The OpenAI CEO says he isn't losing sleep over rogue robots — but the defaults, guardrails, and choices that ripple across hundreds of millions
Al Drago/Bloomberg via Getty Images
Sam Altman says he hasn’t slept well since ChatGPT launched — not because of sci-fi nightmares, but because of product choices that scale to the size of a civilization. “I haven’t had a good night of sleep since ChatGPT launched,” the OpenAI CEO told Tucker Carlson, adding that what keeps him awake are “the very small decisions” about model behavior that, when multiplied across a massive user base, make “the net impact … big.”
That scale is why Altman draws a hard line between what users sometimes perceive and what the systems actually are. Asked whether AI is “alive” or lying, he said the models “hallucinate,” not premeditate (so: a statistical failure mode, not intent) — and that OpenAI has been working with the models to cut down on their falsehoods.
“We’ve already made … in the GPT-5 era a huge amount of progress toward that,” he said, while acknowledging there are still examples of this problem. The point, in his telling, is that better guardrails beat grand theories.
Those guardrails increasingly live on paper. Altman pointed to OpenAI’s “model spec,” the publicly posted playbook for how ChatGPT should behave when the questions turn moral, political, or simply risky for “the rules we’d like the model to follow.” He added that OpenAI “consulted like hundreds of moral philosophers” but ultimately has to make calls — and be accountable for them. Altman says, though, accountability should start at the top: “The person I think you should hold accountable for those calls is me.” OpenAI published a first draft of the “model spec” in 2024 and released a major update this year.
Codifying behavior inside the product is one pillar; pushing for a law is another. Altman said that, if he could pass one policy now, it would be about “AI privilege” — a legal protection that would treat certain chatbot conversations like doctor-patient or attorney-client confidential exchanges so that governments can’t easily subpoena user chats, which can happen today under current law.
“When you talk to a doctor about your health or a lawyer about your legal problems, the government cannot get that information… I think we should have the same concept for AI,” he said. “Right now, we don’t have that.” That idea tracks with months of Altman’s public comments pressing Washington to recognize AI-“client” confidentiality.
Scale is the multiplier. Altman said “hundreds of millions of people” talk to OpenAI’s models daily, which is why defaults matter more than hypotheticals. That includes crisis content: he described tightening responses around self-harm while still wrestling with jurisdiction-specific questions (for instance, countries that permit assisted dying). The through-line is an attempt to reflect a broad user consensus within bright legal lines, with customization at the edges.
Altman said he doesn’t think ChatGPT should be “for” or “against” contested questions so much as it should reflect a broad, evolving user consensus within clear legal bounds. “I don’t… impute my exact moral view,” he said. “What I think ChatGPT should do is reflect that… collective moral view,” while allowing customization and still drawing bright red lines where harm is obvious.
The Carlson interview didn’t just cover product defaults; the two also talked about culture, governance, and rivalry. Altman said people are picking up on ChatGPT’s style — the phrasing, cadence, even punctuation — in real writing. He warned about “unknown unknowns,” from small linguistic drift to more serious risks such as misuse of AI in biological research, disinformation, or cyberattacks.
On deepfakes and authenticity, Altman said he’s against mandating biometrics, instead promoting alternatives such as cryptographic signatures or code words to verify who’s speaking or sending what. And on defense, he ruled out building “killer attack drones” but said he suspects “there’s a lot of people in the military talking to ChatGPT for advice,” a gray area he hasn’t resolved. “I don’t know exactly how to feel about that,” he said.
Altman weighed in, too, on economics and governance. Customer support roles, he said, are the most obviously at risk from automation, while nurses are likely to remain indispensable and programmers face an uncertain middle ground. He framed that churn in historical terms, saying that about half of jobs significantly change every 75 years, though AI could compress that cycle into a sharper burst. The Carlson interview veered into other charged territories — including a contentious exchange about the death of a former OpenAI researcher.
And Altman also talked about his relationship with Tesla $TSLA CEO Elon Musk. The OpenAI CEO said that, for a long time, he “looked up to [Musk] as an incredible hero, a great jewel for humanity,” but now his feelings are more mixed — “different now.” Musk was part of OpenAI’s founding group; now, he runs xAI and has sued OpenAI, accusing it of abandoning its initial mission (and calling it “Closed AI”). Altman, meanwhile, is planning a brain-implant startup to rival Musk’s Neuralink.
For better or worse, since launching on Nov. 30, 2022, ChatGPT is now embedded in work, school, and daily life — and Altman is claiming that the real risk lies in the defaults, not the doomsday. There’s no off switch for a system that fields millions of queries before dawn. For OpenAI’s CEO, that endless hum means every tweak, refusal, or hallucination is magnified in ways no lab test can capture.
Altman can insist that he isn’t haunted by visions of rogue robots — that what keeps him awake are the subtler nightmares of defaults and disclosure. In a world where billions of prompts add up to cultural shifts, the sleep he craves is really shorthand for stability. Sweet dreams, in this case, would mean boring settings, clear boundaries, and a legal framework strong enough to stop the midnight second-guessing.
Join 500,000+ readers who start their day with Quartz.
By subscribing, you agree to our Terms of Service and Privacy Policy.
