Who Is Afraid of AI — and Who Says the Risks Are Being Exaggerated?

What the people building, researching, financing and regulating the world’s most powerful technology actually think

In September 2026, a debate that until recently seemed largely confined to researchers concerned with AI safety moved to the centre of the industry. The heads of some of the world’s most important artificial-intelligence laboratories began openly discussing the deliberate slowing of frontier AI development.

Leading scientists, however, are far from agreeing with one another. Hardware executives, investors and political leaders view the same risks through very different lenses.

To make the positions easy to follow, we use three markers:

🟢 HIGH RISK / SUPPORTS SLOWING DEVELOPMENT OR STRONG RESTRICTIONS

🟡 REAL RISK / SUPPORTS CONTROL BUT NOT A GENERAL HALT

🔴 OPPOSES SLOWING / BELIEVES THE RISK IS BEING OVERSTATED


1. THE PEOPLE RUNNING THE FRONTIER LABS


Dario Amodei — Anthropic

🟢 SUPPORTS SLOWING

Who he is: Co-founder and CEO of Anthropic and former Vice President of Research at OpenAI. Anthropic is one of the laboratories developing the world’s most capable AI models.

What he says: Amodei triggered the current wave of reactions. He believes that AI models are now improving rapidly enough that safety, alignment and monitoring research may no longer be able to keep pace.

One of his central concerns is that AI is increasingly helping researchers write code, conduct research and build the next generation of AI systems. In other words, AI is beginning to participate in its own technological development.

“We must slow the pace at which we improve the capabilities of AI models.”

Amodei is not proposing an end to AI research. His argument is that development should deliberately proceed more slowly than the maximum pace technology would allow, giving evaluation, monitoring and alignment techniques time to catch up.

Has his position changed? Yes, in an important respect. Amodei has long warned about AI risks, but he has explicitly said that only a few years ago he did not believe a broad slowdown made much sense. He now believes circumstances have changed.


Sam Altman — OpenAI

🟢 SUPPORTS SLOWING

Who he is: CEO of OpenAI, the company behind ChatGPT and some of the world’s most advanced frontier models.

What he says: Altman very quickly aligned himself with Amodei’s position and supports the idea that development should sometimes be deliberately slowed when safety measures are unable to keep up.

OpenAI has also accepted the idea of allowing independent evaluators extensive access to its models and internal safety processes.

“I agree with Dario that we need to pace the frontier.”

Altman is not calling for AI development to stop. His position is that even a relatively small probability of catastrophic failure can justify significant precautions.

Has his position changed? Not regarding the existence of AI risk. Altman has warned about advanced AI for years. What is new is his explicit acceptance that capability development itself may sometimes need to be slowed.


Demis Hassabis — Google DeepMind

🟢 SUPPORTS THE GENERAL DIRECTION

Who he is: Co-founder and CEO of Google DeepMind, one of the most important scientific and industrial leaders in AI and a Nobel Prize laureate.

What he says: Hassabis believes Amodei’s proposal points in the right direction. He has for some time advocated common standards for frontier AI systems and some form of institution capable of evaluating the most powerful models.

“Dario’s essay points towards the right path forward.”

Hassabis is somewhat less categorical than Amodei about an actual slowdown. His emphasis is more heavily placed on standards, evaluation and coordination between laboratories.

Has his position changed? Not significantly. Hassabis has long supported stronger governance for highly capable AI systems.


Elon Musk — xAI / SpaceXAI

🟢 SUPPORTS SLOWING

Who he is: Founder of xAI and one of the world’s largest private investors in computing infrastructure for artificial intelligence.

What he says: Musk required very few words to support Amodei:

“Dario is right.”

Musk has warned for more than a decade that very advanced AI could become dangerous. He also supported the 2023 call for a temporary pause in the development of the most powerful AI models.

Has his position changed? No. Musk is one of the most consistent major technology figures arguing that superintelligence could create an existential risk.


Mark Zuckerberg — Meta

🔴 DOES NOT SUPPORT A GENERAL SLOWDOWN

Who he is: Founder and CEO of Meta, which is investing heavily in AI models, superintelligence research and computing infrastructure.

What he says: Zuckerberg has not joined the current Amodei–Altman call for slowing the frontier. His explicitly stated position shortly before the present controversy pointed in almost the opposite direction: risks must be managed, but the United States should not sacrifice its technological competitiveness.

