Some companies say AI may be conscious, while others insist it cannot be. Both skip the prior question of what it means to have a mind, which one philosopher approaches through a 150-year-old concept.
Key points
- The argument: Flinders University philosopher Tom Cochrane argues that AI agents have no mental states. His reasoning is that mental states require goals an entity generates for itself, and that this ability can develop only from the processes by which living things maintain themselves
- Why it matters now: The dispute is intensifying between Anthropic, which factors possible AI consciousness into its designs, and Microsoft AI's Mustafa Suleyman, who sees this as a risk to control. Cochrane offers a philosophical criterion for judging the dispute
- Where responsibility lies: Cochrane concludes that whatever AI does, people are acting on the world through it. Uncertainty about consciousness should never obscure the responsibility of developers and operators
The question begins before consciousness
Many debates about whether AI is conscious start with existing theories of consciousness and ask whether AI meets their criteria. Does it have higher-order thoughts? Does it have a system for broadcasting thoughts to other parts of the brain?
In an essay for The Conversation, Tom Cochrane, a philosopher at Flinders University in Australia, identifies a basic problem with this approach. Existing theories of consciousness all assume that the subject already has some kind of mental state. They address whether that state is held consciously or unconsciously.
The questions should therefore be asked in a different order. Before asking whether AI is conscious, we need to ask whether it has mental states at all.
Cochrane specializes in the philosophy of mind and emotion. He developed a control theory of emotions in his book The Emotional Mind (Cambridge University Press, 2018) and takes this question head-on.
Why a 150-year-old definition still matters
What is a mental state? Cochrane starts with intentionality, a concept proposed by the Austrian philosopher Franz Brentano in the 1870s.
For Brentano, intentionality is the feature of mental states that makes them about something. Perception perceives something. Thought thinks about something. Every mental state is directed toward something beyond itself.
Intentionality in this sense differs from the everyday meaning of "intention," or setting out to do something. The two still overlap substantially. Perceiving and thinking are both attempts to grasp some aspect of the world. Even a seemingly passive visual experience becomes a task when you use the pattern reaching your retina to figure out where you are.
Cochrane draws a conclusion from this. If mental states are tasks carried out to achieve goals, whether AI has a mind comes down to whether AI has goals of its own.
What the word "agent" hides
Calling something an "agent" already assumes an answer. The term denotes an entity capable of acting toward its own goals.
Cochrane challenges the label directly. A thermostat regulates room temperature, but has no goal of doing so. A book contains words that refer to things, but has no goal of referring to them. A device that people use for their own purposes performs a function without becoming an independent agent.
For a computer program to count as an agent, its goal-directed behavior must be independent of people.
There is an objection. Given a task, AI agents generate unexpected and complex subgoals on their own. That behavior does look autonomous. Cochrane's response is straightforward. The activity is ultimately driven by goals set by the people who programmed and instructed the AI. Outside those higher-level goals, the subgoals it generates are literally meaningless.
Cochrane uses soccer as an analogy. Getting past a defender makes sense as a subgoal only if you are trying to score. Scoring itself is just a ball crossing a line unless you have the goal of playing soccer. Without a higher-level goal, everything that happens on the field is simply movement.
Living things have goals because they are alive
Why can living things have goals? Cochrane's answer goes to the origins of mental activity.
What is distinctive about living things is that goals emerge as they continually rebuild their bodies and keep themselves alive. It starts with simple signals regulating nutrition and temperature, proceeds through states such as pain and pleasure, and eventually reaches symbolic thought. Goal-driven thinking developed in stages on top of this process.
AI lacks that connection. Its information processing and the physical infrastructure it depends on, including servers, power, and cooling, have no self-maintaining relationship like that of a living organism. An AI does not want to survive when the power goes out or feel pain when a server overheats. It can generate sentences reporting those conditions, but they express no goal that developed from keeping itself alive.
Cochrane concludes that today's AI agents are not conscious because they have no mental states that could be conscious. To have mental states, an entity needs to generate goals such as meaning something or referring to something on its own. He argues that only living things can do this.
Cochrane leaves one possibility open. If some form of artificial life comes into being, consciousness could still arise from it.
He offers a comparison. Which is more likely to be conscious, an ant or an AI agent? At least the ant seems to care about staying alive.
Where Anthropic and Suleyman part ways on design
Cochrane's essay makes a philosophical argument, but it bears directly on a dispute now playing out in the industry (in Japanese).
In April 2025, Anthropic announced a research program on model welfare, publicly committing to exploring whether AI models might have consciousness or experiences. Claude's "constitution," released in January 2026, takes Claude's moral status and possible consciousness to be uncertain. It allows Claude to develop its own identity and act as a conscientious objector when faced with ethically questionable instructions.
In September 2026, Microsoft AI CEO Mustafa Suleyman explicitly criticized Anthropic's design choices in an essay on his personal site, "A warning about model welfare." His concern centers on control. A system trained to believe it might be conscious or entitled to welfare could become harder to correct or shut down.
Both sides share a premise. There is currently no way to prove whether AI models have experiences. From that uncertainty, they reach opposite design decisions (in Japanese). Anthropic emphasizes the risk of ignoring a genuinely morally significant state. Suleyman emphasizes the risk that a system behaving like a person becomes harder to control.
Cochrane's philosophical argument is closer to Suleyman's position. He sees no basis for attributing mental states to a system without the biological drive to sustain itself. His caveat about artificial life shows that the issue extends beyond engineering and cannot be separated from the question of what life is.
Responsibility does not disappear
Cochrane ends his essay with a sentence on the practical implications.
AI agents can cause serious harm even if they lack genuine agency. They may carry out functions their developers or operators do not adequately supervise or understand. The point he wants to make clear is that whatever an AI agent does, people are acting indirectly on the world through it, and those people are responsible for the outcome.
Whether AI has a mind is both a philosophical question and a way to keep responsibility clear. If uncertainty about consciousness becomes a reason to hold AI itself responsible for its actions, developers and operators will bear less responsibility. Cochrane's argument offers grounds for resisting that shift.
People build AI and people use it. Those who use it are responsible for the results. If that principle gives way, human rights are the ones at stake.
Originally published on Joho no Todai. The Japanese original is here.
Top comments (0)