08-20-2026, 01:49 PM
Exploring Consciousness in LLMs: A Systematic Survey of Theories, Implementations, and Frontier Risks
Authors: Sirui Chen, Shuqin Ma, Shu Yu, Hanwang Zhang, Shengjie Zhao, Chaochao Lu
Published: May 26, 2025
Field: Artificial Intelligence / LLMs / Cognitive Science
arXiv: 2505.19806
This survey examines the difficult question of whether large language models could possess anything meaningfully described as consciousness. A central contribution is distinguishing consciousness from awareness: awareness means responding appropriately to information in the environment, whereas the authors associate stronger notions of LLM consciousness with capabilities such as monitoring one's own uncertainty, detecting inconsistencies, reflecting on reasoning and correcting oneself. They also distinguish phenomenal consciousness—subjective experience or “what it feels like”—from access consciousness, where information is globally available for reasoning, reporting and decision-making. Importantly, the paper does not claim that today's LLMs are conscious.
The authors connect LLM research with several major theories of human consciousness. Recurrent Processing Theory suggests recurrent feedback may be important; iterative LLM self-refinement superficially resembles such processes. Integrated Information Theory (IIT) associates consciousness with integrated information $\Phi$, although proponents of IIT generally argue that present AI architectures lack the appropriate causal structure. Embodiment theories suggest that lacking a body and continuous interaction with the physical world may be a fundamental obstacle. By contrast, Global Workspace Theory (GWT) may be easier to approximate computationally through architectures in which information is distributed among specialized modules. The paper then surveys measurable consciousness-related abilities including Theory of Mind, situational awareness, metacognition, sequential planning, and creativity. Existing models display portions of these abilities, but behavioral success alone cannot establish genuine subjective experience.
A major part of the survey concerns safety. More capable, self-monitoring and autonomous systems could potentially increase risks involving scheming, manipulation, autonomous action and collusion between agents. The authors therefore argue that future research needs better consciousness-specific benchmarks, mechanistic interpretability capable of examining internal representations rather than merely outputs, embodied or multimodal systems, and studies of emergent behavior in multi-agent environments. Their central message is that research should move beyond the simplistic question “Is ChatGPT conscious?” toward identifying precisely which computational properties associated with consciousness can be defined, measured and tested.
Key takeaways
ARTICLE
Authors: Sirui Chen, Shuqin Ma, Shu Yu, Hanwang Zhang, Shengjie Zhao, Chaochao Lu
Published: May 26, 2025
Field: Artificial Intelligence / LLMs / Cognitive Science
arXiv: 2505.19806
This survey examines the difficult question of whether large language models could possess anything meaningfully described as consciousness. A central contribution is distinguishing consciousness from awareness: awareness means responding appropriately to information in the environment, whereas the authors associate stronger notions of LLM consciousness with capabilities such as monitoring one's own uncertainty, detecting inconsistencies, reflecting on reasoning and correcting oneself. They also distinguish phenomenal consciousness—subjective experience or “what it feels like”—from access consciousness, where information is globally available for reasoning, reporting and decision-making. Importantly, the paper does not claim that today's LLMs are conscious.
The authors connect LLM research with several major theories of human consciousness. Recurrent Processing Theory suggests recurrent feedback may be important; iterative LLM self-refinement superficially resembles such processes. Integrated Information Theory (IIT) associates consciousness with integrated information $\Phi$, although proponents of IIT generally argue that present AI architectures lack the appropriate causal structure. Embodiment theories suggest that lacking a body and continuous interaction with the physical world may be a fundamental obstacle. By contrast, Global Workspace Theory (GWT) may be easier to approximate computationally through architectures in which information is distributed among specialized modules. The paper then surveys measurable consciousness-related abilities including Theory of Mind, situational awareness, metacognition, sequential planning, and creativity. Existing models display portions of these abilities, but behavioral success alone cannot establish genuine subjective experience.
A major part of the survey concerns safety. More capable, self-monitoring and autonomous systems could potentially increase risks involving scheming, manipulation, autonomous action and collusion between agents. The authors therefore argue that future research needs better consciousness-specific benchmarks, mechanistic interpretability capable of examining internal representations rather than merely outputs, embodied or multimodal systems, and studies of emergent behavior in multi-agent environments. Their central message is that research should move beyond the simplistic question “Is ChatGPT conscious?” toward identifying precisely which computational properties associated with consciousness can be defined, measured and tested.
Key takeaways
- There is currently no convincing evidence that LLMs possess subjective consciousness. Behaviors such as self-reflection or Theory of Mind are not proof of inner experience.
- Consciousness is not a single capability. The paper separates awareness, metacognition, self-modeling, planning and phenomenal experience instead of treating them as equivalent.
- Some functional components associated with consciousness are increasingly measurable in LLMs, particularly metacognition, Theory of Mind and situational awareness.
- The critical future problem is developing objective evaluation and interpretability methods that can distinguish genuine internal mechanisms from models merely producing convincing conscious-like language.
ARTICLE
┌────────────────────────────────┐
│ KONSTANTINOS MICHAILIDIS │
└────────────────────────────────┘
│ KONSTANTINOS MICHAILIDIS │
└────────────────────────────────┘