Source
- Vishne, Gerber, Knight & Deouell (2023), LIT-644, read in full in NOTE-499: Results (Figs. 1–5), Discussion, Limitations and the STAR Methods on stimuli, HFA, state-space analysis, decoding and statistics.
What was actually shown
The paradigm. Vishne et al. (LIT-644) recorded subdural electrodes in 10 patients with intractable epilepsy (8 right hemisphere only) while they viewed clearly visible greyscale images of faces, watches, other objects and animals for 300, 900 or 1,500 ms (three patients also saw 600 and 1,200 ms). Patients responded only to rare targets, about 10% of trials, which were not analysed, so the analysed trials needed no report (though each image had to be judged a non-target). The signal is high-frequency activity (HFA, 70–150 Hz), a proxy for local firing. Of 907 noise-free electrodes, 430 were visually responsive by a criterion blind to category and timing.
Amplitude decays. With the image still on, HFA fell from peak by 81.8% ± 1.1% by 800–900 ms across occipital, ventral temporal, parietal and prefrontal electrodes, and single-electrode category selectivity by 77.4% ± 1.1% (Fig. 1).
Content does not. Linear decoders on the pattern across electrodes separated face from watch in ventral temporal cortex at a mean AUC of 99.8% over 100–900 ms (occipital 92.4%). A decoder trained at onset worked at the end and the reverse: the temporal generalisation matrix in ventral temporal cortex is a square block, so the discriminating direction in population space is stable in time. Information ended about 450 ms after offset for 300 ms images and 230 ms for the longer ones; the duration contrasts were significant in ventral temporal and occipital cortex and not in prefrontal or parietal cortex (Fig. 4). Exemplar identity and the geometry of exemplars were sustained and stable in ventral temporal cortex through 900 ms, also after category structure was removed, on 60 images seen by 5 patients (Fig. 5). Occipital results were in the same direction and less robust.
Prefrontal and parietal content is an onset burst. Category decoding there was real (peak AUC 84.1% and 86%) but “transient and mostly limited to ~150–600 ms after stimulus onset”, later than in visual cortex, did not generalise across time and did not track duration. GNWT’s predicted offset ignition was not seen: the prefrontal event is at onset only, and nothing marked the image going off.
What could have come out otherwise: ventral temporal decoding could have decayed with amplitude, or survived as a changing sequence of codes (a diagonal generalisation matrix), or ended at a fixed time after onset rather than after offset. Prefrontal decoding could have lasted as long as the image. None did.
Bearing on IIT and the global workspace
The multi-duration paradigm was adopted by the COGITATE adversarial collaboration, whose pre-registered predictions the paper tests. As the paper reports them, IIT’s proponents predicted a stable representation in a posterior “hot zone” lasting as long as the experience, and were “largely agnostic” about prefrontal cortex; the global neuronal workspace’s proponents predicted transient prefrontal onset and offset representations with no sustained component, even without report. Both held, except the offset ignition, and the authors conclude that “these predictions of IIT and GNWT are, in fact, not adversarial”.
That is a result about two sets of paradigm predictions, not about the theories’ formal content. IIT 3.0 (LIT-445, read in NOTE-365) derives consciousness from integrated information in a system’s cause–effect structure, and nothing here computes Φ or tests a postulate; the “posterior hot zone” prediction was made by Tononi et al. for COGITATE, not derived in LIT-445. The workspace side is the theory whose indicators Butlin et al. (LIT-056) adopt, and its ignition prediction is what held. So the paper does not move the record’s standing on either theory. What it shows is that a paradigm chosen to tell them apart did not, because each predicted something about a different region and both were right.
THEORY-023. That account holds that current evidence cannot settle whether an AI system is conscious, partly because architectural indicators presuppose the disputed theory, so the dispute is over what counts as evidence. Its claim is about AI and is neither supported nor contradicted here. The bearing is narrower: even in humans, with intracranial recordings and pre-registered predictions, a test designed to discriminate two theories found them compatible. That is a reason to expect slow progress from indicator-style evidence, not a test of THEORY-023’s argument, and no relation is declared.
