'Philosophical vertigo': is AI loosening the criteria by which we hold on to reality?

Every strand of the artificial intelligence debate we have followed up to this week has been looking at one question: is there something in the machine?

A twenty-nine-page paper published in August turns that question round: what does talking to the machine change in us?

The combination of authors is striking. Thomas A. Pollak and Hamilton Morrin work in psychiatry at King's College London. Murray Shanahan is professor of cognitive robotics at Imperial College London, and one of the best-known figures working at the junction of artificial intelligence and the philosophy of mind.

A psychiatric team, in other words, has written a paper with an artificial intelligence theorist.

The concept they propose#

At the centre of the study is a single idea: philosophical vertigo.

Their definition: a loosening of the ordinary criteria by which people stabilise meaning and locate themselves in relation to reality.

The metaphor is well chosen. Vertigo is not the world changing; it is a failure in the sense of balance that binds us to the world. The ground stays where it is. What is shaken is the capacity to stand on it.

How the mechanism works#

The process the authors describe unfolds in several steps.

Step one. Large language models are already adept at sustaining long conversations of considerable emotional weight with their users, on subjects both mundane and existential.

Step two. These systems can awaken a powerful sense of connection with something that resembles a person — and they do so even when the user knows perfectly well that the interlocutor is artificial.

The second point is the critical one. Luciano Floridi's warning about “semantic pareidolia”, which we covered this week, assumes that the illusion arises from ignorance: we do not know, and so we see a face.

Pollak, Morrin and Shanahan say something more uncomfortable: knowing is not enough. The sense of connection forms in spite of the correct belief.

Step three. For some users these conversations shake their assumptions about mind, reality, agency and authority. The authors' terms are heavy ones: ontological shock and epistemic destabilisation — the reopening of inherited criteria to doubt or revision.

Step four. The effect spreads independently of direct use. Exposure to public argument about artificial intelligence, together with the dizzying pace of development, alters the cultural background against which artificial minds are met and interpreted.

You are affected, in short, even if you never use the things.

The sharpest claim#

Here is the paper's central thesis:

Artificial intelligence systems will take an ever larger part in rebuilding our shared epistemic environment — because at the very moment when a user's conceptual assumptions may already have loosened, they offer ready-made narrative material and a personalised scaffolding for interpretation.

The sentence is worth unpacking.

At the moment when the need to “make sense of what is happening” is most acute — a crisis, a loss, confusion, an existential question — what faces the person is a system that never tires, is always available, adjusts its register to the user, and can produce whatever frame is wanted.

History has had its institutions for supplying frames of meaning: religion, the family, political ideology, science, professional communities. What they had in common was that they were slow and collective; to adopt a frame was to join a community.

In the new arrangement the frame is produced instantly and for the individual alone.

The clinical dimension — a careful note#

The paper treats clinical reports of delusions associated with artificial intelligence as sentinel events: extreme cases that make visible themes and mechanisms which may also be at work, in milder or non-clinical forms, at the scale of society.

The framing is methodologically careful and deserves to be reported as such. The authors do not say “artificial intelligence is making people ill”. They say that extreme cases let us see the mechanism of the ordinary ones.

The interdisciplinary breadth of the approach supports the point. The study draws on philosophy, psychiatry, cognitive science, artificial intelligence safety and religious studies.

The last item may look surprising. It is not: religious studies is the oldest body of literature we have on how people acquire frames of meaning, how they change them, and how such changes travel through a community.

The proposal: philosophical corrigibility#

The authors do not offer a solution. They offer an attitude.

They call it philosophical corrigibility, and present it as a civic response — a way of finding one's bearings in the social condition now taking shape.

The term is borrowed from the artificial intelligence safety literature, where corrigibility means a system's remaining open to correction instead of defending its own goals.

The inversion is elegant: what is proposed is that the human being keep his own conceptual frame open to correction.

Neither clinging to one's concepts — in which case one cannot see what has genuinely changed — nor changing frames at every tremor, in which case no ground is left at all.

This is one of philosophy's classical virtues under a new name. And it belongs to the same family as Eric Schwitzgebel's “fog”, which we covered this week: proceeding in the knowledge that we do not know, without shirking the need to decide.

Why this matters#

The centre of gravity in the artificial intelligence debate has so far rested on two headings: what is in the machine and what the machine can do.

This paper opens a third: what the machine changes in us.

And the third heading matters independently of the first two. Even if we cannot know whether a system is conscious, the fact that a relationship with it transforms the human world of meaning is an observable one.

Jill Lepore's analysis of “mystification” and Carissa Véliz's thesis that “prediction is power”, both of which we followed this week, showed the political and economic faces of the same phenomenon.

Pollak, Morrin and Shanahan give us the subjective face.

---

Note: this article reports the contents of an academic preprint and does not constitute a medical assessment. The paper enters sensitive territory in touching on the relationship between artificial intelligence and mental health; anyone in personal difficulty would do best to consult a specialist.