Liad Mudrik Prediction Maps and Testing Consciousness Theories at MoC7 2026
Liad Mudrik will give a talk at Models of Consciousness 7 (MoC7) titled “Using prediction maps to test theories of consciousness.” The talk runs 40 minutes with 10 minutes for questions, at the University of Copenhagen, October 12 to 16, 2026. Prediction maps are the next step in the adversarial collaboration method Mudrik brought to CS26, and the slide title names the specific tool she will present. A prediction map records where and when a theory predicts a measurable neural signature, in spatial and temporal terms, before the data is collected.
Mudrik is professor of psychology at Tel Aviv University and co-leader of the Cogitate Consortium, the largest open science initiative in the field. The MoC7 programme confirms her as one of eight invited speakers, all of whom feed a collective consensus paper that the conference is attempting for the first time. Her contribution to that paper is the empirical method that distinguishes testing from asserting.
What a prediction map makes explicit
The core problem with most theory testing is that competing theories rarely make claims precise enough to collide. Global Neuronal Workspace theory predicts a broadcast across frontal and parietal regions at the moment a stimulus becomes available to report. Integrated Information Theory claims a sustained posterior spin, a persisting state that outlasts the brief frontal surge. Spoken at that level, the two predictions do not cleanly contradict each other, and each laboratory finds the evidence its own theory expects.
A prediction map removes the verbal hedge. It fixes a coordinate system on the brain and on the time axis, and assigns each theory a load-bearing claim at specific points in both. One theory predicts early posterior activation, another predicts late frontal broadcast, and the map places the predictions on the same grid. Once the grid exists, one experiment can pass or fail several theories at once. This is the adversarial collaboration logic already documented for Mudrik’s CS26 address, and the prediction map is its graphical form.
The intellectual ancestry is the ConTrast database, published by Yaron, Melloni, Pitts and Mudrik in Nature Human Behaviour in 2022. The database records, experiment by experiment, what each theory of consciousness predicts and what the data actually showed. Over 500 experiments are indexed. The database answers a descriptive question, which theories are being tested and how. The prediction map tool answers the prescriptive question, what a fair test of each theory would have to measure and where.
Why the Cogitate result made the tool necessary
The first Cogitate adversarial collaboration, published in Nature in 2025, tested IIT against GNW on the same paradigm across multiple sites. The result was mixed for both. Frontal activity did not appear necessary for conscious perception, which weakens the workspace timing claim, while the late onset of activity rather than early onset tracked consciousness, which does not match the IIT timing prediction. Neither theory was eliminated, and the field read the result as ambiguous.
Mudrik’s position, stated in her CS26 address and in the Cogitate commentary, is that the ambiguity is partly a measurement artifact. Each theory was tested against the assumption it needed and not against the claims it could not afford to make. A prediction map forces both sides to commit to a specific spatial and temporal signature before the next experiment runs. The point is not to make the theories falsifiable in principle, which they already were. The point is to make their differences visible in the same units, so an experiment cannot be read one way by one camp and another way by the other.
The MoC7 setting is tailored to that tool. The conference is run by the Association for Mathematical Consciousness Science, and its stated aim is methodological consensus. A prediction map is the device that converts theoretical disagreement into a common diagram, which is exactly the currency a consensus paper needs.
What this requires of an AI consciousness test
The AI consequence is direct. The 2026 state of the field overview documents that most evaluations of language models rely on behavioral tests or conversational probes that any theory can interpret its own way. A test built on prediction maps would specify, in advance, which internal representations a model would need to produce and where in its processing stack they would appear, before the evaluation runs.
This is harder for artificial systems in one specific respect. A biological brain has a shared anatomy that fixes the reference frame for a prediction map. Two labs can agree on where the posterior cortex is. A language model has no anatomy that every implementation shares, and different models organize the same function in different layers. The prediction map method therefore requires a prior decision about what the coordinate system is, and that decision is itself a theoretical commitment.
Mudrik’s existing work on testing validity anticipates this difficulty. Her 2024 paper in Neuron with Ron Hirschhorn and Uria Korisky argued for increasing the ecological validity of consciousness experiments, moving from contrived laboratory stimuli toward situations closer to real cognition. The same instinct applies to AI testing. A model tested on artificially constructed probes may fail to show signatures that its normal processing does produce, and an ecological test is the one most likely to reveal them. The overlap with the methodological criteria recommended for AI evaluation is laid out in the 14 indicator checklist researchers apply.
Where this sits
Mudrik’s MoC7 talk converts the site’s existing record of her work into a concrete next experiment. The site has already reported the adversarial collaboration at CS26 in detail, and the meaning is continuous. The prediction map is not a new theory of consciousness. It is a measurement discipline that any theory can be subjected to, and that is precisely why the consensus session can use it.
A reader at MoC7 will be able to judge the tool by its most visible test. The Larissa Albantakis and Matteo Grasso analysis of theory-kept assumptions on this site documents what happens when formal theories are evaluated against their own preferred evidence rather than against their load-bearing claims. Prediction maps are the methodological answer to that pattern.
The measure of the talk
The test of the prediction maps talk is simple. It will have succeeded if, by the end of the conference, at least one rival pair of theories has committed to a shared prediction map and a shared experiment that the map defines. Registration for MoC7 closes August 31, 2026. The conference site is amcs-community.org/events/moc7-2026/. The earlier profile of Mudrik’s adversarial collaboration and its application to AI sets out the full context of her method, and the wider 2026 research field is in the consciousness research overview.
Liad Mudrik will deliver “Using prediction maps to test theories of consciousness” at Models of Consciousness 7 at the HC Ørsted Institute, University of Copenhagen, October 12-16, 2026, per the confirmed AMCS programme. The ConTrast database reference is Yaron, Melloni, Pitts and Mudrik, Nature Human Behaviour, 2022, DOI 10.1038/s41562-021-01284-5. The ecological validity paper is Mudrik, Hirschhorn and Korisky, Neuron 112(10): 1642-1656, 2024, DOI 10.1016/j.neuron.2024.03.031.