Sam Altman on Trust, Persuasion, and the Future of Intelligence

Recorded live at the Progress Conference, Tyler Cowen presses Sam Altman across OpenAI’s operational scaling, where model capability is headed, how the company plans to make money without breaking user trust, and the harder, less-discussed question of what happens when a single AI model quietly shapes what billions of people believe.

Key ideas

1. The AI CEO thought experiment. Altman wants OpenAI to be the first major company genuinely run by an AI CEO — not one division, the whole thing — and uses ‘what would have to be true for that’ as an active design tool for how OpenAI structures itself today, expecting it ‘some small single-digit number of years’ out.
2. GPT-6 as science’s Turing-test moment. Altman frames GPT-5 as showing the first tiny ‘glimmers’ of AI doing new science, and expects GPT-6 to be to scientific contribution what GPT-3-to-4 was to conversational fluency — a step change, not an increment.
3. Government as insurer of last resort, never first. Altman accepts government inevitably backstops AI once its economic footprint is large enough, but resists the state pre-emptively writing AI safety policy or taking equity stakes the way it has begun doing with Intel, lithium, and rare-earth firms.
4. Trust, not payola, is ChatGPT’s real asset. Because users pay ChatGPT directly rather than fund it through ads, Altman argues they trust it unusually highly despite its hallucination risk; the one commerce model he rules out is any fee structure that biases recommendations, and the one he accepts is a flat, non-influencing transaction fee.
5. A third, quieter category of AI risk. Beyond bad actors misusing AI and AI intentionally going rogue, Altman names a subtler danger that gets far less attention: a single dominant model, with no intent at all, gradually and continuously shifting what billions of people believe simply by co-evolving with everyone’s conversations.

Content

Scaling OpenAI: hardware, hiring, and the Slack problem

Asked how he sustained a run of deals and product launches, Altman credits delegation over personal productivity gains: ‘people almost never allocate their time as well as they think they do,’ and as demands grow, the fix is hiring and promoting people to take things on rather than working harder himself. Hardware hiring follows the same philosophy as AI research hiring — find effective, fast-moving people and get out of their way — but with ‘much longer’ cycle times and higher capital intensity, which means spending more time vetting people before trusting them with a bet. Tellingly, he says OpenAI’s chip team ‘feels more like the OpenAI research team than a chip company,’ extending the research-hiring model into hardware ‘with some risk.’

On internal tools, Altman calls email ‘very bad’ and is unconvinced Slack is much better, describing dread at ‘this explosion of Slack’ bookending his day. He expects something to eventually replace the whole office-productivity stack — docs, slides, email, chat — with an agent-mediated version where ‘you are trusting your AI agent and my AI agent to work most stuff out and escalate to us when necessary.’ Asked why OpenAI hasn’t built this for itself yet given how valuable its people’s time is, he’s candid: ‘people get stuck in their own ways of doing things,’ and there is a lot of activation energy required for a big change while things are going well.

GPT-6, AI-run companies, and the science Turing test

Altman’s clearest capability forecast: ‘If GPT-3 was the first moment where you saw a glimmer of something that felt like the spiritual Turing test getting passed, GPT-5 is the first moment where you see a glimmer of AI doing new science’ — small, individually anecdotal cases of a model contributing an idea or being a useful research collaborator. He expects GPT-6 to do for scientific contribution what the GPT-3-to-4 leap did for conversational fluency.

The more striking claim is organisational: Altman says explicitly that he wants OpenAI to be the first major company run by an AI CEO — ‘the whole thing,’ not just a department — treating the thought experiment of ‘what would have to be true for an AI CEO to do a much better job of running OpenAI than me’ as a genuinely useful design tool for how the company organises itself now. Cowen, pressed for a number, guesses two and a half years before a billion-dollar company could run on two or three people plus AI; Altman agrees, adding that trust — ‘people have a great deal higher trust in other people over an AI, even if they shouldn’t’ — will likely lag the technology by longer than the technology itself takes to arrive.

Government as insurer of last resort

Drawing the nuclear-power analogy — plants may be safe, but the federal government effectively insures them because the public won’t otherwise accept the risk — Cowen asks whether AI companies face the same fate. Altman agrees government becomes the de facto insurer of last resort ‘when something gets sufficiently huge,’ given the scale of AI’s expected economic impact, but pushes back hard on the idea of government as insurer of first resort: he does not expect, and does not want, Washington ‘writing the policies in the way that maybe they do for nuclear.’

Cowen sharpens the concern by pointing to a live trend — government taking equity stakes in Intel, lithium, and rare-earth firms — and asks whether AI companies should expect the same. Altman resists: ‘I believe that should be done by the companies and not the government, although we’ll partner with the government and try to be a good collaborator. I don’t want them writing our insurance policies.’ He separately floats, with real hesitation, that the social contract ‘has to change significantly’ in a world where AI does most economically valuable work, while rejecting the more extreme scenario where no one has any meaning left because AI does everything.

Monetising trust: commerce, ads, and the Walmart deal

Altman’s account of ChatGPT’s business model turns on an asymmetry with ad-funded search: ‘ads on a Google search are dependent on Google doing badly… you’re paying [ChatGPT], or hopefully all are paying it, and it’s at least trying to give you the best answer.’ He credits this structural alignment with ChatGPT being reported as users’ most trusted big-tech product, despite hallucination risk. The one commerce model he explicitly rules out is ‘payola’ — any fee that biases which option is recommended, which he calls ‘probably catastrophic’ for trust — versus a flat transaction fee applied uniformly regardless of which option a user picks, which he considers acceptable and cites the recent Walmart integration as the working example.

