Liad Mudrik Adversarial Collaboration and Empirical Theory Testing at CS26 2026

Liad Mudrik, professor of cognitive neuroscience at Tel Aviv University and principal investigator of the Cogitate Consortium, presented a major address at Consciousness Science 2026 in San Diego on the future of theory testing in the science of consciousness. Mudrik has spent the past seven years leading the largest open science initiative in the field, organizing direct adversarial collaborations designed to resolve long-standing disputes between competing neuroscientific accounts of subjective experience.

Her address highlighted an acute methodological crisis in cognitive science. Over the past three decades, dozens of distinct theories of consciousness have accumulated, each supported by an independent body of confirmatory experiments published in peer-reviewed journals. Because individual laboratories typically design paradigms customized to their favored framework, experiments rarely eliminate rival hypotheses. Mudrik’s adversarial collaboration methodology, detailed in Science (2024, DOI: 10.1126/science.adl2737) and demonstrated in the landmark Cogitate Phase 1 results published in Nature (2023, DOI: 10.1038/s41586-023-05748-7), forces theoretical proponents to preregister mutually agreed upon experimental designs, point predictions, and interpretation criteria before data collection begins.

The mechanics of adversarial collaboration

Traditional scientific progress in psychology and neuroscience often follows an uncoordinated pattern. A laboratory supporting Integrated Information Theory (IIT) designs an experiment emphasizing posterior cortical synchrony, while a laboratory supporting Global Neuronal Workspace theory (GNW) builds a visual masking task highlighting prefrontal ignition. Both publish positive findings, leaving the broader scientific community with contradictory claims and no clear adjudication mechanism.

Mudrik’s framework introduces five structural constraints that alter the dynamic of experimental testing. Proponents of competing theories work with neutral arbiters to design shared protocols that test conflicting predictions on identical tasks. Multiple independent data collection sites execute the experiments using standardized equipment, blinding experimenters to trial conditions. Analysis pipelines are fully preregistered and locked before unblinding. When results disagree with prior commitments, theorists are obligated by the preregistered agreement to acknowledge specific empirical anomalies rather than shifting definitions post hoc.

In the initial Cogitate investigation, teams led by Stanislas Dehaene (GNW) and Giulio Tononi (IIT) agreed to test visual awareness using combined functional magnetic resonance imaging (fMRI), magnetoencephalography (MEG), and intracranial electroencephalography (iEEG). The outcomes produced significant challenges for both camps. GNW predicted that prefrontal ignition would track conscious perception regardless of task relevance, but frontal activations diminished when stimuli were task-irrelevant. IIT predicted sustained synchronization across posterior sensory areas throughout the duration of a conscious percept, yet sustained signals were observed only transiently.

Comparing adversarial protocols with standard research paradigms

The shift from solitary laboratory investigation to multi-site adversarial testing represents a substantial structural evolution in empirical rigor.

The comparative structure in the table demonstrates why this approach is gaining traction across cognitive science. By removing the latitude for selective reporting, the methodology converts theoretical disagreements into quantifiable empirical milestones.

Extending adversarial testing to artificial intelligence

In the flagship overview of consciousness science 2026, the proliferation of unverifiable claims regarding machine consciousness is identified as a primary hazard for the field. Many evaluations of large language models rely on anecdotal prompting or anthropomorphic conversational behavior, which provide no genuine epistemic validity.

Mudrik’s CS26 presentation argued that the artificial intelligence research community urgently requires its own adversarial collaboration protocols. Current debates over whether transformer models exhibit functional analogs of global workspaces or higher-order monitoring suffer from the same confirmation bias that plagued cognitive neuroscience. A developer claiming an architecture possesses proto-conscious features can easily construct prompts or behavioral tests that appear to demonstrate agency, while a skeptic can devise simple counter-examples showing failure modes.

An adversarial framework for AI would require AI research laboratories and independent philosophers of mind to establish preregistered evaluation batteries. Such batteries would subject model internals to mechanistic interpretability probes under agreed upon activation conditions, testing whether internal representations meet specific mathematical criteria for information integration, continuous recurrent feedback, or reality monitoring before inspecting test results.

The open-source research initiative documented at github.com/tlcdv/the_consciousness_ai aligns with this emphasis on open, verifiable standards. Investigating whether artificial systems instantiate the specific functional properties posited by neurobiological theories requires transparent, reproducible testing rather than proprietary self-evaluations.

Epistemic challenges and theoretical plasticity

During the plenary discussion at CS26, philosophers and neuroscientists examined the limits of adversarial collaboration. One persistent challenge is theoretical plasticity. When empirical findings contradict a core prediction, theorists rarely abandon their broader framework. Instead, they introduce auxiliary hypotheses, arguing that the specific experimental parameter failed to capture the true operational conditions of their theory.

Mudrik acknowledged this limitation, observing that no single adversarial experiment will instantly resolve foundational philosophical questions. However, she emphasized that forcing theories to accumulate auxiliary modifications increases their complexity penalty. Over multiple rounds of testing, frameworks that continually alter their parameters to explain away disconfirming evidence lose scientific parsimony relative to models whose core predictions survive unaltered.

A second challenge involves construct validity. Proponents of different theories often define consciousness differently. GNW emphasizes conscious access and verbal reportability, while IIT emphasizes phenomenal presence independent of behavioral report. Adversarial protocols must carefully delineate which specific definition is being evaluated to avoid talking past one another.

Future directions for Cogitate and cognitive science

The Cogitate Consortium is currently advancing Phase 2 of its research agenda, expanding testing to incorporate Predictive Processing models and Higher-Order Thought frameworks. This expanded scope addresses how predictive error signaling interacts with recurrent workspace dynamics during perceptual transitions.

Mudrik’s presentation at CS26 made clear that the value of adversarial collaboration extends beyond specific experimental outcomes. It establishes a sociological and scientific model for conducting inquiry in contested domains. For the study of consciousness, whether in biological organisms or computational architectures, the transition from partisan defense of isolated theories to collaborative empirical elimination marks an indispensable step toward scientific maturity.