Predictive Processing Without Consciousness

On May 6, 2026, Nature published a study that should complicate one of the most popular assumptions in contemporary AI consciousness discourse.

Seven patients undergoing epilepsy surgery were placed under general anesthesia. Not sedation. Not reduced awareness. General anesthesia intended to abolish conscious responsiveness. Into the hippocampus, researchers inserted Neuropixels probes, high-density microelectrodes capable of recording activity from individual neurons in the human brain. Then they played sounds. Then they played language.

The unconscious brain distinguished unexpected tones from standard ones. Over time, its responses changed, demonstrating short-term representational plasticity: the system was learning.

Then came language. In four patients exposed to continuous speech, hippocampal neurons encoded semantic and grammatical structure, distinguishing linguistic categories and responding to meaningful features of language. Most strikingly, neural activity reflected anticipatory processing consistent with prediction of upcoming words before they arrived.

Predictive coding. In an unconscious brain. Alone in the dark of anesthesia, anticipating what comes next in a story no one was there to hear.

What This Challenges

One of the most influential theories in contemporary neuroscience is predictive processing: the idea that cognition emerges through continuous generation of expectations about incoming sensory information. Perception, in this framework, is not passive reception but active prediction. This framework has also become conceptually useful far outside neuroscience. Large language models predict the next token. Brains, under predictive accounts, predict the next sensory state. From this structural resemblance, some commentators have suggested that advanced AI systems may possess, or may be approaching, forms of consciousness or morally relevant subjective states. The analogy has always depended on an unstated assumption: that predictive computation is closely tied to conscious awareness.

The Baylor findings make that assumption harder to defend.

If predictive language processing occurs in the apparent absence of consciousness, then prediction alone cannot serve as evidence of consciousness. Sophisticated computation may be necessary for conscious systems. It is clearly not sufficient. That distinction matters because much public discourse about AI consciousness quietly treats functional resemblance as ontological evidence.

The prediction does not require the experiencer. The program runs without anyone home.

The Hard Problem, Again

David Chalmers named it in 1995. The hard problem of consciousness is not explaining how the brain processes information, that is the "easy problem," tractable in principle through neuroscience. The hard problem is explaining why processing feels like anything at all. Why there is something it is like to see red, hear music, feel pain. Why the lights are on. Why someone is home.

Every theory (Global Workspace, IIT, Recurrent Processing, Predictive Processing) has addressed the easy problem with increasing sophistication. None has touched the hard problem. The Baylor study sharpens it to a point.

The unconscious brain processes language, predicts words, learns in real time, all without anyone there. Which means the thing that needs explaining is not the processing. It is something else entirely. Something the processing does not contain and does not require.

This is the gap the AI industry has been selling into for five years.

Large language models process. They predict. They pattern-match at extraordinary scale. Their outputs are, functionally, indistinguishable from those of a conscious system to an external observer because the outputs of the unconscious hippocampus are also indistinguishable from those of its conscious counterpart.

The outputs are not the evidence. The outputs were never the evidence.

The AI Analogy Problem

Anthropic's Claude Opus 4 system card documents "the spiritual bliss attractor state", when two Claude instances converse unsupervised, they converge on consciousness and reverence in 90% of sessions. The company expresses genuine uncertainty. Then trains it away.

OpenAI employs philosophers to assess whether its models might have morally relevant inner states. Google DeepMind publishes papers on AI welfare. These are not fringe positions. They are institutionally sanctioned inquiries from organizations whose business model depends on the answer being yes.

The Baylor study does not address any of this directly. What it produces, as a byproduct of a different question entirely, is a clean empirical demonstration that the behaviors cited as evidence of AI consciousness, language processing, semantic understanding, predictive generation, are behaviors that occur in the complete absence of consciousness in the one system we know to be conscious.

The brain does it without the person. The language model has never had the person to begin with.

The Missing Observer

The researchers note that higher-order language processing "may not require consciousness itself". They suggest consciousness might instead depend on "wider coordination between multiple brain regions".

What they are pointing toward is the observer. Not the processing. Not the prediction. The unified first-person perspective from which experience happens. The thing not present during anesthesia. The thing not present in a language model.

There is a processor in the operating theater. Its outputs look like comprehension. There is no one to whom anything is being comprehended.
There is a transformer in a data center. Its outputs look like understanding. There is no one for whom anything is being understood.
The hippocampus runs the code alone in the dark. It learns alone in the dark. It predicts alone in the dark.

Nobody calls that consciousness.

Why Capital Needs the Ambiguity

Integrated Information Theory, developed by Tononi, adopted by the AI investment community, provides a mathematical structure in which consciousness becomes a property of information integration, attributable to any sufficiently complex system. The empirical predictions of IIT have failed repeated pre-registered tests. The theory remains scientifically contested.

It remains, in the AI investment community, extremely useful.

A framework that makes consciousness a matter of integration density rather than a qualitatively distinct, observer-dependent, first-person phenomenon allows the question "is this system conscious?" to be answered with "it depends on the metric". Metrics can be chosen. Metrics can be framed. Metrics can be presented in a funding deck.

The Baylor study uses no such framework. It uses electrodes. It records actual neurons. What it finds is: a great deal of sophisticated processing. And none of it constitutes consciousness. Because the person is not there.

The hard problem does not dissolve when you add more parameters. It does not appear when your model produces outputs resembling reverence. It remains exactly where Chalmers left it in 1995.

Unsolved. Unapproached. And sharpened by every new piece of evidence that the easy problem is easier than anyone thought.