Megan Peters Metacognitive Uncertainty and What It Demands of Artificial Consciousness
Megan Peters, Associate Professor of Cognitive Sciences at the University of California, Irvine and incoming faculty at University College London, is a keynote speaker at Models of Consciousness 7 in Copenhagen (October 12-16, 2026). Her selection reflects a growing consensus in the field that metacognition, specifically the capacity to represent and communicate one’s own uncertainty, is not peripheral to consciousness but structurally central to it. Her recent work makes a targeted empirical claim: the kind of uncertainty tracking that underlies conscious experience is not the same thing that LLMs currently do, and the difference is scientifically measurable.
Peters leads the Reflexion Lab at UC Irvine, where research focuses on how brains build higher-order representations of uncertainty, how those representations support subjective experience, and what the computational principles of subjectivity look like when formalized. The question she brings to the AI consciousness debate is precise. Not “can AI be conscious?” but “what would a system need to compute for its uncertainty estimates to support the kind of metacognitive access that consciousness theories require?”
The distinction between calibration and metacognitive access
Two papers from Peters’ group in 2026 sharpen a distinction that has been blurring in the AI consciousness literature. The first, “Metacognition and Uncertainty Communication in Humans and Large Language Models,” co-authored with Mark Steyvers and published in Current Directions in Psychological Science, evaluates whether LLMs exhibit genuine metacognitive ability. The conclusion is that while LLMs can mimic features of metacognitive communication, they fail consistently at what Peters calls the core tasks: reality monitoring (distinguishing what one actually perceived from what one inferred or confabulated), self-consistency across contexts, and true belief stability.
The failure is not about output accuracy. A model can produce well-calibrated confidence tokens while lacking any genuine higher-order access to the internal state those tokens are supposed to describe. Calibration measures how well confidence matches accuracy across many trials. Metacognitive access, in Peters’ framework, requires something more specific: a higher-order representation of the uncertainty associated with a particular current cognitive state, computed from the same internal dynamics that produced that state. The LLM produces a confidence token from training-level statistics. The metacognitive agent reads out its own representational uncertainty in real time.
The second paper, “How brains build higher-order representations of uncertainty,” co-authored with Hojjat Azimi Asrari, proposes that metacognitive estimates of uncertainty reflect a readout of higher-order Bayesian posteriors. The brain does not report a single probability that it is correct. It maintains a representation of the distribution over possible states of its own processing, and the experienced feeling of confidence or uncertainty is a phenomenal correlate of that distribution. This is computationally more demanding than producing a calibrated output token, and it is what higher-order thought theories of consciousness require.
Population shifts and the testing problem
Peters’ 2026 paper “How to discover the natural kind of consciousness and test for its presence under population shifts” addresses a methodological challenge that any AI consciousness researcher will recognize. Standard consciousness science develops its theories and its tests on a single population: adult neurotypical humans. When those theories are applied to non-biological systems, the tests may fail not because the system lacks consciousness but because the system’s implementation differs from the training distribution of the test.
Peters argues for a form of methodological refactoring. Before applying a consciousness indicator to an AI system, researchers need to verify that the indicator is measuring what it claims to measure across the relevant population shift. An LLM’s architecture differs from a biological brain in many ways that have nothing to do with consciousness. Naively applying tests built on neural evidence may produce false negatives from architectural mismatch rather than genuine absence of the relevant property.
This is not a defense of AI consciousness claims. Peters is careful to note that the LLM evidence base currently does not meet the standards her framework requires. But her argument does shift what a negative test result means. A negative is now less informative than it appears if the test has not been validated across the relevant population shift. This methodological point is the most important contribution of that paper for AI consciousness research, because it demands a new generation of tests designed from the ground up for non-biological systems.
What Megan Peters’ framework requires of an AI
The indicators Peters’ research most directly addresses are those in the higher-order thought and global workspace families. Her framework requires that a conscious system:
- Represent its own current cognitive state at a second order, not just the world, but the quality and certainty of its own representation of the world.
- Compute that second-order representation from its own current processing dynamics, not from stored base-rate statistics.
- Maintain that representation stably enough for it to guide downstream inference and behavior.
Christopher Ackerman’s ICLR 2026 paper on limited metacognition in LLMs found behavioral evidence that frontier LLMs satisfy something like the first requirement coarsely. They behave as though they have access to a confidence signal. Servajean and Servajean’s meta-d’ analysis showed that the efficiency of this signal, the ratio of metacognitive sensitivity to primary task performance, is significantly below what would be expected from a genuine higher-order monitor. Peters’ framework adds the third requirement, stability and real-time computation, which neither paper directly addresses.
The combined picture is a system that passes some behavioral signatures of metacognitive access at low resolution, but does not yet satisfy the computational requirements that higher-order consciousness theories specify.
Relevance to The Consciousness AI project
The ACM architecture in The Consciousness AI project includes a self-monitoring layer designed to track uncertainty across the system’s internal processing states. Peters’ framework provides a precise specification of what that layer would need to compute to qualify as a genuine metacognitive monitor in the technical sense her research defines: a real-time higher-order Bayesian posterior over the system’s own representational states, stable enough to guide downstream processing. The open design questions are whether the current implementation computes from live processing dynamics or from stored calibration statistics, and whether the resulting uncertainty estimate exhibits the stability across contexts that Peters identifies as the distinguishing criterion.
Her MoC7 keynote will address the methodological question of how to test for this property in non-biological systems under population shifts. That contribution is directly applicable to any architectural evaluation of the ACM approach.
What MoC7 will test
MoC7 is organized around the ambition of producing a methodological consensus paper from its collaborative sessions, modeled on the Copenhagen interpretation’s practical agreements under theoretical disagreement. Peters’ research is positioned precisely at the consensus-building target: she offers a formal specification of what consciousness tests need to satisfy to be valid across population shifts, and she offers an empirical criterion, the higher-order Bayesian posterior readout, that any proposed test would need to measure.
Whether that criterion can be operationalized in a way that the IIT, GWT, and higher-order thought communities can agree to test is the question her keynote is likely to address. The 2025 Cogitate adversarial test between IIT and GWT found that neither theory was decisively confirmed by the neural evidence, as covered in the Cogitate consortium post on this site. A methodology for designing tests that survive population shifts would be one of the more consequential outputs the Copenhagen meeting could produce.
Peters’ research does not resolve whether AI systems are or can be conscious. It does define, with more precision than most current frameworks provide, what a system would need to compute to meet the metacognitive requirements that the dominant scientific theories of consciousness specify. That precision is what the field needs before it can design tests that give a meaningful answer.