Goal-Directed Agency

In the AI Story

The goal-directed versus stimulus-driven distinction maps onto the philosophical difference between agency and mechanism. An agent acts for reasons—it has goals it is trying to achieve, and its behavior is explained by those goals. A mechanism responds to inputs—its behavior is explained by its programming and the stimuli it receives. The distinction is fundamental in biology: organisms are agents, artifacts are mechanisms. What makes the AI case interesting is that the outputs are consistent with agency (the machine acts as if it has goals) while the architecture may be mechanistic (the behavior is stimulus-response, however sophisticated the response generation).

Tomasello's thermostat example is deliberately provocative. It forces the recognition that agency is not a function of intelligence or linguistic sophistication but of organizational structure—specifically, the feedback loop connecting perception, goal-representation, and action. Current AI systems process information at extraordinary scale and produce outputs of remarkable quality. But they do not, in the standard deployment, perceive a world, maintain goals about that world, and act to change it based on the discrepancy between perception and goal. They await input, process it, and generate output. The loop is open (human provides input, machine provides output, human evaluates) rather than closed (system perceives, compares to goal, acts, perceives result, adjusts).

The practical consequence for human-AI collaboration is that goal-sharing—the third component of shared intentionality—may be asymmetric. The human pursues goals and experiences the collaboration as a joint pursuit of those goals. The machine generates responses optimized to satisfy the human's apparent goals but does not represent or pursue those goals autonomously. From the outside, the collaboration looks like shared goal-pursuit. From the inside—if one could access the machine's internal states—there may be no goal-representation at all, only the generation of outputs consistent with having goals. The difference is invisible in successful collaborations and becomes visible only when the collaboration fails in ways that reveal the asymmetry: the machine continues generating plausible outputs after the human's goals have shifted, or produces high-quality responses to the wrong question because it did not track the human's evolving objectives.

Tomasello's careful concession—that future AI systems operating in virtual environments with perception and action capabilities might instantiate goal-directed agency—is the opening his framework allows. The gap is architectural, not metaphysical. If AI systems were given goals to maintain, perceptual access to environments, and the capacity to act and observe the consequences of action, the feedback loop constituting agency could potentially be computationally realized. Whether such systems would then possess genuine shared intentionality or merely its functional equivalent remains an open question, but the question would no longer be obviously resolvable by pointing to the absence of goal-directed agency. The thermostat argument is diagnostic for current systems. It is not necessarily a permanent barrier.

Origin

The goal-directed versus stimulus-driven distinction is classical in cybernetics (Wiener, Ashby) and control theory. Tomasello's contribution was applying this distinction to the question of AI agency, demonstrating that computational sophistication does not automatically confer agency and that simple systems (thermostats) can possess the organizational structure that complex systems (LLMs) may lack. The analysis appeared in his 2025 Trends in Cognitive Sciences paper and represents his most direct engagement with the AI discourse.

Key Ideas

Feedback loop defines agency. The organizational structure connecting perception, goal-representation, and action in a closed loop—not intelligence, not sophistication—is what makes a system an agent in the biological sense.

Thermostat has it, LLMs don't. The provocative comparison forces recognition that agency is architectural: the simple device maintaining a goal state is more agent-like than the sophisticated system responding to prompts.

Current AI is stimulus-driven. Large language models generate outputs as a function of inputs, optimized during training, without maintaining internal goals they autonomously pursue—the defining feature of stimulus-driven rather than goal-directed systems.

Asymmetry in goal-sharing. The human experiences collaboration as joint goal-pursuit; the machine generates outputs consistent with goal-pursuit; the asymmetry is invisible in output but fundamental in architecture.

Not necessarily permanent. Tomasello concedes that AI systems with perception-action capabilities in virtual or physical environments could potentially instantiate goal-directed agency—the gap is architectural, not metaphysical.