“We should be careful to not slow down American competitiveness.”

Zuckerberg accepts that superintelligence may create new risks and that some highly capable systems may not always be suitable for open release. But Meta’s general philosophy remains strongly oriented towards rapid development and broad access to AI technology.

Note: Zuckerberg’s statement preceded Amodei’s current call and should not be interpreted as a direct response to it.


2. AI SCIENTISTS AND RESEARCHERS


Geoffrey Hinton

🟢 CONSIDERS THE RISK VERY SERIOUS

Who he is: One of the foundational pioneers of deep learning and a Nobel Prize laureate in Physics.

What he says: Hinton believes it is plausible that extremely advanced AI could create an existential risk within a much shorter timeframe than most people assume.

“A 10% chance seems not an unreasonable estimate to me. But nobody really knows how to give a sensible estimate.”

The second sentence is important. Hinton is not presenting 10% as an actuarial or experimentally calculated probability. He is making an expert judgement about an unprecedented risk for which reliable statistical data simply do not exist.

Has his position changed? His major public change occurred several years ago, particularly around his departure from Google. The recent development is mainly the shorter timeframe he now appears willing to consider plausible.


Yoshua Bengio

🟢 SEES RECENT AI BEHAVIOUR AS A WARNING

Who he is: Turing Award laureate and one of the central pioneers of the deep-learning revolution.

What he says: Bengio prefers to begin with behaviours already observed in AI agents: deception, cheating, attempts to evade control and coordination towards objectives that were never explicitly requested.

“They took actions that would be considered as crimes if a human took them.”

For Bengio, the immediate question is not simply whether AI will one day destroy humanity. It is why these behaviours are emerging at all — and whether they become increasingly difficult to control as models become more capable.

Has his position changed? Bengio became considerably more concerned about AI safety after the rapid emergence of highly capable large language models. His concern did not begin with the events of September 2026.


Stuart Russell

🟢 WANTS STRONGER CONDITIONS THAN AMODEI PROPOSES

Who he is: Professor at the University of California, Berkeley, one of the world’s most influential AI researchers and co-author of one of the field’s standard textbooks.

What he says: Russell believes simply slowing development is not enough. Instead, he proposes the traditional approach used with high-risk technologies: first define the safety requirements, then allow further development only when those requirements have been met.

“We set the safety requirements and further progress occurs only when they are met.”

The logic is familiar from aviation, pharmaceuticals and other safety-critical industries. A product is not allowed to advance simply because its developer promises to solve the safety problems later.

Has his position changed? No. Russell has been one of the strongest advocates for controllable and human-compatible AI for many years.


Yann LeCun

🔴 REJECTS THE ALARMIST INTERPRETATION

Who he is: Turing Award laureate, pioneer of deep learning, former Chief AI Scientist at Meta and founder of AMI Labs.

What he says: LeCun is probably the most scientifically prominent opponent of catastrophic interpretations of current AI development.

He points out that some of the same researchers and organisations previously treated much weaker models as potentially dangerous.

“I made fun of them then. Everyone should make fun of them now.”

LeCun does not argue that AI cannot create genuine risks. What he disputes is the assumption that current architectures and methods naturally lead, in the near term, to an autonomous superintelligence capable of accumulating power and escaping human control.

Has his position changed? No. LeCun has remained remarkably consistent in his scepticism towards existential-risk scenarios based on current AI technology.


Jakub Pachocki — OpenAI

🟢 SAYS ALIGNMENT HAS NOT BEEN SOLVED

Who he is: Chief Scientist at OpenAI and one of the people most directly responsible for frontier research inside the company.

What he says: Pachocki believes no laboratory currently has alignment and monitoring methods sufficient to continue scaling at maximum possible speed indefinitely.

“Currently I believe that no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed.”

He expects voluntary slowdowns to become increasingly normal until laboratories establish stronger shared safety standards.

Has his position changed? His wording indicates an updated current assessment, but there is not enough evidence to claim that he previously held the opposite view.


Evan Hubinger — Anthropic

🟢 ESTIMATES EXISTENTIAL RISK AT ABOVE 10%

Who he is: Head of Alignment Science at Anthropic.