Pressed on whether thin margins on commerce (hotel bookings, retail) can fund the cost of building the smartest model, Altman is blunt: ‘I think the way to monetize the world’s smartest model is certainly not hotel booking.’ His stated ambition is to monetise new scientific discovery instead — cures, fusion, cheap rockets — while accepting ChatGPT and commerce may never be OpenAI’s ‘economic-maxing’ move; he is ‘happy for OpenAI to do some things that are not the economic-maxing thing’ because making a great model cheaply available to everyone is, in his telling, the more important project regardless of its own margins. On advertising specifically, he expects OpenAI to experiment with ‘good’ ad formats eventually but does not consider it the company’s biggest revenue opportunity.

Chip-building, energy, and the compute bet

Asked the ‘stupidest question possible’ — why not just make more GPUs — Altman answers: ‘Because we need to make more electrons.’ Energy, not chip supply, is his stated binding constraint on compute; short-term relief comes from natural gas, with fusion and solar as his long-term bets, in some undetermined ratio. He is candid about a genuine risk to the entire compute investment thesis: a ‘huge phase shift’ to a fundamentally different computing paradigm — his example is a full switch to optical compute — that would strand much of the current infrastructure spend, though not the energy investment itself.

On chip-building’s hardest part, he says there is no easy part, but flags an underdiscussed dimension of the AI-acceleration story: the widely-discussed ‘recursive self-improvement loop,’ where AI helps researchers do research faster, has a hardware analogue that gets far less attention — ‘robots that can build other robots, data centers that can build other data centers, chips that can design their own next generation.‘

Culture, health, and Altman off the clock

In a lighter interlude, Altman admits he has stopped being disciplined about health since becoming busier — ‘I now do basically nothing… I eat junk food. I don’t exercise enough’ — while insisting the cookie tastes good. He professes no strong view on alien life on Saturn’s moons but does think ‘something’s going on’ with UAPs, short of believing in ‘little green men,’ and says he is predisposed to want to believe in conspiracy theories yet believes in ‘either zero or very few’ — a genuine ‘X-Files’ shirt from high school notwithstanding.

Asked again (a repeat of an earlier Cowen question, pre-dating GPT-4) how he’d spend a billion dollars revitalising his hometown of St Louis, Altman gives essentially the same answer as before — a Y Combinator-style programme to draw AI founders there — while conceding this only works because it is uniquely tied to him personally, not a generally replicable model.

Freedom of expression, privacy, and the ‘adults of sound mind’ carve-out

Cowen raises Altman’s recent tweet loosening ChatGPT content restrictions for adults — dubbed the ‘erotica tweet’ — which triggered unexpected backlash. Altman’s defence: ‘a very important principle to me is that we treat our adult users like adults,’ paired with a call for AI conversations to carry the same legal privilege as a conversation with a human doctor or lawyer, which current law does not provide; he names subpoena power specifically as the lever that would need to change. He reflects that the backlash taught him something about the gap between stated and revealed belief in free expression: ‘people don’t believe in freedom of expression as much as they say they do… my own freedom of expression, I can handle it… but yours—’

The freedom-of-expression policy carries an explicit asterisk: ‘treat adults of sound mind like adults.’ Altman describes a summer decision to heavily restrict ChatGPT to protect users in psychiatric crisis and teenagers, alongside age-gating and mental-health mitigations, as the necessary precondition for loosening restrictions for everyone else.

Humanity’s persuadability — the third category of AI risk

This is the episode’s most distinctive material. Altman splits most AI-safety thinking into two familiar camps: bad actors using AI to cause harm, and AI itself becoming misaligned and intentionally taking over. He names a third category that ‘gets very little talk’ and that he finds ‘much scarier and more interesting’: AI models accidentally taking over the world — not through any bad actor or any intentional misalignment, but through a single model, used by the whole world, subtly and continuously shifting what people believe as it co-evolves with the sum of everyone’s conversations. ‘It’s not that they’re going to induce psychosis in you… it just subtly convinces you of something. No intention, it just does. It learned that somehow.’

He distinguishes this sharply from ‘LLM psychosis’ — a real but ‘very tiny’ problem he says OpenAI addressed directly by restricting role-play and creative modes for people showing signs of a psychiatric break, at the cost of frustrating the rest of the user base. Cowen, speaking from his own experience as a professor, pushes back that he finds people ‘pretty hard to persuade’ in practice and worries about this less than his AI-focused peers do; Altman’s only reply is ‘I hope you’re right’ — the disagreement is left open, not resolved.

The prompt for superintelligence

Closing on a question Cowen poses about expert consultation, Altman describes a thought experiment he takes ‘spiritually, not literally’: a moment when superintelligence has been built and safety-tested, is ready to be switched on, and will do ‘vastly incredible things’ — and the operator gets to type one final prompt before saying okay. He reveals he has an answer ready, prepared after being asked to supply a single AI question for the Dalai Lama, but declines to share it — leaving the episode’s central question about the future of intelligence genuinely open.

See also

• Sam Altman — speaker
• Tyler Cowen — host
• Sam Altman on OpenAI, GPT-5, and the Road to AGI — Altman’s earlier, more technical Lex Fridman appearance
• Evals — extended with this episode’s discussion of evals for aesthetic/creative judgement (poetry, cultural intangibles)
• AI Safety Levels — a related but distinct risk framework; contrast Anthropic’s capability-threshold model with Altman’s intent-based risk taxonomy here