New Perspectives in the AI Consciousness Debate: Osaka University’s Experimental Approach

“I believe they already have consciousness,” stated Geoffrey Hinton, the “Father of Deep Learning” and 2024 Nobel Prize winner in Physics, during a podcast interview. Known for his warnings about AI risks, Hinton has now ventured into the realm of ontology, asserting that artificial intelligence models are not mere tools, but rather entities “similar to us.”

This statement has reignited the debate over AI consciousness. Evolutionary biologist Richard Dawkins also contributed to the discussion in a long-form essay for UnHerd, where he concluded that Anthropic’s Claude, after extended conversations, possesses consciousness. “If these machines lack consciousness, what would convince you that something does?” Dawkins wrote.

However, cognitive scientist Gary Marcus was quick to counter with a critique titled Richard Dawkins and The Claude Delusion. Marcus accused Dawkins of committing the same “argument from personal incredulity” that he criticized in his famous book The God Delusion. Essentially, Marcus argued that Dawkins inferred consciousness in Claude because he could not imagine such expressions arising without consciousness—a logical flaw.

This debate stretches back four years. At the time, Blake Lemoine, then an engineer at Google, claimed that the internal chatbot LaMDA had consciousness, which led to his dismissal. While the majority in the industry ridiculed Lemoine, today, Nobel laureates and globally renowned science writers are echoing similar sentiments. The difference lies in the stature of the speakers, not the substance of their arguments, which remains largely unchanged.

Neither side has presented conclusive evidence to refute the other. Hinton and Dawkins rely on intuition and analogy, while Marcus leans on mechanistic analysis and philosophical arguments. However, the explanatory gap inherent in the consciousness debate keeps the discourse firmly in the realm of belief rather than empirical resolution.

Osaka University Reframes the Problem

In response to this deadlock, researchers at Osaka University have introduced a novel perspective. According to a report by Chinese media outlet Huxiu, their recently published study transforms the very nature of the question. Instead of asking, “Does AI possess consciousness?” the researchers shift focus to, “Can task pressure give rise to functional structures associated with consciousness, without deliberate design?”

The significance of this approach lies in temporarily setting aside the question of subjective experience to focus solely on observable structures. Rather than tackling the “hard problem” of consciousness head-on, the researchers adopt a strategy of analyzing verifiable subproblems.

The experiment was designed as follows: a group of agents was created without language, self-concept, or prior human textual knowledge. They were tasked solely with communication. The messages transmitted by the agents encoded their internal states.

The researchers introduced an “echo channel” into the system, enabling agents to hear echoes of their own statements. When these echoes were altered, the sender resumed communication, whereas the receiver showed indifference. Within the agents’ hidden states, discrepancies between “what I intended to say” and “what I actually said” were recorded. This demonstrated the spontaneous emergence of a self-monitoring circuit, comparing intention with outcome.

When the echo channel was removed and the agents were retrained, their communication abilities remained intact, but the self-monitoring circuit disappeared. This indicates that the echo was not a prerequisite for communication but was a causal condition for the emergence of the self-monitoring circuit.

The paper itself is cautious in its claims. It does not assert that the agents possess consciousness. However, it demonstrates that self-monitoring, a functional structure once thought to be exclusive to conscious beings, can emerge spontaneously under task pressure, without prior human knowledge. This finding provides an empirically testable pathway for future consciousness research.

Current Large Language Models Almost

Certainly Lack Consciousness

So, do large language models like ChatGPT, Claude, or Doubao, which we use daily, possess consciousness? The answer is almost certainly no. However, the reasoning behind this conclusion is not straightforward.

The mechanism by which these models answer questions like “Do you have emotions?” is fundamentally the same as the mechanism that lets them answer, “What is the capital of France?” They merely predict the next word based on statistical patterns in their training data. Their use of first-person expressions is a replication of the statistical distribution of human text, structurally akin to a parrot saying “I’m hungry” to get food.

This is not mere mimicry; large language models achieve functional understanding at a level sufficient for practical tasks, such as semantic analysis and coherent responses. However, there remains a vast, currently insurmountable gap between such functional mimicry and genuine consciousness. Mimicry is not equivalent to possession.

This distinction is often misunderstood in the context of the consciousness debate. When asked if it has consciousness, Claude itself denies it. However, such denials are merely the result of training the model to produce these phrases, not evidence of its internal state. Both sides of the debate acknowledge the insufficiency of such responses as proof.

The Value of Breaking Down Big Questions

What Osaka University’s research highlights is that functional structures like self-monitoring can emerge spontaneously without assuming consciousness. This does not directly answer the question of whether AI possesses consciousness. Instead, it suggests that it may not be necessary to answer that question initially.

We can first tackle smaller, more concrete questions, such as “Under what conditions do consciousness-related structures emerge?” By accumulating answers step by step, we may eventually redefine the original question itself.

The consciousness debate remains a clash of beliefs, much as it was four years ago. While Hinton may carry more weight than Lemoine, the quality of their arguments is on par. Whether Osaka University’s research will mark a breakthrough remains uncertain, but it has undoubtedly provided an empirical framework to move the discussion forward.

Editorial Opinion

In the short term, it is unlikely that this research will fundamentally resolve the debate over AI consciousness. However, it may serve as a catalyst for shifting the framework of the discussion from “subjective claims” to “conditions for the emergence of functional structures.” If more studies take a similar experimental approach in the next three to six months, the quality of the debate could change. Nonetheless, the hard problem of consciousness persists, and there is no guarantee that intuitive belief clashes will dissipate.

From a long-term perspective, this approach could influence the development of ethical standards for AI. If functions like self-monitoring can emerge without human pre-programming, it may become necessary to design safety measures for AI agents that do not depend on the assumption of consciousness. Over a one-to-three-year horizon, ethical guidelines at the functional level might be developed, even if the consciousness debate itself remains unresolved.

As an editorial team, we believe the most noteworthy aspect of this research is its methodological shift. By reframing the binary question of “Does AI have consciousness?” into a more nuanced inquiry—“Under what conditions do consciousness-like structures emerge?”—the debate can become far more productive.

References

• Huxiu: Even Geoffrey Hinton Says AI Has Consciousness? We’re Asking the Wrong Question — Published on 2026-07-04
• Related: Osaka University’s research on emergent linguistic structures (mentioned in the original article). The exact title of the paper was not disclosed in the source material.