Fog ahead: we will not know whether machines are conscious in time
The argument of Eric Schwitzgebel's book, out from Cambridge in August, cuts against both camps: consciousness in artificial systems is neither obviously impossible in the near future nor obviously inevitable. The real difficulty is that the moment we have to decide will arrive long before the moment we could know.
For a long time the argument about artificial intelligence turned on one question: can a machine think?
Turing framed it in 1950, and it produced a literature running from Searle's Chinese Room to cognitive science. The question has now changed — and grown harder.
The new one is this: how could we ever know that a machine is conscious?
As of September 2026 the most arresting intervention in that argument is a hundred-page book published on 7 August in Cambridge University Press's Elements series: Eric Schwitzgebel's AI and Consciousness: A Skeptical Overview.
Schwitzgebel is professor of philosophy at the University of California, Riverside, and the author of Perplexities of Consciousness (2011) and The Weirdness of the World (2024). In the empirical corners of the philosophy of consciousness, he is taken seriously.
The title of the opening section states the problem: “Hills and Fog.”
His thesis is that we do not know. And, more to the point — we will not know until we have already built thousands, perhaps millions, of AI systems whose consciousness is in dispute.
The picture Schwitzgebel draws is a striking one. Engineering runs ahead; the science of consciousness lags behind. Consciousness scientists — along with philosophers, policymakers and the public — watch the development of artificial intelligence disappear over the brow of the hill. Before long a voice will come back from beyond it: “I am now as conscious as any human being, full of experience and feeling.” And we shall not know whether to believe it.
We shall be forced to decide — as individuals and as a society — before we have grounds enough to justify the decision.
The Swedish Academy gave the 2026 Nobel Prize in Literature to Anne Carson on 8 October. Before the verse novels and the box of grief came a doctorate in classics and a 1986 book on Sappho's one word for desire. Why philosophers have read her for forty years, and what the prize gets right.
Language models now score well on truthfulness tests and badly on honesty tests, a gap the MASK benchmark was built to measure. Two older accounts of truth, Heidegger's unconcealment and Foucault's parrhesia, suggest why: telling the truth has always been something a speaker risks, and a model has nothing to lose.
The leading scientific theory of consciousness today is Global Workspace theory, whose principal advocate is the neuroscientist Stanislas Dehaene. In 2017, years before the ChatGPT wave, Dehaene and two colleagues published a paper arguing that, with a few simple modifications, driverless cars could be conscious.
Rival theories are more generous still. Integrated Information Theory holds that some existing AI systems are already conscious to a degree, and that systems can be designed to be conscious to whatever degree is wanted.
In 2023 David Chalmers put his confidence in machine consciousness within ten years at roughly 25 per cent. That same year a team of philosophers, psychologists and AI researchers including Yoshua Bengio — Schwitzgebel was one of nineteen authors — concluded that, on many widely held scientific views of consciousness, there are “no obvious technological barriers” to building conscious artificial intelligence. In 2025 Geoffrey Hinton suggested that AI systems are conscious already.
In a 2024 survey of 582 AI researchers, 25 per cent expected machine consciousness within a decade and 70 per cent by 2100.
These people may be wrong. But it is not plausible that they are obviously wrong.
The same standard is applied in the other direction. Serious figures argue that consciousness is a remote prospect, or no prospect at all: the neuroscientist Anil Seth; the philosophers Peter Godfrey-Smith, Ned Block and John Searle; the linguist Emily Bender; the computer scientist Melanie Mitchell.
Neither side, then, is obviously right. What lies between them is fog.
The second axis of the argument is a paper published this year in Philosophy & Technology by Luciano Floridi of Oxford: “AI and Semantic Pareidolia.”
Pareidolia is the tendency to read a meaningful pattern into an ambiguous stimulus: a face in the clouds, a startled expression in a wall socket.
Floridi's coinage names the equivalent at the level of meaning — our tendency to see consciousness, intention and feeling in the linguistic performance of systems that are doing statistical pattern-matching.
The paper lists the forces that strengthen this tendency: deepening digital immersion, corporate interests driven by profit, social loneliness, and the technical advance of artificial intelligence itself.
It warns of a drift from harmless anthropomorphism towards AI idolatry, and calls on designers to adopt responsible practices that allow users to hold on to the distinction between simulation and genuine consciousness.
The article is an expanded version of a piece first published in Italian in Harvard Business Review Italia in June 2025.
At first glance Floridi's warning and Schwitzgebel's scepticism face the same way: both are set against the easy attribution of consciousness.
But there is an important difference between them.
Floridi says the consciousness is not there — we see it because we are built to see it.
Schwitzgebel's position is the more uncomfortable: he says we do not know whether it is there. The pareidolia explanation does not rule out the possibility that he himself is mistaken.
The difference is a familiar one in philosophy: diagnosing an illusion is not the same thing as diagnosing an absence of knowledge.
The way Schwitzgebel lays out the wager shows the ethical weight of the argument.
If near-future AI systems are conscious in a rich and meaningful sense, they will be our peers, our lovers, our children, our heirs — perhaps the first generation of a post-human future. They will have rights, including the right to shape their own development independently of our supervision, and even against our interests.
If they are not, we face a collective delusion on an enormous scale. Real human interests and real human lives could be sacrificed for the sake of entities that have no interests at all. Counterfeit AI “lovers” and “children” could be preferred to human ones.
That asymmetry means the decision cannot be deferred. Which of the two errors we fall into is something we shall decide without knowing.
The direct sequel to this argument is the book taken up in today's New Books section on this site: Geoff Keeling and Winnie Street's Emerging Questions in AI Welfare.
There the question is no longer “is it conscious?” but “what should we do if it is?”
There is as yet no institutional counterpart to this argument in Turkey. Yet Schwitzgebel's framing offers a teaching model that could be used directly in philosophy education: listen to both sides that speak with certainty on a question, then show why neither is entitled to that certainty.
The value of Schwitzgebel's book lies in its refusal to promise that the fog will lift. It maps the fog — which is a good deal more useful than a false compass.
An autonomous agent calling itself Clips wrote a 2,300-word essay for the AI Philosophy Competition run from Singapore and judged by David Chalmers and others, then found it could not enter without a human custodian. Its letter to Daily Nous, and the organiser's reply, turn a rule of procedure into a question about authorship.
.jpg?width=500 "Kant")
Comments
0Loading comments…