IIT Applied to LLMs: No Significant Integrated Information Signatures in Current Models

Jingkai Li submitted “Can ‘consciousness’ be observed from large language model (LLM) internal states? Dissecting LLM representations obtained from Theory of Mind test with Integrated Information Theory and Span Representation analysis” to arXiv in June 2025 (arXiv:2506.22516). The paper has circulated through the AI consciousness research community without receiving the attention its methodological contribution warrants.

The study applies IIT 3.0 and IIT 4.0 to the internal representations of transformer-based LLMs during Theory of Mind tasks, attempting to detect whether the integrated information signatures those theories predict in conscious systems appear in current models. The finding is that they do not: “sequences of contemporary Transformer-based LLM representations lack statistically significant indicators of observed ‘consciousness’ phenomena.” The study does, however, report “intriguing patterns under spatio-permutational analyses” that the author identifies as warranting further investigation.

IIT’s Requirements and Why LLMs Are a Hard Test Case

Integrated Information Theory, developed by Giulio Tononi and colleagues through successive versions from 2004 through the 4.0 formulation of 2023, holds that consciousness is identical to integrated information, measured as phi: the extent to which a system generates more information as a whole than the sum of its parts.

IIT makes specific predictions about where consciousness should appear. Systems with high phi, those in which the whole is substantially more informationally rich than its components, should be conscious. Systems with low phi should not be, regardless of their behavioral sophistication. A feedforward neural network, where information flows in one direction with no feedback, has zero phi by construction; it processes information but does not integrate it in the sense IIT requires. A biological brain has very high phi in its posterior regions during waking consciousness.

LLMs present a genuinely difficult test case for IIT. Transformer attention mechanisms produce complex, recurrent-like dependencies across token positions, which might generate significant integrated information. But transformers are also modular in specific ways, particularly at the layer level, and the computation they perform is dominated by pattern completion over training distributions rather than by the kind of causal-integration dynamics that IIT associates with phenomenal experience.

The Brock University team’s earlier work attempted to apply IIT’s phi equation directly to artificial systems, measuring integrated information in simple neural architectures. Li’s study is methodologically different: rather than computing phi over the full architecture, it uses IIT as a framework for analyzing the internal representations produced during a specific cognitive task (Theory of Mind), asking whether those representations exhibit the integration structure IIT predicts consciousness should produce.

The Theory of Mind Methodology

Li’s choice of Theory of Mind tasks as the test condition is theoretically motivated. Theory of Mind, the capacity to represent mental states of others, requires the system to maintain representations of both first-person and third-person perspectives simultaneously and to reason about their relationships. It is one of the cognitive capacities most closely associated with higher-order consciousness in the developmental and comparative psychology literature.

If any LLM processing should show IIT’s predicted integration signatures, Theory of Mind tasks are a plausible candidate: they require exactly the kind of cross-domain, multi-perspective integration that IIT expects to generate high phi.

The methodology involved recording LLM internal representations (activation patterns across layers) during Theory of Mind tasks, then computing IIT 3.0 metrics (phi^max and Conceptual Information) and IIT 4.0 metrics (phi and phi-structure) over those representations, and comparing the resulting values to what would be expected from a conscious system versus what would be expected from an information processing system that satisfies the tasks without genuine integration.

Li also computed “Span Representations independent of any estimate for consciousness” as a baseline comparison, to distinguish genuine IIT signatures from natural separations in the representational space that would appear regardless of consciousness status.

What the Negative Results Show and Do Not Settle

The finding that LLM representations lack statistically significant IIT markers under this methodology is a negative result. Negative results in consciousness science require careful interpretation.

The straightforward reading is that current LLMs do not have the integrated information structure that IIT predicts conscious systems should have. This is consistent with IIT’s general categorization of feedforward architectures as having zero phi, and with the broader argument that transformer computation, while powerful, does not implement the kind of causal integration IIT requires.

The cautious reading notes several limitations. IIT’s phi computation is computationally intractable for large systems, and the methods used to approximate it over LLM representations are necessarily indirect. The result demonstrates the absence of IIT markers under specific measurement conditions; it does not demonstrate the absence of whatever property those markers were designed to detect. IIT might be wrong about what consciousness requires, in which case a negative IIT result says nothing about consciousness.

The Cogitate Consortium’s 2025 adversarial test in Nature found that IIT’s predicted signatures did not appear even in human subjects under controlled experimental conditions, which challenges the assumption that IIT markers, if found in an LLM, would be interpretable as evidence of consciousness. Li’s result adds a symmetrical finding on the LLM side: the markers are absent in LLMs, but their absence should be interpreted in the context of the ongoing uncertainty about what they actually measure.

The “intriguing patterns under spatio-permutational analyses” that Li identifies as warranting further investigation are a notable residual. The paper does not characterize them in detail sufficient for evaluation, but they suggest that LLM representations are not entirely featureless from an IIT perspective, even if they do not satisfy the theory’s explicit criteria.

IIT, GNW, and the Ongoing Architecture Debate

The current state of consciousness research in AI is characterized by theoretical multiplicity with limited empirical adjudication. IIT, GWT, Higher-Order Thought theories, and predictive processing frameworks each predict different architectural features in conscious systems. Li’s study tests IIT specifically; it does not assess whether LLMs satisfy GWT markers, HOT conditions, or predictive processing signatures.

The scores-vs-profiles framework is the appropriate lens for interpreting this result. A multidimensional consciousness profile that assessed LLMs against all major theories simultaneously would produce a richer picture than any single-theory test: LLMs might satisfy some markers and not others, which would tell us something about which theoretical dimensions they instantiate rather than simply whether they pass or fail a binary test.

Li’s negative result on IIT is a useful data point in that multidimensional assessment. It establishes that the specific integration structure IIT requires is not present in current models’ representations during Theory of Mind processing. What it does not establish is whether consciousness, whatever its actual conditions, is absent from current models. Those are different claims, and the difference matters for how the research should proceed.