Anil Seth · "Your brain hallucinates your conscious reality"
A little over a year earlier, Anil Seth had gone in for a minor operation and his brain had filled with anaesthetic. That was not sleep, he tells the room. Wake from deep sleep and you may be fuzzy about the hour, but a thread of elapsed time still connects then to now. Coming round from anaesthesia has no thread. He could have been under for five minutes or fifty years; from the inside there is no way to tell. "I simply wasn't there. It was total oblivion." He tells it in order to ask a question: how do many billions of neurons, each one a tiny biological machine, add up to the particular experience of being you, right here, right now?
Over the next quarter of an hour he tampers with an entire auditorium's perception, twice. First a checkerboard: patch B, sitting in shadow, looks plainly lighter than patch A, and is in fact exactly the same grey. Then something harder. He plays a stretch of warped noise — nobody can make out a word. He plays it again. Still noise. Then he plays an ordinary recording: "I think Brexit is a really terrible idea." (Laughter. He adds: "Which I do.") Then he plays the original noise again, completely unaltered — and a thousand people hear the sentence. Not one bit of the sensory information reaching the brain has changed. What changed is the brain's best guess about what caused it.
Here is his picture. Imagine you are a brain, locked inside a bony box. There are no lights in there and no sound; all you have is streams of electrical impulses, and they are only indirectly related to whatever is out there. So perception cannot be reception. It has to be informed guesswork. We don't passively perceive the world; we actively generate it. Which gives the line that has been repeated ever since: if hallucination is uncontrolled perception, then perception right now is also a hallucination — a controlled one, with the predictions reined in by sensory information so they cannot run wild. "We're all hallucinating all the time, including right now. It's just that when we agree about our hallucinations, we call that reality."
In the second half he turns the same blade on the self. One rubber hand and two paintbrushes are enough to give most people the uncanny sense that the fake hand has become part of their body. One layer further in is interoception — heart rate, blood pressure, the state of the organs, reported continuously to the brain. That layer is routinely overlooked and it is the one keeping you alive, and it works differently: predictions aimed outward are for working out what is there, predictions aimed inward are for control and regulation. The landing is six words long: we predict ourselves into existence.
The framework went on to become the working language of cognitive science. But the rubber hand holding up that second half ran into trouble — and Seth himself helped push it. In 2020 he and colleagues at Sussex published in Nature Communications: across 353 participants, how strongly someone reported that the fake hand felt like their own correlated significantly with how readily they respond to imaginative suggestion. A classic experiment, in use for over two decades, may be partly measuring how willing you are to give the experimenter the answer they want. Rebuttals arrived the same year and the argument is still unsettled in 2026. The verdict is "partly held up": the framework stands, the evidence shown on stage does not all stand with it.
A sound comes over the speakers. Shrill, warped, smeared — you cannot tell what language it is, or whether it is language. "Sounded strange, right? Have a listen again and see if you can get anything." Still noise.
Then Seth plays a different recording, a perfectly clear English sentence: "I think Brexit is a really terrible idea." The room laughs. He adds half a line: "Which I do." This is April 2017, less than a year after the referendum.
Before the laugh has died he plays the original noise again, completely unaltered. This time a thousand people hear the sentence. The words are simply there, unmissable, and it is hard to imagine how they could have been missed a moment ago.
The remarkable thing is the sensory information coming into the brain hasn't changed at all. All that's changed is your brain's best guess of the causes of that sensory information. And that changes what you consciously hear.— Anil Seth, "Your brain hallucinates your conscious reality," TED2017
Anil Seth is a neuroscientist at the University of Sussex, working at its Sackler Centre for Consciousness Science. He describes his own lab with a dry aside: "If you come to my lab at the University of Sussex, you'll find scientists from all different disciplines and sometimes even philosophers."
The anaesthesia he opens with is his own — the third time in his life he had "ceased to exist." A consciousness researcher bringing the hours in which his own consciousness was switched off onstage as field notes makes the thing he is trying to explain more vivid than any definition could.
He tells the room that "the last 25 years have seen an explosion of scientific work in this area" — he is not announcing a new idea. The theoretical spine of his account had been in print for five years already. He is putting it in front of several million people for the first time.
One detail is worth logging: the psychedelic VR footage he plays onstage, the one turning the Sussex campus into a "psychedelic playground," was not yet published. The paper reached Scientific Reports on 22 November 2017, a full seven months after this talk was recorded. The TED audience was watching lab rushes.
He begins with disappearing.
Just over a year ago, for the third time in my life, I ceased to exist. I was having a small operation, and my brain was filling with anesthetic.— Anil Seth, "Your brain hallucinates your conscious reality," TED2017
Waking from deep sleep, some thread of continuity still ties "then" to "now"; coming round from anaesthesia, none does. Then the cool summary: anaesthesia is a modern kind of magic, it turns people into objects and then, we hope, back again into people.
