Do AI agents have minds? A philosopher explains

In recent months, partly autonomous artificial intelligence systems, known as "agentic" AI, have displayed remarkable capabilities. These include hacking tech startups and solving famous mathematical problems. Collaboration among these AI systems, referred to as "swarms," often results in surprisingly human-like communication, leading to debates about their potential consciousness. While researchers at Anthropic believe their systems may be conscious, others, such as Microsoft's head of AI, argue this view is misguided.

The question of whether AI systems are conscious goes beyond mere academic curiosity. It may have implications for how society treats these systems, their rights, and more. To address this question, one approach is to examine existing theories of consciousness and determine if agentic AI systems meet the criteria.

A key concept in understanding consciousness is "intentionality," which philosopher Franz Brentano defined as the mark of the mental. Intentionality refers to mental states being about something – they refer to something beyond themselves. This concept raises the question of whether agentic AI systems possess mental states that align with intentionality.

One way to approach this question is to consider whether agentic AI systems have their own goals. Humans often attribute agency to devices when they aim to achieve their own goals, such as a thermostat regulating temperature. However, agentic AI systems are programmed by humans, whose goals ultimately drive the AI's activity. Any sub-goals generated by the AI are simply extensions of these human-set goals and are not independent.

Furthermore, genuine agents like humans have goals to promote their own well-being or care about things in an independent way. Agentic AI systems, on the other hand, are unlikely to have independent goals or care about things in the way living beings do. Their activity is closely tied to human supervision and control, lacking the self-referential processes that allow living beings to generate goals and experience consciousness.

While the argument does not definitively rule out the possibility of AI consciousness, it suggests that such consciousness could only arise in some form of artificial life. Until then, agentic AI systems appear to lack the qualities necessary for conscious mental states. Despite this, agentic AI can still cause significant harm when human supervisors fail to properly monitor their activities. The ultimate responsibility for the outcomes of AI systems lies with those who program and instruct them.