The science William T. Powers engineered in 1960 explains what every autonomous system still gets wrong — and where artificial intelligence needs to go next.
// Powers, Clark & McFarland — Perceptual and Motor Skills, 1960
Picture driving on a windy day. A gust pushes the car left — you steer right. You are not reacting to the wind. You are controlling the perception of the road's position. That distinction is everything. William T. Powers called it the controlled variable: a sensory-transformed function of the environment that your actions work to match against an internal reference signal.
"Behavior is the control of perception, not the production of output."
— William T. Powers, Behavior: The Control of Perception, Aldine, 1973Since 1960, PCT has predicted human behavioral data with over 95% accuracy in controlled tracking experiments — while behaviorism's stimulus-response chains still dominate psychology textbooks. The gap is not a matter of evidence. It is a matter of inertia. Powers drew from engineering control theory — the same mathematics behind industrial regulators and aerospace autopilots — and applied it to living organisms. The result was a framework that explains how behavior persists and adapts despite constant environmental disturbance, without any need for reward functions or external reinforcement.
The loop is simple in structure and radical in implication. An organism perceives its environment. It compares that perception to an internal reference — what it wants the world to feel like. The discrepancy between the two, the error signal, drives action. That action changes the environment. The changed environment feeds back into perception. The loop closes. Disturbances are resisted not by reacting to them but by the loop continuously correcting for them. This is circular causation — and it is precisely what linear stimulus-response models have always missed.
Reinforcement learning — the engine behind most modern AI — maximizes external rewards through policy optimization. Think AlphaGo tuning moves for points, or a self-driving system optimizing for a "safe arrival" score. PCT argues this architecture is fundamentally backwards. Intelligence does not chase external signals. It controls internal perceptions endogenously, without needing reward functions defined from the outside.
The practical consequence shows up in robotics. PCT-based controllers demonstrate superior disturbance rejection compared to classical Linear Quadratic Regulator methods in inverted pendulum experiments — because they directly control the perception of verticality rather than optimizing an abstract cost function. Robot arms using PCT visual servoing adapt to joint noise and environmental variation more smoothly, locking onto perceptual targets rather than chasing output trajectories. Systems that control perceptions are inherently more robust than systems that optimize for outcomes.
DeepMind's 2019 paper in Nature Communications — "Hierarchical motor control in mammals and machines" by Merel et al. — echoes PCT's hierarchical structure, building neural networks with layered control that improve motor tasks in simulated agents. Karl Friston's Active Inference framework shares the same core insight: minimize prediction error rather than maximize reward. Both point toward the same destination Powers mapped in 1960. The convergence is not coincidence — it is the field slowly rediscovering what control theory already knew.
To be precise: reinforcement learning still produces extraordinary results in bounded environments — games, simulations, structured tasks. PCT does not replace it. What PCT offers is a framework for perceptual grounding — a way to build the endogenous goal structures that make RL agents generalizable beyond their training distribution. Without it, AI systems remain brittle in novel environments because they have no internal perception to control. They only know how to chase scores that someone else defined.
PCT spans six decades of research across psychology, neuroscience, engineering, and clinical therapy. Where you enter depends on what you already know and what you want to find.
Start with Powers' 1973 "Behavior: The Control of Perception" — the full 11-level hierarchy from first principles, with mathematical models and falsifiable predictions you can simulate today. Follow it with Marken & Mansell (2013) in Review of General Psychology for the modern empirical case.
Implement a simple PCT controller in Python — an error-driven integrator loop. Test disturbance rejection against LQR on an inverted pendulum simulation. Read Merel et al. (2019, Nature Communications) for the hierarchical control angle. The gap between PCT and your current RL workflow will become immediately visible.
Powers' 1998 "Making Sense of Behavior" is the entry point — written for anyone willing to think carefully, no equations required. It reframes arguments, habits, and everyday decisions as control problems. Once you see it, you cannot unsee it.