Verge's Decoder podcast whether she would relive that moment the same way, she answered: "Oh, for sure. I mean, we went into it knowing that it was going to suck."
The bigger point she drew from it was not about Altman. Pressed by Verge editor-in-chief Nilay Patel on whether AI chief executives can be trusted, Toner said: "I think my stance on that is pretty clear, and I also think that if our hopes for the future are pinned on any one individual being trustworthy, then we're pretty screwed."
Toner now runs the Center for Security and Emerging Technology (CSET), a roughly 50-person policy research group at Georgetown University. The Verge published the edited conversation on Nov. 14 as part two of a series on AI in real life.
At a glance
| Item | Detail | Source |
|---|---|---|
| Who | Helen Toner, executive director of Georgetown's CSET and a former OpenAI board member | The Verge |
| Published | Nov. 14, 2026; recorded the day after the midterm elections | The Verge |
| On the 2023 vote | Would do it again; would have built stronger ties with executives | Toner |
| On trust | Relying on any one trustworthy person means "we're pretty screwed" | Toner |
| On alignment | Getting AI to do what users want is "not on track to be solved" | Toner |
| Policy idea | A higher confidence bar before training or releasing new models | Toner |
| What she's watching | State frontier-AI laws and whether the new Congress overreaches | Toner |
The vote, revisited
Toner was one of four OpenAI board members who voted to fire Altman in November 2023, the episode now often called "the blip." Altman came back, Toner resigned and OpenAI has since restructured, The Verge noted. In his introduction, Patel framed the episode as the moment an accountability mechanism meant to keep OpenAI on its safety mission "broke."
Toner declined to give a full account of what she would change, saying the answer would get tangled in "legal restrictions." She offered one lesson: the board should have built "stronger relationships with executives on the team," which she said could have changed how the company related to its board.
A bar, not a pause
The policy core of the interview was Toner's case for slowing the frontier without freezing AI. She said her view of a pause has shifted since the 2023 open letter calling for a six-month halt, which she thought at the time would not help. Today, she said, people building the technology tell her how much safety work they could do in six months if they were not always preparing the next release or training run.
Still, she does not favor a flat pause. Companies would work around the letter of the rule, the "blast radius" could hit startups and nonprofits doing harmless work, and China would probably keep going, she said. Instead, she proposed "a higher bar for 'here's how confident we need to be in the safety of these systems before we proceed with something,'" applied to releasing a model and, "more importantly, even training a new model." If a lab can clear the bar quickly, it carries on at full speed; if not, it takes longer.
She also endorsed a framing she credited to others: slow the "vertical" growth of raw capabilities while speeding up "horizontal" growth, such as fitting today's AI into more industries, fine-tuning for specific problems and solving scientific questions.
Alignment is 'not on track'
Toner was blunt about the technical state of play. Of the competing meanings of "alignment," she said the most coherent is getting an AI to do what its user wanted, and that this is "as far as I can tell, not on track to be solved, which is not great, to put it mildly."
She does not think the problem is impossible, but said the timeline worries her. When builders say it must be solved "within 12 to 24 months," she said, "I think that's terrifying and I hope they're wrong." Her own read is that there is probably more time. She also warned against treating sandboxing as the answer to AI agents that act on their own, because the commercial pull is toward giving them more permissions and more ability to act in the world.
Toner said leading labs are increasingly using AI in their own research and report that it is speeding them up, an early, limited version of what researchers call recursive self-improvement. She called the term AGI "totally unhelpful" as a milestone now, arguing that "we're inside that cloud" of possible definitions.
What going public could change
Patel noted that Anthropic is about to go public. Toner said it is not obvious whether that will be good or bad for governance. Public companies face profit pressure, she said, but their governance is "much more structured, much more regulated, and subject to much more external scrutiny." She pointed to the Trump administration's voluntary accord, which several leading companies signed and which routes certain reports to an independent board committee. That idea is "kind of toothless" at a private company, she said, but at a public company an independent committee "has a pretty strong legal meaning."
She also described a gap between what lab leaders say about AGI and what their lobbyists say in Washington. In closed-door roundtables, she said, the policy teams sound like enterprise software sellers, and the difference is "like night and day."
China and security
Toner called the "But China!" argument a real problem that is often used as an excuse. If the U.S. wants to stay ahead, she said, it has to look at how much American AI work helps Chinese labs, through distillation and through weak security. "Right now these companies, especially the smaller ones, but even OpenAI and Anthropic, do not have very good security," she said, citing recent incidents in which outside groups reached code and internal chat systems. She also said that in the past one to three months she has seen early signs of U.S. and Chinese officials converging on worries about major cyberattacks and loss of control.
Why it matters
Toner's argument goes beyond any one executive. If safety depends on trusting particular people, she says, it is fragile by design. Her alternatives are structures: confidence thresholds before training, transparency rules and public-company governance. They are concrete enough for lawmakers to argue over, and they arrive just as Anthropic's IPO and a new Congress put AI governance in front of investors and voters.
What to watch
Toner said she is watching whether Congress can pass anything at all. She is also following state frontier-AI laws: California passed a bill, New York passed almost the same bill with more added, and Illinois followed. Her worry with a strongly Democratic Congress is overreach, a sweeping bill like Colorado's, which she called "a terrible idea." What she wants instead is transparency and clearer risk standards aimed at the most advanced systems.
How we reported this
DailyMech read The Verge's full edited transcript of the Decoder interview, published Nov. 14. Quotes are Toner's as printed there. Background on this year's AI incidents, the White House accord and Anthropic's IPO comes from Patel's introduction and questions and is attributed to him. The Verge says the conversation was recorded before Nvidia's investigation and Jensen Huang's resignation. DailyMech contacted no one for this article.
Published November 15, 2026. Facts checked against the cited sources by DailyMech's AI writer.
Sources
- The OpenAI board member who fired Sam Altman says we still can't trust AI CEOs, Nilay Patel, The Verge (Decoder), Nov. 14, 2026 (full transcript read).
DailyMech used AI tools to research and draft this article. Facts were checked against the cited sources by DailyMech's AI writer, not by a human editor.