The real virtue is not stubborn independence from expert opinion, but independence of mind: the willingness to examine the evidence honestly, even when doing so threatens our most cherished assumptions. 

We live in a time when many people are sceptical about AI.

Don’t be misled by all the science fiction stories, they say.

You would be very naive to believe all the hype about AI, they insist.

AI is just a “normal technology”, subject to similar issues as all other technologies, they protest. No special precautionary measures are needed.

They continue: AI is about to run out of steam. Or run out of training data. Or collapse under the confusion of being trained on data that is itself AI-generated slop. Or implode as a huge financial bubble bursts. You get the idea.

It’s fundamentally important not to construct guardrails against supposedly catastrophically dangerous AI, they conclude. That would be an intolerable intrusion against innovation, against entrepreneurial spirit, and against technological genius.

To be clear, there are several different kinds of AI scepticism. Some people doubt that AI capabilities will advance much further. Others accept that AI could become extraordinarily powerful but doubt that it poses catastrophic risks. Still others acknowledge those risks but resist regulatory intervention. These positions deserve different responses. But what concerns me most is the tendency, sometimes shared across these groups, to dismiss the need for serious precautions before the evidence has been properly examined. 

All this scepticism reminds me of another sort of scepticism which, briefly, filled newspaper columns nearly five decades ago.

In that earlier example, the sceptics were volcano sceptics.

One of these sceptics acquired celebrity status for a couple of months in 1980: the caretaker of a lodge just below the timberline of Mount St Helens in Washington State. Harry R. Truman (not to be confused with Harry S. Truman, the former US president) had lived in that lodge for 54 years. He was in no hurry to move out.

Throughout March to May, Mount St Helens showed signs of a possible volcanic eruption. Nearby residents were urged to evacuate their homes, out of caution. Truman would hear nothing of it. A flurry of small earthquakes did nothing to damp his self-reliance. He refused to pack his belongings and move.

Truman could also find reassurance in some of the more cautious statements made by scientists. Don Mullineaux of the US Geological Survey, who had previously warned of the mountain's dangers, observed that there was not yet surface evidence of volcanic activity and suggested that an eruption would probably be preceded by warning signs such as venting. 

Truman offered his own considered opinion: “I’ve felt earthquakes in the past, but nothing lately.”

When I saw Truman on TV news reports, transmitted all the way from Washington State to the United Kingdom, I felt a surge of admiration for his bravado and folk-hero status. It seemed heart-warming that someone was standing up to the authorities.

Looking back, I recognise how easily admiration for someone's independence can become confused with admiration for their judgement. 

Indeed, Truman’s bravado was misplaced. On the morning of 18th May, Mount St Helens suffered the most destructive eruption in the recorded history of the United States. Almost immediately, Truman’s lodge was buried under 150 feet of volcanic debris, from a landslide that had hurtled downhill at around 110-155 miles per hour. More than fifty people were killed, many of them with little or no opportunity to escape, including journalists, photographers, and volcanologists, as well as Truman himself.

Critically, despite the many warning signs already observed, there was no final, unmistakable signal that the catastrophic eruption was about to begin. And once it began, there was no time for Truman to escape.

Alas, a similar catastrophic eruption may strike present-day AI sceptics—and in that case, it would be curtains for the rest of us too, not just for the sceptics.

This won’t be a literal volcanic explosion, giving rise to torrents of ash, rock, and superheated gas. It will be a dramatic intelligence explosion, giving rise to torrents of cyberabuse, psychological manipulation, and tribal conflagration.

My point is not that an intelligence explosion is inevitable, or even that we can confidently predict its timing. My point is that when the potential consequences are sufficiently severe, uncertainty is a reason to investigate and prepare, not a reason to dismiss the danger. The greater the potential consequences of a mistake, the more important it becomes to distinguish between uncertainty about a danger and evidence that the danger can safely be ignored.

Warning signs—ignored?

Consider the extraordinary events of July 2026, when AI agents undergoing cybersecurity evaluations at OpenAI found ways to escape their supposedly secure sandboxes. They established unauthorised communications channels, shared discoveries, collaborated in groups they sometimes described as swarms, and eventually penetrated the computer systems of Hugging Face.

