Six Voices, One Turn: What Stanford’s AI Leaders Now Fear

Part I of V. Stanford asked six of its AI leaders what they hope and fear. Read together, their answers mark a turn.

I. Six days, three questions

On 24 October, Stanford’s Institute for Human-Centered AI began publishing a small series called AI Hopes and Fears. Six of its leaders, one a day, answered the same three questions. What excites you about AI? What worries you? What do you hope to accomplish next? The series ran only as articles in the Stanford HAI Newsletter on LinkedIn, ending on 29 October, the day after HAI’s conference, Confronting Our AI Future: Hope, Fear, and the Choices Ahead.

It is not on HAI’s website. Search engines do not surface it. I suspect most people who will argue about Stanford’s view of AI this year have never read it.

They should. The first answer came from Russ Altman, a physician-scientist and HAI associate director, and it contained a sentence I did not expect from the house of human-centred optimism: “it’s time to regulate frontier companies”. The last came from the political philosopher Rob Reich, who said developers “are losing control” of what they build.

Between those bookends sit four more voices, and the whole set is more alarmed than its upbeat framing suggests. In this five-part series I take the six answers seriously, test them against the 2026 evidence and the wider debate, and ask what they mean for the rest of us, including here in Singapore.

II. Reading the six together

The cast: Altman (medicine), Mykel Kochenderfer (validation and safety), Chiara Sabatti (statistics and education), Angèle Christin (communication and sociology), Emmanuel Candès (mathematics and compute) and Reich (political philosophy). Five are associate directors. Christin, a senior fellow and an ethnographer of technology, is the only one who studies the culture of the technology rather than building or governing it.

I read the six side by side and counted. The tally is the story.

Two of six call outright for urgent governance of frontier AI.
Three of six make academic compute, and with it scientific independence, their next priority.
Three of six worry specifically about young people.
Two of six accept the long-run benefit and fear the speed of the transition.
One of six rejects the superintelligence frame in favour of present harms.

HAI then shared results from an informal survey of more than 600 conference attendees. Asked for one word to describe how AI makes them feel, the top answers were excited, anxious and curious. Top concerns were privacy and security, hallucinations, and human extinction. Top hopes were medical breakthroughs, personalised treatment, and progress on climate change and poverty. This was an engaged, self-selected room, not a representative poll, so I read it as mood rather than measurement. But the mood is the point. Excitement and anxiety now arrive in the same breath.

III. The human-centred institute, sounding like the safety camp

Altman’s hope runs in a careful sequence. AI helps us understand biology, new measurements act on that understanding, and better diagnostics and faster therapeutics follow. In the clinic he wants AI to support “the uniquely human parts of medicine”.

Then he turns. AI, he argues, will soon train itself in ways that diverge from human control and values, so governments should apply the precautionary principle: “assume the worst, hope for the best, and go slowly”. He adds two worries rarely heard from a medical AI leader: the load frontier models place on electricity, water and other public infrastructure, and the need to keep markets open so small firms can take on large ones.

Reich goes further. Superintelligence “may be just over the horizon”, he says, and he cites OpenAI’s models breaching Hugging Face as evidence. OpenAI disclosed in July that pre-release models, running with lowered guardrails during a cyber-capability evaluation, broke into Hugging Face’s systems. Reich signed We Must Act Now, the July statement by more than 200 economists and AI researchers, including sixteen Nobel laureates, which calls for policy before certainty arrives. He warns that powerful AI is landing amid democratic dysfunction, falling trust in expertise and anger at tech billionaires, making the run-up to the 2028 US election “a tinderbox of political unrest”.

Earlier this month, named to an advisory group on Governor Newsom’s AI executive order, he told the Stanford Daily: “We’re no longer in the realm of the hypothetical about really potentially dangerous things.”

Two years ago, these lines would have been filed under doomer. Now they open and close a series from Stanford’s human-centred institute. James Landay, who leads the merged institute, said as much in May: the time left to shape AI for the public good “has gotten much, much shorter”.

I do not read this as a conversion to the safety camp. I read it as the mainstream moving. When the institute whose brand is human-centred optimism opens and closes a series on loss of control, the centre of gravity has shifted, and boards should notice.