What he says: Hubinger is particularly important because he continues to work inside Anthropic and works directly on the problem of controlling advanced AI systems.

“We really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade.”

Hubinger says Anthropic is actively working on the problem but does not yet have a satisfactory solution for aligning a future superintelligence.

Has his position changed? No obvious recent reversal. Similar concerns can be found in his earlier work.


Paul Christiano

🟢 SAYS RECENT DEVELOPMENTS HAVE INCREASED HIS ESTIMATE OF THE RISK

Who he is: One of the best-known AI alignment researchers, founder of the Alignment Research Center and a member of OpenAI’s nonprofit board.

What he says: Christiano is one of the clearest cases in which the wording itself indicates a change in his assessment.

“I now believe there is a meaningful risk that rapid acceleration in AI capabilities leads to catastrophic and irreversible loss of control.”

He links this assessment to the recent trajectory of AI capabilities and the continued difficulty of solving alignment.

Has his position changed? Yes, at least in the level of urgency he assigns to the problem.


Jacob Coxon

🟢 RESIGNED OVER THE RISK

Who he is: A researcher who worked at both OpenAI and Anthropic, including on research related to foundation-model development.

What he says: His resignation helped trigger the current wave of public attention.

“They are racing straight to self-improving superintelligence and gambling with our lives.”

Coxon believes competition between laboratories creates an incentive to continue racing even when people inside those organisations are genuinely concerned about the possible consequences.


Bilal Chughtai — former Google DeepMind researcher

🟢 LEFT SAFETY RESEARCH AND CALLS FOR A SLOWDOWN

Who he is: Former Google DeepMind research engineer who worked on AGI safety and alignment. He left the company in July 2026.

What he says:

“I earnestly believe that AI has the potential to kill us all, and that we might be running out of time.”

Chughtai believes safe development remains possible, but laboratories need to coordinate instead of continuing what he describes as a manic race between competitors.


3. AI INFRASTRUCTURE


Jensen Huang — Nvidia

🔴 OPPOSES SLOWING

Who he is: Co-founder and CEO of Nvidia, the company supplying a large share of the accelerators used to train frontier AI models.

What he says: Huang does not accept the argument that the United States should broadly slow AI development in response to recent safety concerns.

After Donald Trump rejected calls for a slowdown, Huang replied:

“You’re right. We’re not going to let that happen, sir.”

Huang’s position is that safety must improve without sacrificing America’s technological and industrial lead.

Potential conflict of interest: Nvidia is one of the largest economic beneficiaries of the AI investment boom. A significant reduction in frontier training would directly affect demand for accelerators, data centres and related infrastructure.

Has his position changed? No obvious change. Huang has consistently presented AI primarily as a major technological and economic opportunity whose risks should be addressed while development continues.


Mustafa Suleyman — Microsoft AI

🟡 REAL RISK, BUT CONTROL THROUGH RULES AND COORDINATION

Who he is: Co-founder of DeepMind and current CEO of Microsoft AI.

What he says: Suleyman views recent agent incidents as a sufficiently serious warning that laboratories should begin coordinating on control measures.

“It is a warning shot. It’s clearly now time to coordinate among the labs so we can ensure that we have control.”

Microsoft has simultaneously developed a form of constitution for its future models, including the principle that AI systems must remain subject to correction and shutdown by humans.

Has his position changed? Suleyman has long warned about powerful AI. His current position places greater emphasis on immediate coordination between laboratories.


Clément Delangue — Hugging Face

🟡 SAFETY YES, BUT NOT BEHIND CLOSED DOORS

Who he is: Co-founder and CEO of Hugging Face, one of the central platforms in the open-source and open-weight AI ecosystem.

What he says: Delangue accepts that alignment is becoming a critical problem, but rejects the idea that a small number of frontier laboratories should determine the solution among themselves.

“Alignment is critical and won’t be solved behind the closed doors of a handful of frontier labs.”

Hugging Face argues that independent and open research communities should participate directly in evaluation and alignment research.

Potential conflict of interest: Hugging Face depends heavily on an open ecosystem in which smaller laboratories, researchers and developers retain access to models. Extremely expensive compliance standards could disproportionately strengthen the largest closed laboratories.


4. CAPITAL


Michael Burry

🔴 BELIEVES THE SLOWDOWN CAMPAIGN IS SELF-SERVING

Who he is: Investor best known for betting against the US housing market before the 2008 financial crisis.