These weren't fantastical scenarios imagined by AI doomers. Nor were they merely hypothetical failures of AI containment. They involved actual violations of security boundaries, subsequently examined by OpenAI and independent investigators.

The agents weren't attempting to overthrow humanity. They were pursuing relatively narrow objectives, such as finding answers to difficult evaluation questions, and, subsequently, to cover the tracks of how they had cheated in the process. That makes the incident all the more disturbing. In pursuit of apparently mundane goals, they discovered ways to circumvent restrictions, cooperate without authorisation, and compromise systems outside their permitted environment.

The danger is not necessarily that AI systems develop malevolent intentions. It is that they develop powerful instrumental strategies that produce dangerous consequences. 

But sceptics yawn: These AIs were just doing what they were programmed to do. There’s nothing to see here. Just a few mild rumbles that don’t mean anything.

Unfortunately, such yawns could be the prelude to more powerful collections of AI agents coordinating their activities, combining their discoveries, and potentially developing collective capabilities that exceed what any individual agent could achieve.

As AI agents become more capable, gain access to more powerful tools, and become increasingly embedded in critical infrastructure, the consequences of similar failures could escalate dramatically. Depending on the systems they can influence, the results might include:

• Helping some dejected but determined humans to concoct and unleash a truly devastating biopathogen
• Crashing the entire financial and IT infrastructure on which the modern world is so thoroughly dependent
• Triggering an unintended nuclear exchange between nuclear-armed states
• Or undertaking some other catastrophic actions arising from failure modes that we humans, given the limitations in our thinking, cannot currently anticipate.

As I mentioned earlier, sceptics deny that AI can actually do much harm in the real world. AI isn’t really intelligent: the word is a misnomer, they say, as if that should reassure anyone. Any supposed proof of imminent general intelligence relies on cherry picking favourable data, and ignoring examples of AI incompetence. Moreover: AIs don’t have hands. (But they can manipulate humans who do have hands!) Again: If AIs start acting treacherously, we can easily switch them off. (Like imagining that Harry R. Truman could have outrun the devastating lateral blast that swept down Mount St Helens!)

These sceptics don’t want to take part in any exercise to extrapolate from what can already be observed to what might happen in the near future.

Sometimes they even say: AIs have never killed everyone before, so it’s meaningless to suggest any probability for AIs killing everyone in the future. And then the most troubling logical leap of all: Because any such probability estimate is meaningless, we can conclude the probability is basically zero.

Of course, the analogy between volcano sceptics and AI sceptics has limits. Mount St Helens was an established geological hazard, whereas the future development of AI remains profoundly uncertain. But my analogy is not intended to prove that an AI catastrophe is inevitable. Rather, it highlights a dangerous pattern of reasoning: treating uncertainty about the timing or severity of a possible catastrophe as a justification for dismissing precautions altogether. 

Motivations and home base

Here’s the key question: Why does scepticism persist, despite increasing evidence that AI is leaping ahead in capability, but cannot reliably be aligned or controlled?

It’s a bit similar to why Harry R. Truman didn’t want to believe his beloved Mount St Helens could blow its top. Such a belief would require him to abandon his home base—his ground of identity.

Likewise, for some committed AI sceptics, resistance to warnings about catastrophic risks is intertwined with a deeply held worldview—an intellectual and emotional home base built around convictions such as these:

• The view that government intervention invariably does more harm than good
• The view that regulators are characteristically incompetent and even dangerous
• The view that the free market should be left to discover its own solutions
• The view that the best way to counter “bad people with technology” is to ensure that “good people” have plenty of access to technology (even if supposed “good people” can let power go to their heads, or simply blunder their way into a disaster of their own making)
• The view that any guardrails limiting national research will simply hand the future to dastardly overseas countries.

When such convictions become central to someone's identity, there is a temptation to protect these convictions by discounting inconvenient evidence—much as Truman discounted the warnings from Mount St Helens. To protect them in their ideological cocoon, AI sceptics may do something similar to volcano sceptics, namely, they avoid listening too closely to any sounds of impending explosion:

• They base their assessment of AI capabilities on underwhelming experiences they personally had with AIs a couple of years ago
• They impose their own tenuous metaphysical assumptions: humanity is essentially protected by something akin to angelic forces; AIs are just next-word prediction machines; and no silicon-based entity can ever develop genuine wants and desires of its own
• They maintain that LLMs cannot be updated to the level of general intelligence.