The paper’s own reading is conditional. If experience of a steady image is continuous, the sustained ventral temporal code fits it; if experience is discrete, the transient frontoparietal burst fits it better. Which region is the correlate depends on the temporal nature of experience, which the data do not settle. Dennett and Kinsbourne’s Time and the Observer (LIT-431, read in NOTE-367) is the paper’s second reference for that question; nothing here tests their argument.
What this does not say
- It does not say the code is the percept. Every image was clearly seen and stayed the same, so percept and input never came apart. The same code could carry an unseen image. The authors name binocular rivalry as the test.
- It does not say prefrontal cortex holds no sustained content. Prefrontal coverage was sparse and mostly one hemisphere, and only HFA was measured; low-frequency activity or synaptic states could hold it. The authors say the absence “should be taken with more caution than the presence of activity”.
- It does not say what the “experience subspace” does. The authors use “representation” for patterns that correlate with stimulus descriptors, with no mechanistic role implied. Their proposal that position in a stable subspace fixes perceptual quality is speculative.
- It is not single-neuron evidence. HFA is a population proxy, and the main multivariate analyses pool electrodes across patients into one pseudo-population per region.
- It does not reach beyond 1,500 ms, or say how the code relates to representational drift over longer times, which the authors leave open.
Connections
- Verzhbinsky et al. (LIT-633, THEORY-097). Also human intracranial high-frequency signals, but about co-occurrence across regions, not content within one. Prefrontal content arriving after sensory content and lasting a few hundred milliseconds is the kind of event a cross-region coordination mechanism would have to carry; neither paper tests the link.
- THEORY-100. Filed alongside: in crows reporting near-threshold stimuli, neurons in the prefrontal analogue followed the reported percept rather than the stimulus (second-hand, through LIT-650). That is the dissociation of percept from input this account lacks, but it was found in a report paradigm and in the frontal analogue, where here, without report, frontal content was only an onset burst. The two do not conflict; report may be what recruits sustained frontal coding. Neither tests that.
Note of 2026-10-12: bearing on THEORY-096’s wording
This finding meets the onset prediction the workspace’s proponents made, and THEORY-096 states the workspace account. But THEORY-096’s statement says ignition “sustains” the content. A sustained prefrontal code was not found here, and the proponents did not predict one. A note on THEORY-096 records the gap. No relation is declared.
Note of 2026-10-14: COGITATE replicates it for category, and narrows the title
COGITATE (LIT-973, read in NOTE-812) is the multi-duration result the first kind of promote_when names: an independent intracranial cohort with 576 prefrontal electrodes. It finds a sustained posterior category representation, and prefrontal category content confined to onset with nothing at offset. It does not hold for every dimension. Face identity was not sustained in posterior cortex, and face orientation decayed after about 0.5 s though the image stayed on. The title said “the content” of the image; it now says “the category and object identity”, which is what both cohorts show. The first kind of promote_when is partly met: replicated for category, not for every dimension. The second kind, percept and input apart, is untouched, since every image was seen. Only 15% of face-selective electrodes were sustained, which fits “amplitude decays”, but COGITATE did not test a stable population code against decaying amplitude as Vishne et al. did.
Narrowed and promoted (2026-10-21)
Batch 076 of the promotion audit narrowed the title to Vishne et al.’s measurement and promoted it under ADR-047 v3: one study read in full, with the decoding and amplitude figures stated in its text. The general reading the old title carried, that the content of a sustained seen image lives in a stable sensory code while frontoparietal cortex marks only its onset, is held as THEORY-830, Proposed, with this entry’s old promote_when. COGITATE (LIT-973, note of 2026-10-14 above) replicates the category result in an independent cohort but finds face identity not sustained in posterior cortex. That limits the general claim, not this measurement, whose exemplar result is stated for its own 60 images and five patients.