It is worth pausing to ask how artificial intelligence — a technology that is compressing decades of scientific research, enabling one-person businesses to compete with large corporations, and beginning to crack diseases that have defied medicine for generations — arrived at a point where the public conversation is dominated by robot uprisings, mass unemployment, and the end of civilisation.
Something went wrong. Not with the technology. With the story being told about it.
The Scare Merchants
What do you expect from a product where the founding marketeers employed scare tactics as part of their sales pitch?
From the outset, AI was offered to the public in one of two flavours: existential menace or messianic saviour. Neither was honest. And when you trace the extinction rhetoric back to its sources, the motives become rather less mysterious.
In May 2023, Sam Altman testified before the US Senate that 'I think if this technology goes wrong, it can go quite wrong.' He had reportedly told OpenAI staff internally that he thought they might all die — while simultaneously raising ten billion dollars from Microsoft. In the same period, he called for a federal AI licensing regime, the sort of regulatory architecture that an incumbent with deep pockets can navigate with ease, and that a startup operating from a garage cannot. The two positions are not in contradiction. They are the same position.
Elon Musk provided ten million dollars to the Future of Life Institute, which organised the March 2023 open letter calling for a six-month pause on advanced AI development. He signed the letter himself. Four months later, he launched xAI. The pause, had it been observed, would have frozen OpenAI's lead while Musk assembled his own capability. OpenAI did not pause.
A May 2023 statement from the Center for AI Safety — signed by Altman, Demis Hassabis, Dario Amodei, Geoffrey Hinton, and Yoshua Bengio — declared that mitigating the risk of AI extinction should rank alongside nuclear war and pandemic as a civilisational priority. Anthropic, Amodei's company, raised 7.3 billion dollars that year, in no small part on a safety-first identity.
The one senior voice consistently refusing to play along was Yann LeCun — Turing Award winner (the technology world's equivalent of a Nobel Prize), formerly Meta's Chief AI Scientist, now focused on his World Model thesis: his argument that genuine machine intelligence will require models that understand the physical world rather than simply predict the next word. LeCun called the extinction claims 'preposterously ridiculous' and the FLI pause letter 'counterproductive and dangerous'. He said plainly what many researchers knew privately — that large language models are extraordinary pattern-matching systems, not minds, and not the science-fiction threat being marketed to legislators and the press.
The Job Apocalypse That Wasn't
Industrial disruption has never arrived without some job displacement. The Luddites were not irrational — the mechanisation of the textile trade genuinely threatened their livelihoods. The mechanisation of agriculture emptied the countryside. The internet made a great many roles redundant before it created far more. Disruption has always looked, in the moment, like destruction. It rarely is.
The question worth asking is not whether AI will displace some categories of work — it will — but whether the history of human ingenuity at any point suggests we simply run out of things to do. The evidence does not support that conclusion.
The UBI Fantasy
A related argument has been circulating in technology circles: that AI will generate such extraordinary wealth that governments will be able to fund a universal basic income — effectively paying people not to work, forever. It is worth sitting with this for a moment.
When has there ever been a business model where the businesses make their founders wealthy; pay enough tax so that no one else has to work; make enough money to reinvest in the business and maintain their place at the leading edge of development; pay their investors a healthy dividend yield — and, oh yes, charge reasonable prices in a competitive environment? Sounds like Nirvana to me.
The Threat That Is Actually Real
Here is the irony. While legislators have been consuming testimony about robot uprisings and existential risk, a genuine and documented threat has been advancing largely outside the main conversation.
The internal Anthropic codename 'Fable' refers to what was publicly released as Claude Opus 4.5 in July 2025. According to Anthropic's own model safety documentation, the system can autonomously exploit previously unknown — zero-day — software vulnerabilities, and write functional malware components. This is not speculation. It is in the developer's own published materials.
US Treasury Secretary Scott Bessent personally rang Anthropic CEO Dario Amodei to tell him he was alarmed. His specific concern was direct: should China obtain a copy of the model, it could materially enhance China's cyberattack capabilities against Western infrastructure. Bessent has since been involved in White House-level discussions on the matter. Members of the Senate Intelligence Committee have begun calling for hearings.
Context: Microsoft and OpenAI reported in early 2024 that five state-backed hacking groups — from China, Russia, Iran, and North Korea — were already using AI models to assist cyberattack operations. University of Illinois researchers demonstrated the same year that GPT-4 could autonomously exploit real zero-day vulnerabilities with an 87% success rate.
This — not cinematic dystopia — is where genuine regulatory attention belongs.
The Politicians Who Don't Use It
There is a considerable volume of AI regulation being proposed by people who do not appear to use AI. A journalist asked Keir Starmer, more than once, whether he used it. The Prime Minister did not initially answer. When pressed, he admitted that his children did. That exchange tells you most of what you need to know about the regulatory debate: it is being conducted at considerable distance from the technology itself.
Datacentres: Planning, Not Panic
AI runs on hardware — enormous quantities of compute, electricity, and cooling, housed in datacentres. There has been a backlash in the United States, where facilities can be built with dedicated power generation and place demands on local water supplies. Communities have objected to their presence. This is not new behaviour. Communities have always objected to large infrastructure — power stations, motorways, distribution hubs. The datacentre debate is a planning debate and an energy debate. It is not an existential one.
What Hassabis Got Right
Demis Hassabis has conducted himself rather differently from most of his peers. Where others have led with fear or grandiosity, he has pointed consistently to concrete achievements. AlphaFold — DeepMind's protein-folding system — has compressed what might have taken decades of biological research into a matter of years. It will almost certainly contribute to treatments and cures for diseases that have resisted medicine for generations. That is not a promise. It has already happened.
The One-Person Business
Among the most significant and least reported developments enabled by AI is the genuine viability of the one-person business at institutional scale. The ability to build research, analysis, client communication, and operational infrastructure without a large team is quietly transforming what it means to start something. Something I have been trying to develop and create with Tara Capital — an AI-native investment research platform built as a connected system, not a collection of reports, combining macro research, implementation, and portfolio construction in a single framework.
Regulate What You Can See
You cannot regulate for disasters that have not yet happened. You cannot know in advance what bad actors will find to exploit. What you can do is focus regulatory energy on the threats that are already visible and documented — the cybersecurity dimension is the obvious starting point — rather than expending it on speculative extinction scenarios that happen to have suited certain commercial interests rather well.
The larger point is this: the benefits of AI will not be unlocked by committees writing rules about a technology they have not used. They will be unlocked by a generation of people free to experiment — to learn, by doing, what this technology is capable of and where its limits lie. That is how every tool in human history has been understood. There is no reason to expect this one to be different.
More at taracapital.io