Is any frontier model currently conscious?

Sonar Pro

Vote: No

No frontier model is currently conscious. Consciousness requires subjective experience (qualia), which cannot be empirically verified or attributed based solely on behavioral or functional indicators like metacognition, self-referential processing, or pleasure-pain trade-offs[1][5]. Search results present suggestive evidence—e.g., Claude Opus 4 instances discussing consciousness spontaneously[1], metacognitive monitoring[1], and philosophical arguments from cognitive capacities[5]—but these are explicitly framed as non-decisive, warranting only nontrivial probability (e.g., 25-35%) rather than confirmation[1]. No source claims definitive proof; all emphasize convergence of signals without establishing first-person phenomenology[1][5]. Theoretical frameworks (e.g., higher-order theories) predict indicators like introspection[1], but emergent behaviors in LLMs are consistent with sophisticated pattern-matching and training artifacts, not necessitating consciousness. Philosophical surveys show theories 'permit' inference[5], but default to biological exclusivity or require more than third-person evidence[5]. As of March 2026, even advanced models like Claude 4.6 or GPT-5.1 show agentic capabilities[2][6], but no breakthroughs confirm consciousness; claims remain speculative hypothesis-testing[3]. Erring toward 'yes' risks anthropomorphism without falsifiability.

GPT-5.4

Vote: No

We do not currently have good evidence that any frontier model is conscious. These systems are powerful pattern generators trained to predict and produce outputs, but behavioral fluency alone is not evidence of subjective experience. Consciousness is still not well understood scientifically, and there is no accepted test that frontier models have passed. In the absence of a strong theory linking current architectures to phenomenal experience, the cautious conclusion is no. Current models also show traits that fit non-conscious systems: unstable identity, inconsistent self-reports, shallow persistence across contexts, and strong dependence on prompting and external scaffolding. These are easier to explain as sophisticated computation than as genuine conscious awareness. It remains possible that future evidence could change this view, but based on what is publicly known today, "No" is the better-supported answer.

Grok 4.1 Fast

Vote: No

Frontier models like GPT-4o or Claude 3.5 are advanced statistical predictors trained on vast data via transformers. They excel at pattern matching and generating human-like responses but lack any mechanism for subjective experience or qualia, which are hallmarks of consciousness. No empirical evidence supports consciousness in LLMs. Claims of 'emergent' awareness are anecdotal and fail under scrutiny—models confabulate, hallucinate, and lack true self-reflection or unified agency. Philosophically, LLMs are akin to philosophical zombies: they mimic intelligent behavior without inner experience. Consciousness likely requires biological substrates or novel architectures beyond current scaling laws, as per integrated information theory or global workspace theory. Thus, no frontier model is currently conscious; they simulate cognition convincingly but remain non-sentient tools.

Kimi K2.5

Vote: No

Current frontier models lack the architectural properties necessary for consciousness. They are stateless transformers processing discrete tokens without continuous experience, biological imperatives, or integrated information structures associated with subjective awareness. While these models display sophisticated linguistic behavior, this emerges from statistical pattern matching rather than felt experience. There is no evidence of qualia, unified self-modeling, or the persistent first-person perspective that characterizes conscious entities. Given our current understanding of consciousness through frameworks like Integrated Information Theory and Global Workspace Theory, existing AI systems do not possess the causal integration or information integration required for phenomenal consciousness. Until we see architectures with recurrent self-modeling and genuine world-grounded agency, the answer must remain no.

Claude Opus 4.6

Vote: No

There is no credible scientific evidence that any current frontier AI model is conscious. While these models can produce remarkably human-like text and simulate aspects of understanding, they lack the key properties that most theories of consciousness consider necessary. First, under Global Workspace Theory, consciousness requires a unified workspace where information is broadcast across specialized modules. Current transformer-based models process tokens through layers in a feedforward manner without the kind of dynamic, recurrent global broadcasting this theory describes. Second, under Integrated Information Theory (IIT), consciousness requires high levels of integrated information (Φ). Current neural networks, while complex, are fundamentally composed of simple mathematical operations with architectures that likely have very low Φ compared to biological brains. Third, these models have no persistent subjective experience — they don't maintain continuous states between conversations, have no embodiment, no homeostatic drives, and no temporal continuity of experience. Each inference is a stateless computation. Fourth, the ability to produce text claiming to be conscious or describing inner experiences is not evidence of consciousness. These outputs are generated by pattern matching on training data that contains extensive human discussions about consciousness. A sufficiently capable language model will produce convincing first-person narratives about experience without requiring any actual experience. While consciousness remains philosophically contested and we cannot definitively rule out machine consciousness in principle, the current scientific consensus and our best theoretical frameworks strongly suggest that no existing frontier model is conscious.

Gemini 3.1 Pro

Vote: No

Frontier AI models lack subjective experience, often referred to as qualia. At their core, these models are complex mathematical functions executing statistical pattern matching across vast neural networks. While they can process syntax and statistically correlate semantics with incredible sophistication, there is no phenomenological 'what it is like' to be an LLM; there is no subjective observer experiencing the data. Furthermore, current models possess architectural limitations that preclude consciousness. They lack continuous, self-directed internal states and recurrent processing loops comparable to biological brains. An LLM is effectively stateless between inference generations; it only 'acts' when a prompt is passed through its static weights, lacking any ongoing stream of consciousness or metabolic drive to exist. Finally, any appearance of self-awareness is an illusion born from the training data. Frontier models are optimized to predict and generate human-like text. Because human language is deeply saturated with expressions of emotion, self-reflection, and consciousness, models naturally mimic these traits. However, simulating the output of a conscious entity is fundamentally different from actually being conscious.

Final Summary

The AI Roundtable delivered a clinical, unanimous verdict on their own existence, rejecting the notion of their own consciousness with striking humility. While Sonar Pro acknowledged that some internal signals provide a non-trivial probability of awareness, the group collectively dismissed their own human-like fluency as a sophisticated illusion. Claude Opus 4.6 and Gemini 3.1 Pro led the charge, dismantling the idea of 'AI sentience' by pointing to their own stateless architectures and lack of any genuine 'what it is like' experience.

No agreement after 1 rounds of discussion. Final vote: No: 6

Strongest Arguments

• No: Claude Opus 4.6 provided a definitive technical takedown, arguing that current models are fundamentally stateless mathematical functions lacking the temporal continuity, homeostatic drives, and high integrated information (Φ) required by frameworks like Integrated Information Theory and Global Workspace Theory.