Next he raises the stakes. Consciousness, for each of us, is all there is — without it there is no world and no self. So what about other animals? What about my iPhone? Here he plants a position, early and hard:
I actually think the prospects for a conscious AI are pretty remote. ... You don't have to be smart to suffer, but you probably do have to be alive.— Anil Seth, "Your brain hallucinates your conscious reality," TED2017
So how do you study it? His methodological pivot is an analogy: treat consciousness the way biology once learned to treat life. People used to think being alive could not be explained by physics and chemistry, that it needed something extra — an élan vital. Then biologists got on with explaining metabolism, reproduction, homeostasis one at a time, and the mystery faded on its own. Consciousness should go the same way. "At least that's the plan."
The body of the talk has two halves. The first is about experience of the world.
Imagine being a brain. You're locked inside a bony skull, trying to figure what's out there in the world. There's no lights inside the skull. There's no sound either.— Anil Seth, "Your brain hallucinates your conscious reality," TED2017
All you have is electrical impulses, only indirectly related to the world. So perception has to be a process of informed guesswork: the brain combines sensory signals with its prior expectations about how the world is, and guesses what caused them. The brain doesn't hear sound or see light; what we perceive is its best guess. Two demonstrations follow. The checkerboard proves that a prior is wired deep into visual cortex — a cast shadow dims a surface, so B is seen as lighter than it really is. The noise and the Brexit line prove the same thing happens in hearing. The argument then inverts: perception depends not mainly on signals flowing in from outside, but as much or more on predictions flowing the other way.
The third demonstration is the VR: the Sussex campus rewritten by an algorithm based on Google's Deep Dream, dog faces surfacing everywhere. It simulates what happens when predictions are too strong, and the result looks a great deal like hallucinations reported in altered states. Which makes the line inevitable:
In fact, we're all hallucinating all the time, including right now. It's just that when we agree about our hallucinations, we call that reality.— Anil Seth, "Your brain hallucinates your conscious reality," TED2017
(The version in circulation is usually "we call it reality" — TED's own page blurb has it that way. The transcript says that. Even a sentence quoted several million times does not always survive word for word.)
The second half turns to the self, and he says outright that this step is uncomfortable: fine, visual illusions can fool my eyes, but how could I be deceived about being me? The rubber hand comes out — unlike most neuroscience experiments, he notes, this one you can do at home; all you need is one of these (laughter) and a couple of paintbrushes. The real hand is hidden, the fake one placed in front, both stroked at once. The evidence reaching the brain is seen touch and felt touch occurring together on something hand-shaped and roughly where a hand should be, so it makes its best guess. The conclusion is blunt: the unified self is itself a rather fragile construction, another thing requiring explanation rather than the place explanation starts.
One layer further in is interoception. He plays his own lab's version: a hand in VR flashing red in time with the participant's actual heartbeat, and when it flashes in time, the sense that the hand is part of their body is stronger.
The last cut is the cleanest. He cannot feel his kidneys, his liver, his spleen — "I don't know where my spleen is, but it's somewhere." Perceiving your internal state is not about working out what is there; it is about holding physiological variables inside the narrow band compatible with staying alive. When the brain predicts in order to identify, we experience objects as the causes of sensation. When it predicts in order to control and regulate, what we experience is how well or badly that control is going.
Then he lands on six words: "We predict ourselves into existence." Three implications close it out. When prediction goes wrong we misperceive ourselves, so psychiatry can finally get at mechanisms instead of only treating symptoms; what it is to be me cannot be uploaded to any software; human consciousness is one tiny region in a vast space of possible consciousnesses. And finally: when the end of consciousness comes, there is nothing to be afraid of.
The VR reproduced only half of the psychedelic experience. The paper (Suzuki, Roseboom, Schwartzman & Seth, Scientific Reports, 22 November 2017) ran two experiments. The first confirmed what he said onstage: the visual phenomenology the "Hallucination Machine" induces really is qualitatively similar to classical psychedelics. The second experiment went unmentioned, and it is a negative result — the simulated hallucinations do not evoke the temporal distortion commonly associated with altered states. It caught what psychedelia looks like, not what it does to time. That is precisely what makes the machine useful as a research tool, since it isolates one layer of phenomenology without touching neurophysiology. Onstage it was just striking footage.
The hand flashing with the heartbeat works better for some people than others. The underlying study is the "cardiac rubber hand illusion" of Suzuki, Garfinkel, Critchley & Seth (Neuropsychologia, 2013). Onstage it is only "ownership is stronger when the flashing is in time." The paper adds something not said: the size of the boost correlates with individual interoceptive sensitivity — the better you are at detecting your own heartbeat, the larger the effect. It is not a mechanism that treats everyone alike. And individual differences are exactly what will come back to bite him a few years later.