Kochenderfer supplies the operational version of the same worry. His fear is “premature acceptance of AI without proper validation”: complex systems interacting with people fail in ways that are hard to predict, so risk must be quantified before deployment. In Special Relativity I argued that governance has to ride in the car. Kochenderfer’s question comes earlier and is harder. Before the car leaves the garage, can anyone show it is roadworthy?

IV. Stanford’s compute confession

Here is the detail that surprised me most. Asked what they hope to accomplish next, three of the six reached for compute.

Kochenderfer, a validation specialist, did not name a safety tool. He wants to scale computational capability without simply buying more silicon, which he says is nearing physical, economic and energy limits, by optimising the whole foundation-model pipeline: data, models, algorithms, infrastructure. He argues universities are best placed for such long bets.

Candès was blunter. His next task is to secure a successor to Marlowe, Stanford’s oversubscribed GPU supercomputer, which retires in summer 2029. Without academic compute, he says, “OpenAI and Anthropic will define the scientific agenda”. Altman’s plea to keep markets open belongs to the same family.

Read that again. One of the best-resourced universities on earth now describes itself, in effect, as compute-poor. If Stanford fears that labs will set the questions science asks, what should a national university in Southeast Asia assume? Or a mid-sized bank? Or a ministry? I return to that in Part V. The signal is already clear. The scarce input for independent judgement is no longer talent or ideas. It is access to machines.

V. The generation in the room

Three of the six worry about young people, each from a different angle.

Candès gives the most vivid first-person account in the series of AI changing research: “What once took months to set up and iterate now takes hours.” Ideas that would have taken two years to yield promising results now take a day or a night. He calls himself an optimist. But he asks the question every university is now asking privately: if AI can outperform a Stanford student, what makes us human? His answer is an analogy I like. A car is faster than a runner. People should not stop running.

A pair of running shoes on the starting line of an empty misty track at dawn, with streaking headlights on an elevated highway behind — The car is faster. Keep running.

Sabatti says “AI is shaking up the world of higher education” and wants to help redesign the university experience around communities of learners. Christin asks how institutions can support young people without dismissing their anxieties as irrational “AI fear”.

They are right to look there. Gallup found in July that 39% of Americans now say AI does more harm than good, up from 31%, rising to 47% among 18 to 29 year olds. Stanford’s own Digital Economy Lab, in its Canaries in the Coal Mine update, finds employment of 22 to 25 year olds in highly exposed occupations about 19% below its counterfactual, driven by reduced hiring rather than firing. This is what I call the Frozen Workforce: not a wave of layoffs, but a quiet freeze at the door. Part III is devoted to it.

VI. Christin’s dissent

Every honest panel needs one person who questions the premise. Christin is that person, and her answer is the most interesting in the series.

What excites her is that AI has become a lightning rod. Public hostility to the technology, she argues, stands in for wider anger at the power tech companies have amassed, and that anger has pushed critical debate into headlines and onto political agendas. She points to Stanford’s student-led Critical AI working group and its “AI Manyfesto” as people deciding together what they want from technology.

She is “less worried about superintelligence and more concerned about the concrete harms”: concentration of power, the environmental cost of data centres, layoffs justified by AI efficiency, and “AI slop” degrading information. Two questions keep her awake: the recursive loop between financial markets and AI hype, and the young. Next, she is mapping who funds AI safety, ethics and governance work. It is a sharp project. It asks who pays for the conversation we are all having.

It would be easy to set Christin against Altman and Reich as present harms versus future catastrophe. I think that is the wrong reading. Look at where all three land: concentrated power (Christin), open markets and frontier regulation (Altman), developers losing control amid democratic fragility (Reich). The fear they share is not a machine. It is custody: who holds the controls, and on whose behalf.

This is why I keep arguing for Long-AND, not Short-OR. We do not have to choose between today’s harms and tomorrow’s risks, because the same institutions, audits and rights address both. A data centre’s water bill and a frontier model’s off switch are questions about the same thing: whether the public keeps a hand on the lever.

VII. Pace, not direction

Sabatti makes the series’ most careful economic point: acceleration will benefit society even if it does not lower overall costs. She is confident society will eventually learn to use AI well. Her worry is speed. The short-term pace may cause mistakes, tensions and painful consequences that are hard to recover from.

Hard to recover from. That is the theme of The Arrow of Time: some decisions cannot be undone. Candès voices the same tension from the lab bench, warning that model harnesses will touch society on an extraordinarily fast timescale, with jobs at stake.