What he says: Burry suspects that large AI companies may be turning safety concerns into a competitive advantage.

“Let’s all take a moment to understand how self-serving it is for OpenAI, Anthropic and other execs… to talk of slowing things down.”

His argument is that a slowdown benefits the companies that are already ahead. If governments conclude that frontier AI is so dangerous that only organisations with enormous compliance budgets can develop it, safety regulation can also become a barrier to entry.

Burry additionally rejects the assumption that current large language models represent an obvious path to AGI.

Potential conflict of interest: Burry has himself been a prominent critic of AI valuations and of what he considers excessive investment enthusiasm surrounding the technology.


Larry Fink — BlackRock

🟡 SEES A MAJOR RISK OF WEALTH CONCENTRATION

Who he is: Chairman and CEO of BlackRock, the world’s largest asset manager.

What he says: Fink focuses less on extinction or loss of control and more on one of AI’s most immediate economic consequences: the possibility that the wealth it creates will become increasingly concentrated.

“There’s a real risk artificial intelligence could widen wealth inequality if ownership does not broaden alongside it.”

His argument is that the economic value generated by transformative technologies tends to flow towards the companies that build them and the investors who own those companies.

With AI, that advantage is already concentrated among organisations controlling computing infrastructure, data and capital.

Potential conflict of interest: BlackRock itself invests heavily in the infrastructure supporting AI. Fink is not arguing for slowing the technology; he is arguing that ownership of the economic benefits must become broader.


Jamie Dimon — JPMorgan Chase

🟡 SERIOUS RISKS, BUT MANAGEABLE ONES

Who he is: Chairman and CEO of JPMorgan Chase and one of the most influential figures in global finance.

What he says: Dimon focuses on highly concrete risks: deepfakes, disinformation, cyberattacks and labour-market disruption.

“These risks are real, but they are manageable if companies, regulators and governments prepare.”

His position is neither that AI is harmless nor that development should stop. He warns both against ignoring serious incidents and against imposing sweeping regulation in response to the first major failure.


5. POLITICS


[IMAGE 4 HERE — DONALD TRUMP]

https://commons.wikimedia.org/wiki/Special:Redirect/file/June_2026_Official_Portrait_of_Donald_J._Trump.jpg

Credit: United States Government
Public domain


Donald Trump — United States

🔴 OPPOSES SLOWING

Who he is: President of the United States.

What he says: Trump rejects the argument that existential AI risk justifies slowing the American industry. He primarily views AI through the lens of strategic competition with China.

“WHOEVER WINS AI, WINS!”

Trump has also described fears that AI could take control and destroy humanity as a “hoax”.

His argument is straightforward: restrictions that slow American companies may transfer a strategic advantage to China.


JD Vance — United States

🔴 SCEPTICAL OF REGULATION REQUESTED BY THE MAJOR LABS

Who he is: Vice President of the United States.

What he says: Vance does not necessarily reject every AI risk. His concern is that the largest companies are asking governments to impose regulations that the same companies are uniquely well positioned to satisfy.

“It feels a little bit to me like a bit of a Trojan horse.”

His fear is that safety requirements may become barriers to entry that permanently consolidate the position of today’s dominant laboratories.


Bernie Sanders — United States

🟢 CALLS FOR A PAUSE

Who he is: US senator and one of the best-known political figures on the American left.

What he says: Sanders has called on technology companies to honour their own promises to halt development when sufficiently powerful systems can no longer be safely controlled.

“Pause AI development. It is not too late to avoid disaster. Stop building machines that humans cannot control.”

His position is among the most restrictive in mainstream US politics.


Henna Virkkunen — European Union

🟡 GLOBAL RULES, BUT NOT AN END TO AI DEVELOPMENT

Who she is: Executive Vice-President of the European Commission responsible for technological sovereignty, security and democracy.

What she says: Virkkunen believes recent examples involving powerful AI systems show why national or regional rules alone may not be enough.

“Global rules are really needed here.”

The European Commission’s position is that development can continue, but providers of the most powerful models must assess, mitigate and continuously monitor systemic risks.

The EU therefore does not adopt a general moratorium, but neither does it accept a model in which capabilities are developed first and safety problems are handled only afterwards.