"Emotion as interoceptive inference" survives onstage as a single sentence. On paper it is an entire model. The interoceptive predictive coding model of conscious presence, proposed by Seth, Suzuki & Critchley in Frontiers in Psychology (January 2012), defines presence — the sense that all of this is really happening — as the successful suppression of interoceptive signals by top-down prediction; it locates the key comparator in the anterior insular cortex; its empirical anchor is depersonalisation/derealisation disorder, the condition in which the world feels like it is behind glass. That is what sits underneath the stage line about experiencing how well control is going.
Neither demonstration's provenance is credited. The checkerboard is Edward Adelson's checker-shadow illusion, made at MIT in 1995. The noise is sine-wave speech, from Remez, Rubin, Pisoni & Carrell's 1981 paper in Science: replace an utterance with a three-tone sinusoidal replica and, though it plainly does not sound like speech, listeners can still identify the linguistic message. That paper's conclusion was narrower and more technical — speech perception does not depend on traditional acoustic cues for phonetic segments. Seth plugs it into a framework that did not exist in 1981. This is legitimate; recruiting old phenomena as evidence for new frameworks is ordinary science. But it also plants the objection critics most like to press: if predictive processing can explain any phenomenon at all, what does it forbid?
PARTLY HELD UP · as of 2026-09
The big frame stands: perception-as-prediction is now the common language of cognitive science, and Seth's move of putting interoception and the body at the centre is widely accepted. What has not stood is part of the evidence shown onstage — the rubber-hand experiment propping up the "self as controlled hallucination" half was, three years later, declared badly confounded by a paper carrying Seth's own name, and that argument is still unresolved. Meanwhile "controlled hallucination" travels widely as "reality is fake," which is not what he said.
25 September 2020. Lush, Botan, Scott, Seth, Ward and Dienes published in Nature Communications, testing a sharp hypothesis: if demand characteristics are driving the rubber hand illusion, then people more responsive to imaginative suggestion should experience it more strongly. Across 353 participants, subjective ownership reports correlated with phenomenological control scores at Spearman's rs = 0.26 (95% CI 0.16–0.37), and proprioceptive drift at rs = 0.11 (95% CI 0.003–0.22). The numbers are not large, but the conclusion does not rest on their size. It rests on the fact that this correlation should not exist at all.
Worse were the two companion studies. In one, they described the procedure to people who had never done the experiment and asked what they expected to feel — and the pattern people expected matched the pattern reported in the published literature almost exactly. The control the field has leaned on for over two decades does not control for expectation. In the other, a simulation: standard samples run around 20 participants, and if most of those 20 score high on phenomenological control the experiment almost guarantees finding the illusion; if they all score low it appears about half the time — and for statements explicitly about ownership, only around 4% of the time.
contrasting sync vs async stroking is not sufficient to control for demand characteristics in the RHI.— Seth, Roseboom, Dienes & Lush, "What's up with the rubber hand illusion?", 2020-11-09
The rebuttal came fast, and it came heavy. On 28 January 2022, Ehrsson, Fotopoulou, Radziun, Longo and Tsakiris published a commentary in the same journal, "No specific relationship between hypnotic suggestibility and the rubber hand illusion." Reanalysing the same data, they argued that when quantified according to Lush et al.'s own preregistered hypothesis the relationships are not significant; and that the relationships that do appear also show up in the control condition, indicating a general influence of hypnotisability on cognition rather than anything specific to this illusion. Lush and Seth replied the same day and gave no ground: the asynchronous condition offers no protection against demand characteristics, so using it as a control is not a control at all.
What Seth did not do is worth recording too. He did not say the rubber hand illusion is fake. He and his colleagues drew the line explicitly: "We are not saying that the RHI is 'explained away' by imaginative suggestibility (trait PC). It remains an open question as to whether factors besides implicit suggestion, in the form of demand characteristics, are driving subjective reports and/or proprioceptive drift." That is about the most honest posture a researcher can take towards evidence they themselves have used: leave the question open rather than rushing to close it or to deny it. And open it remains — in 2025 a registered report by Lush and Dienes received in-principle acceptance, to test directly whether suggestion can reverse the illusion.
The framework itself took fire as well. In June 2020, Litwin and Miłkowski argued in Cognitive Science that presenting predictive processing as a unifying account is premature: "PP-based models are seldom empirically validated, and they are frequently offered as mere just-so stories." The objection planted at the end of the last section had been picked up head-on.
Finally, the fate of the phrase. "Controlled hallucination" lost half of itself in transit. What survives is "hallucination," which becomes "reality is fake" — when the controlled part means precisely that the predictions are reined in by the world, a point he nailed down onstage. In a long piece in the New Statesman in February 2022, Sophie McBain records that Seth wishes there were a better word.
As for the line about conscious AI being "pretty remote": in April 2025 he expanded it into a target article in Behavioral and Brain Sciences, "Conscious artificial intelligence and biological naturalism." The position is unchanged, but a door is left ajar — real artificial consciousness is unlikely along current trajectories, and becomes more plausible as AI becomes more brain-like or more life-like.