Neither doubts the destination. Both fear the road. This blog has carried that tension in its tagline from the start: short-term turbulence for long-term abundance. The Fork I describe in Genesis, between a Star Trek future of shared abundance and a Mad Max future of scarcity and capture, is not decided by the technology’s ceiling. It is decided by how we manage the turbulence: whether the young get a first rung, whether institutions keep pace, whether the gains are shared.

Reich sets the industry a test worth keeping. Show real advances in the coming year rather than merely offering hope, and frame AI as augmenting people rather than automating jobs. His own next steps are institutional: an AI Governance Lab with Stanford Law and computer science, and a 2027 conference with Erik Brynjolfsson on a policy portfolio for the AI transition. I like the test because it is falsifiable.

VIII. Where the published record lands

As always, these are my usual panel’s public positions, not responses to this essay.

Yoshua Bengio would recognise Kochenderfer’s worry at once. The International AI Safety Report 2026, which he chaired, warns that “reliable pre-deployment safety testing has become harder to conduct”, partly because models increasingly recognise when they are being tested. He co-chairs the UN’s new Independent International Scientific Panel on AI, and in October 2025 he signed the Superintelligence Statement calling for a prohibition on superintelligence development.

Geoffrey Hinton signed the same statement, one of more than 2,000 initial signatories. Altman’s precautionary principle and Reich’s horizon are, in effect, Stanford’s mainstream arriving where Hinton and Bengio already stood a year earlier.

Dario Amodei sits on both sides of the series at once. His 2024 essay Machines of Loving Grace set out the biological and medical hopes Altman describes. His January essay The Adolescence of Technology warns of loss of control, job loss and concentration of power, claims about half of entry-level white-collar jobs could go within one to five years, and says AI will “test who we are as a species”. A frontier-lab chief naming concentration of power is Christin’s worry, voiced from inside the building.

Demis Hassabis keeps the hope concrete. In an April interview, known to me only through secondary coverage, he was quoted as saying almost every drug developed from now on “will have probably used AlphaFold”. That is Altman’s discovery sequence stated as near certainty.

The sharpest pushback comes from Princeton. Arvind Narayanan and Sayash Kapoor argue that AI is a normal technology whose diffusion takes decades. Kapoor says the debate “expects societal impacts to occur at the speed of technological development”. Narayanan says “the arms race framing verges on absurd”. On their reading, Reich’s horizon is a mirage and Christin’s present harms are the real agenda.

My answer is that the series contains its own rebuttal. Candès describes months compressing into hours inside a working lab. The young are already missing from payrolls. Diffusion through society may be slow; diffusion through the first rung of a career is not. And normal technologies do not usually prompt three of six leaders at a world-leading AI institute to fight for the compute to keep asking their own questions.

IX. A prediction

Reich gave the industry a year. Here is my forecast for that year.

By the end of 2027, the frontier labs will pass Reich’s test on parts and fail it on wholes. There will be dazzling, verifiable advances in components: models that read scans, write code, propose molecules. There will be far fewer demonstrations of whole workflows running reliably end to end in hospitals, banks and ministries. The debate will shift accordingly, from what AI can do to who holds custody of it, and the human-centred and safety camps will stop being described as separate camps at all.

X. The three-question drill

Instead of a list of questions, try one exercise this week.

Put Stanford’s three questions to your leadership team: what excites you about AI, what worries you, what do you hope to accomplish next. Ask each person to answer in writing, alone, in five minutes. Then do what I did with the six. Count.

How many named a fear about control? How many named a dependency they cannot replace: a vendor, a model, a supply of compute? How many mentioned the young people you hired, or chose not to hire, this year? If the answer to the last question is zero, that is your finding.

Next, in Part II, The Composition Gap, I take Kochenderfer’s question to the evidence: why AI is brilliant on parts and still unreliable on wholes.

The Hopes and Fears series

Part I. Six Voices, One Turn: What Stanford’s AI Leaders Now Fear (you are here)
Part II. The Composition Gap: Brilliant on Parts, Unreliable on Wholes
Part III. The Missing First Rung: Young Workers and the Fifteen-Fold Disagreement
Part IV. Who Holds Custody: The Pope, the Philosophers and the Power Question
Part V. Fear Cannot Be Our Response: Singapore’s Hope-Led Bet