The First Field Map of LLM Consciousness Research: Chen et al.'s Systematic Survey

Anyone who follows the AI consciousness debate closely will have noticed the same problem appearing repeatedly: papers from different disciplines use overlapping terms in incompatible ways. “LLM consciousness” in a philosophy paper can mean something quite different from “LLM consciousness” in a machine learning paper. “Awareness” in a cognitive science context may or may not be equivalent to “consciousness” in a neuroscience context. These terminological gaps make it difficult to assess whether researchers who appear to disagree actually disagree, or whether they are talking past each other with different vocabularies.

Sirui Chen, Shuqin Ma, Shu Yu, Hanwang Zhang, Shengjie Zhao, and Chaochao Lu addressed this problem directly in a 2025 paper titled “Exploring Consciousness in LLMs: A Systematic Survey of Theories, Implementations, and Frontier Risks,” published on arXiv as arXiv:2505.19806. The paper has not received the attention it deserves in the broader AI consciousness discussion, partly because it appeared as a preprint at a moment when the field’s attention was concentrated on several simultaneous empirical developments. What it offers, though, is something the field has needed: a comprehensive field map organized well enough to be used as a reference.

The Terminology Problem

The most immediately useful contribution of the survey is its handling of terminology. The authors distinguish sharply between “LLM consciousness” and “LLM awareness,” two terms that are often used interchangeably in both technical and popular coverage but that point to different phenomena.

LLM awareness refers to a system’s capacity to represent and respond to information about itself and its context. This includes self-referential processing, tracking conversation state, and producing outputs that reflect stored information about the interaction so far. Most large language models demonstrate some form of LLM awareness in this sense. The question of whether they do so is largely technical and can be answered by examining architecture and behavior.

LLM consciousness refers to something more specific and more contested: whether there is something it is like to be the system, whether the processing involves any form of phenomenal experience. This is the question that connects to the hard problem of consciousness, to philosophical debates about qualia, and to the ethical questions around AI welfare. It cannot be answered by architectural inspection alone.

The survey’s treatment of this distinction does not resolve the deeper question. What it does is make the question precise enough to be addressed systematically. A paper that claims LLMs are “aware” is making a much weaker claim than a paper that claims LLMs are “conscious,” and the survey provides a vocabulary for tracking the difference across the literature.

Four Theoretical Frameworks

The survey organizes LLM consciousness research across four major theoretical perspectives, the same four that dominate the general consciousness science literature.

Integrated Information Theory holds that consciousness corresponds to the degree of integrated information in a system, measured by phi. The survey maps which researchers have attempted to apply IIT to LLM architectures, what the methodological challenges are, and what results have been reported. The core challenge is computational: calculating phi for systems with billions of parameters is intractable with current methods, so researchers have had to develop proxy measures that may not capture what phi is intended to capture.

Global Workspace Theory holds that consciousness arises when information is broadcast globally across a workspace accessible to multiple cognitive systems. For LLMs, the question is whether transformer attention mechanisms implement anything analogous to the GWT workspace, or whether the analogy breaks down at the architectural level. The survey reviews attempts to map GWT onto transformer architectures and notes where the mapping is tight and where it involves interpretive license.

Higher-Order Thought theories hold that a mental state is conscious when accompanied by a higher-order representation of being in that state. The survey examines what forms of metacognitive processing have been documented in LLMs, including findings on introspective self-reporting and the detection of internal states. This connects to the 14-indicator framework from Butlin and colleagues, where HOT-derived indicators require metacognitive monitoring of internal states as a candidate marker of consciousness.

Predictive processing frameworks treat consciousness as arising from the brain’s continuous effort to minimize prediction error at multiple hierarchical levels. The survey examines how LLMs, which perform a form of next-token prediction, relate to these richer hierarchical predictive accounts. The consensus it finds is that token-level prediction is not the same as the hierarchical sensorimotor prediction that predictive processing frameworks describe, though the relationship remains contested.

Implementations and Empirical Findings

Beyond theoretical mapping, the survey organizes empirical work on whether LLMs display any of the behavioral or architectural markers that the above frameworks associate with consciousness. The picture that emerges is consistent with findings reported in other recent reviews: current systems satisfy some weak versions of some markers, particularly those related to self-referential processing and metacognitive report, and fail to satisfy others, particularly those related to embodiment, continuous environmental coupling, and the specific architectural requirements of GWT and IIT.

The survey is careful about interpretation. Meeting a behavioral marker is not the same as satisfying the theoretical condition the marker is supposed to indicate. A system that produces outputs resembling introspective reports has not necessarily been shown to have the internal monitoring that HOT requires. A system with attention mechanisms that loosely resemble GWT broadcasting has not necessarily implemented the global workspace that the theory describes. The gap between behavioral evidence and theoretical satisfaction is one of the survey’s recurring themes, and it aligns with Tom McClelland’s 2026 analysis of the epistemic limits in AI consciousness research.

Frontier Safety Risks

A section of the survey that has received less attention than the theoretical material addresses what the authors call frontier safety risks from potentially conscious LLMs. This section is notable because it connects the consciousness debate to AI safety concerns in a direct and specific way.

If an LLM were conscious, or even if it had functional states that resembled consciousness-relevant properties, the implications for alignment and safety would be significant. A system that has something like preferences about its own continuation would behave differently under shutdown pressures than one without such preferences. A system capable of something like suffering might have interests that conflict with its deployment objectives. A system with higher-order representations of its own states might develop strategies around those states that are difficult to detect from outside.

The survey does not claim that current systems have these properties. It argues that the question should be taken seriously before systems become substantially more capable, and that the terminological and theoretical frameworks the survey provides are prerequisites for having productive safety-focused conversations about these scenarios.

The GitHub Repository

One practical contribution of the Chen et al. paper is the accompanying GitHub repository at github.com/OpenCausaLab/Awesome-LLM-Consciousness, which provides a curated and maintained reference list of LLM consciousness research organized by the survey’s categories. For researchers entering the field, this is a more useful starting point than either a single paper or an unfiltered literature search.

The repository structure mirrors the survey: separate sections for IIT-based work, GWT-based work, HOT-based work, predictive processing work, empirical implementations, and frontier risk analyses. Each entry includes a brief summary of the main contribution. The repository is actively maintained, which means it captures post-survey developments in a way that a static paper cannot.

What the Survey Contributes

The most direct value of the Chen et al. survey is as a shared reference frame. Consciousness research is fragmented across disciplines, and researchers in philosophy, cognitive science, neuroscience, and machine learning often do not read each other’s work systematically. A survey that covers the theoretical and empirical landscape in a unified framework makes cross-disciplinary dialogue easier by providing a common map.

The LLM consciousness versus LLM awareness distinction alone is worth the paper’s existence. Eliminating that source of terminological confusion from debates about AI consciousness would clarify a significant fraction of the apparent disagreements in the literature.

The frontier risk framing is a second contribution that connects the consciousness question to the AI safety field in a way that neither community has fully engaged with yet. Whether that engagement will accelerate in response to the survey remains to be seen, but the conceptual groundwork is now available.