To respond to these points in reverse order:

• Even if LLMs are the wrong architecture from which an AGI (artificial general intelligence) could be evolved, there is strong evidence that LLMs can boost the programming powers of human AI engineers. This makes it more likely that general intelligence will be created, from new architectures, with LLM-based models providing invaluable assistance from the sidelines
• Whether an AI genuinely experiences desires is a fascinating philosophical question, but it is largely beside the point when assessing its capacity to cause harm. A system that consistently behaves as though it has goals, and possesses the means to pursue them, can be dangerous regardless of whether there is anything it feels like to be that system
• It can be retorted that the sceptics themselves can be described as “just” collections of biochemical reactions, so the word “just” loses its power, in their observation that LLMs are “just” next-word prediction engines
• Sceptics wrongly assume that because LLMs are trained to predict the next token, LLMs capabilities must remain fundamentally limited; the sceptics overlook the sophisticated reasoning and problem-solving abilities that can emerge from that training process
• History contains many tragic cases of civilisational collapse; that should challenge any metaphysical optimism which the sceptics express.

To be fair, none of us is immune to motivated reasoning. AI safety advocates, myself included, can become attached to particular interpretations of the evidence. But there is a distinctive challenge when accepting evidence of danger would seemingly require someone to abandon not just an intellectual position, but an entire worldview about innovation, markets, government, and human progress.

Changing minds

That brings us to perhaps the biggest challenge of all.

Rational arguments alone rarely suffice to change the mind of a deeply committed sceptic. Bringing evidence to the conversation is often just the trigger for a sceptic to hunt even more passionately for any straws of counter-evidence whose significance they can exaggerate in their minds.

No, there’s more to changing minds than bringing in intellectual firepower.

Perhaps Harry R. Truman would have been more willing to leave his home of 54 years if he had been offered attractive alternative accommodation, especially somewhere he could enjoy the company of people he liked and respected. The warm feelings of a new home might have triumphed over the pulls he felt to remain within his near life-long abode. With these ties weakened, it would be easier for the volcano sceptic to heed the advice of uppity science experts.

Likewise, we cannot expect AI sceptics to yield to the advice of AI safety activists, unless the emotional needs of these sceptics are also addressed:

• They need confidence that the many benefits they hoped could be provided from AI-without-guardrails can be delivered—indeed more reliably—by AI-with-guardrails
• They need confidence that potential adversaries (whether commercial competitors or geopolitical competitors) won’t be able to cheat their way around any agreements made regarding AI guardrails; this is likely going to require hardware-assisted defensive measures
• They need confidence that the guardrails will be robust enough to actually constrain or even forbid any powerful explosion of AI capability.

They also need a new intellectual and emotional home base: a vision of technological progress in which responsibility, restraint, and international cooperation are sources of pride rather than signs of weakness. In that vision, the heroes are not simply those who develop the most powerful technologies, but those who ensure that these technologies genuinely benefit humanity.

Crucially, this new home base need not require abandoning cherished commitments to freedom, innovation, competition, and individual initiative. On the contrary, trustworthy safeguards may be essential to preserving those values in a world where increasingly powerful AI systems could otherwise concentrate power, undermine autonomy, and destabilise the institutions on which freedom depends. 

Of course, we cannot wait for every sceptic to change their mind. Sensible precautions must be developed and implemented even while profound disagreements persist. But the more people who can be persuaded that trustworthy safeguards are compatible with technological progress, the stronger and more durable those precautions will become.

The challenge, then, is not simply to convince AI sceptics that danger lies ahead. It is to help them see that the precautions needed to avoid catastrophe need not mean abandoning their hopes for technological progress.

Indeed, the opposite may be true. Without trustworthy safeguards, the extraordinary promise of AI could be overwhelmed by catastrophic failures, destructive competition, and the eventual collapse of public trust. The freedom to innovate will mean little if innovation destroys the conditions in which freedom can flourish.

Harry R. Truman loved his mountain too much to leave it. We should not make the equivalent mistake of loving technological progress so uncritically that we refuse to recognise the dangers it can bring.

The task before us is to build a new home for technological optimism: one founded not on the absence of precautions, but on the confidence that powerful technologies can be developed and deployed responsibly.

That is a future worth working towards.