Does it Really Matter if AI Systems Exhibit Autonomy?

Braden Cooper

Historically, the development of artificial intelligence (AI) has been targeted, in part, at creating autonomous AI systems. Paradigmatic autonomous systems carry out their functions independently of human guidance. The requirements for capturing the autonomy criterion (i.e., independent function) utilized by AI theorists are at odds with the ways philosophers typically designate the term. There is a tension, then, with how autonomy is represented in philosophical literature, and thus, according to several philosophers, paradigmatic autonomous AI should not be considered autonomous systems. This tension has recently been extended to the philosophical literature on the development and deployment of Large Language Models (LLMs), which behave rather differently than traditional autonomous AI systems. I argue that the autonomy tension breaks down when considering behaviors of LLMs, including typical behaviors (e.g., learning, reasoning, planning, memory retention, problem solving, etc.), and atypical behaviors (as in the documented cases of alignment faking in certain model instances). I show that LLMs are already exhibiting the characteristics of autonomous systems accepted under multiple philosophical frameworks, and that they circumvent various problems leveled against the general possibility of developing autonomous artificial systems. Then, I turn the conversation over to the question of whether these foundational questions about AI systems really matter when considering various practical issues. I elucidate why the autonomy question initially arises but argue that it is not particularly fecund. Instead, I argue that philosophers should look more precisely at the comparative measurable dynamics of various systems when making practical decisions (e.g., decisions about policy, pursuit-worthiness of model development, etc.). LLMs are rapidly evolving and are already being implemented in social institutions. As such, I argue it is more promising to look at the apparent implications of these instances, rather than bother with foundational issues such as whether systems are really autonomous.