I am, I freely admit, the wrong man to ask where to eat. My tastes run to the esoteric. I would choose the Royal Nawaab in West London — a cavernous, clattering Pakistani grill hall — over the celebrated Benares on Berkeley Square without a second’s hesitation, and I have on occasion failed to talk even my wife round to the merits of my choices. On finance infrastructure, its plumbing and its reform, I will back my judgement against most people’s. On where to have dinner, you should almost certainly ignore me. The two facts sit together perfectly comfortably — and in that comfortable coexistence lies the single most useful habit anyone can bring to artificial intelligence.
For the habit is this: authority does not transfer. Being right about one thing confers no warrant to be trusted about another, and the reflex to assume it does — to let eminence in one field vouch for opinions in a second — is among the most reliable ways to be led astray. A gifted surgeon is not thereby an authority on tax policy; a Nobel physicist pronouncing on interest rates is merely a clever person with a view. We all know this in the abstract, and yet in practice the halo travels: we take a specific credential as a general licence. Nowhere is that temptation stronger, or the cost of indulging it higher, than on a subject as crowded, as consequential and as loud as this one.
Describe and predict
What makes artificial intelligence especially treacherous is that the trap is laid not only between fields but within this one. It is tempting to imagine that “an AI expert” is a single, coherent kind of authority — that those who built these systems can therefore tell you what the systems will do to your industry, your job, or the world. They cannot; or rather, when they do, they are performing a different act from the one their expertise underwrites. The cleanest way to see this is to separate two questions that are constantly, and damagingly, run together.
The first is description: how do these systems work, and what can they do today? Here genuine expertise exists, and — crucially — it is checkable. A claim about what a model does can be tested against the model itself. The second is prediction: where is this going, how fast, and with what consequence? Will it arrive at something like general intelligence; will it hollow out employment; does it threaten catastrophe? Here there are no experts, only better- and worse-informed opinions, because the future has not yet happened and cannot be tested. The costliest error in the whole discourse is to grant someone’s authority over the first question — which may be absolute — as though it settled the second, on which they are guessing like the rest of us, if more expensively.
There is no sharper demonstration of this than the founders of the field themselves. Geoffrey Hinton, Yoshua Bengio and Yann LeCun shared the 2018 Turing Award for the very breakthroughs that made the modern era of AI possible; between them they can describe how these systems work with an authority almost nobody alive can match. And they now disagree, profoundly, about what comes next. Hinton left Google in 2023 and, having since won a Nobel Prize, has become one of the most prominent voices warning that the technology may pose a catastrophic, even existential, risk. Bengio, the most-cited researcher in the field, shares much of that concern and has founded a research organisation dedicated to making the technology safer. LeCun, who spent more than a decade as Meta’s chief AI scientist before leaving in 2025 to pursue a wholly different architecture, regards the existential alarm as overblown and the dominant paradigm — the large language model — as, in his words, “a dead end”. Three people; equal and unimpeachable authority to describe the thing; three irreconcilable forecasts of where it goes. Their disagreement is not a scandal. It is the clearest possible signal that prediction here is opinion — and that anyone selling certainty about the future of AI, in either direction, is selling something the evidence cannot cover.
A balanced plate
None of which means all voices are equal, or that one should simply throw up one’s hands. It means the opposite: reading well requires assembling a spread of serious people who disagree, and weighting each for the specific thing they genuinely know. I would not presume to offer a canon — that would rather miss the point — but a few illustrations of the kinds of voice worth keeping in the mix. The three laureates, precisely because they diverge, are the best available map of the live debate. Gary Marcus, the cognitive scientist, has argued for years against the prevailing enthusiasm that deep learning alone will not deliver reliable reasoning — the standing sceptic every diet needs, if only to stress-test the hype. Ethan Mollick, of Wharton, studies empirically, and without evident agenda, what these tools actually do inside real organisations — nearer most readers’ concerns than any frontier speculation. And Paul Graham — no kind of AI researcher, which is precisely the point — is worth reading not for expertise in the technology but for clarity of thought, and for a long view of how new technologies actually spread once the initial noise dies down. The value lies in none of them alone. It lies in the triangulation: several honest, well-informed people who do not agree, read against one another.
An information diet
Which points to the discipline itself, for anyone obliged to form a view on this without living inside it. Build an information diet as you would build any other: from the best sources, in the right order. Nearest the truth sit the primary ones — the laboratories’ own technical reports and model documentation, which describe, checkably, what a system is and does. Next, the honest interpreters, who explain and contextualise with no product to sell. Furthest, and to be taken with the most salt, is the noise: the vendor talking its own book, the headline reaching for ecstasy or alarm, the supremely confident thread from someone who has skin in neither the research nor your outcome. And over all of it, run the two tests this piece has pressed. Is this a claim about what the technology is and does — checkable, and the province of real expertise — or about where it is going — unprovable, and everyone’s guess, dressed to the standing of its author? And whatever the eminence of the source, is it eminence in this specific thing, or a halo borrowed from some other?
There is a pleasing symmetry to end on. The same discipline I would ask you to apply to Hinton, or to me, applies to everyone — myself and this series included: worth hearing on some matters, to be read sceptically on others, and never to be trusted merely for having been right about something else. I remain wholly confident on finance infrastructure and wholly unreliable on restaurants, and I have made my peace with both. The trick — on artificial intelligence as on dinner — is not to find the one voice that is always right, because there is no such voice. It is to know whose judgement to trust on precisely what, to tell the checkable from the speculative, and to keep enough genuine disagreement in the room that no single confident voice, however garlanded, can do your thinking for you. Knowing where your own lane ends is not a limitation. On a subject this loud, it is very nearly the whole of the skill.