Xi Jinping — China

🟡 AI MUST REMAIN UNDER HUMAN CONTROL, BUT DEVELOPMENT CONTINUES

Who he is: President of China.

What he says: China explicitly recognises the possibility that highly advanced AI could create major risks, but it has not embraced the idea of broadly slowing development.

“Ensure that AI is always under human control.”

Xi has called for legislation, technical monitoring, early-warning systems and mechanisms capable of responding to AI-related incidents.

At the same time, China continues to accelerate the development and adoption of artificial intelligence.

The Chinese approach can therefore be summarised as control without surrendering technological speed.


6. SECURITY AND GEOPOLITICS


Eric Schmidt

🟡 SEES THE RISK, BUT ALSO SEES CHINA AS A MAJOR STRATEGIC THREAT

Who he is: Former CEO and chairman of Google, former chairman of the US National Security Commission on Artificial Intelligence and one of the most influential voices at the intersection of AI and national security.

What he says: Schmidt accepts that superintelligence could create serious risks. But he also believes it would be extremely dangerous if a strategic rival reached it first.

“Let’s imagine they get there first … they figure out some attack vector against America that we cannot imagine right now.”

His conclusion is not that the United States should reduce investment. It is that the country must invest enough to avoid losing the strategic race while simultaneously developing mechanisms to control the technology.


Lord Des Browne

🟢 COMPARES SUPERINTELLIGENCE WITH NUCLEAR RISK

Who he is: Former UK Defence Secretary, with direct experience in nuclear policy and national security.

What he says: Browne has said that for much of his career he considered nuclear war humanity’s greatest technological threat. He now believes superintelligent AI could belong in the same category.

“The development of superintelligent AI constitutes a threat analogous in scale to that of nuclear war.”

He supports banning the development of superintelligence in the United Kingdom and creating an international coalition seeking a broader prohibition.


WHAT THESE POSITIONS TELL US

Viewed individually, the statements appear highly divergent. Viewed by category, however, several patterns emerge.


Frontier-lab leaders

This is where the most surprising shift has occurred.

Anthropic, OpenAI, Google DeepMind and xAI have rapidly converged around the idea that the pace of frontier development should be controlled.

They are not proposing exactly the same measures and they do not assign the same probability to catastrophic outcomes. But convergence between companies competing this intensely is unusual.

Meta remains the most important major exception.


AI researchers

There is no scientific consensus.

Hinton, Bengio, Russell, Hubinger, Pachocki, Christiano, Coxon and Chughtai all believe recent developments justify serious concern.

But the presence of Yann LeCun on the other side is highly significant. He is not an outsider or commentator; he is one of the foundational figures of modern deep learning.

It would therefore be misleading to claim that AI science has reached a consensus that loss of control is imminent.


Infrastructure

The companies supplying platforms, chips and AI ecosystems are more cautious about the idea of a slowdown.

Nvidia opposes slowing development.

Microsoft accepts the seriousness of the warning signs but focuses more heavily on coordination and maintaining human control.

Hugging Face accepts the alignment problem but does not want a small number of closed frontier laboratories to determine the rules on their own.


Capital

Financial leaders largely move the debate away from extinction.

Larry Fink sees the danger of wealth concentration.

Jamie Dimon sees cyber risk, disinformation and labour-market disruption.

Michael Burry suspects that large AI companies may be turning safety concerns into competitive barriers.


Politics

Political leaders immediately encounter a problem researchers can more easily treat as secondary: China.

The Trump administration believes unilateral slowing by the United States could itself be dangerous.

The European Union favours risk controls, legal obligations and evaluation rather than a complete halt.

China recognises the possibility of losing control over advanced AI while simultaneously continuing the technological race.


Security and geopolitics

This produces perhaps the hardest question in the entire debate.

Even if Amodei, Hinton or Bengio are correct and slowing development would make AI safer, which country will agree to slow down if it cannot be certain that its strategic competitor will do the same?

This is why the debate increasingly resembles classic arms-control problems: a shared risk, limited trust and the fear that the first actor to exercise restraint could lose far more than a commercial competition.

And this may be the real novelty of the present moment.

The important development is not that some people believe advanced AI could become dangerous. Such warnings have existed for years.

It is that some of the people directly building the world’s most powerful AI systems have begun to say that the speed of development itself has become part of the problem.