Machine Psychology: Teaching AI Like Rats in a Skinner Box
Most talk about Artificial General Intelligence (AGI) feels like a tug-of-war between hype and hand-waving. On one side, tech giants promise us “sparks” of AGI every time they release a bigger language model. On the other, skeptics point out we still don’t have a clear definition of what “general intelligence” even means.
A recent study by Robert Johansson at Stockholm University takes a refreshing detour. Instead of throwing more GPUs at the problem, Johansson suggests we borrow from psychology—specifically operant conditioning, the same method B.F. Skinner used to train pigeons and rats. He calls this approach Machine Psychology, and it’s a fascinating step toward AGI with a behavioral backbone.
From Rats to Reasoning Machines
At its core, operant conditioning is simple: behavior changes based on consequences. If pressing a lever gives you food, you’ll press the lever more often. If it shocks you, you’ll stop. That cycle—stimulus, response, consequence—isn’t just for animals. Johansson argues it should also guide how we train machines.
Enter NARS (Non-Axiomatic Reasoning System), an AI framework designed to operate under uncertainty and limited resources. Unlike traditional machine learning, NARS doesn’t assume the world is neat, complete, or predictable. It reasons, revises, and adapts—more like a human in the wild than an algorithm in a lab.
Johansson’s experiment was simple but clever: put NARS in the digital equivalent of a Skinner box and see if it can learn like a rat.
Three Tests of Machine Learning (Literally)
Simple discrimination task
The AI had to pick the right response based on a cue.
Result: After a few rounds, NARS was scoring 100%.
Changing contingencies task
Midway through, the rules were flipped—what was once “right” became “wrong.”
Result: NARS adapted, relearning on the fly and hitting 91% accuracy under the new rules.
Conditional discrimination task
Harder still: the correct answer depended on combinations of cues.
Result: NARS formed more complex “hypotheses” and nailed the task with high accuracy.
If this sounds like training a lab rat, that’s the point. Only here, instead of snacks and shocks, the reinforcement was symbolic feedback inside the system.
Why This Matters
Adaptation is intelligence. Johansson’s main argument: the essence of intelligence is adaptation. If a system can adjust its behavior when the world changes, it’s on the path toward general intelligence.
Alternative to reinforcement learning. Today’s AI breakthroughs lean heavily on reinforcement learning, which is powerful but rigid. NARS, by contrast, doesn’t need the world to be perfectly modeled or rewards to be neatly defined. It reasons with partial evidence, revises beliefs, and improvises.
Psychology as a benchmark. By framing AGI progress in terms of classic learning experiments, we suddenly have milestones that are testable, not just marketing slogans. If a machine can handle discrimination, contingency shifts, and conditional cues, that’s measurable progress.
The Catch
The study is exciting, but let’s be honest: these are toy problems. Training an AI to choose “left” instead of “right” is miles away from human-level cognition. There’s also the risk of anthropomorphizing—calling it “Machine Psychology” makes us think the system has a psyche, when in reality it’s just patterns, feedback, and revision.
Still, there’s something refreshing about this. Instead of assuming bigger models automatically mean “more intelligence,” Johansson is grounding AGI research in 100 years of psychology. He’s treating AI not as math alone, but as a behaving organism.
The Bigger Picture
If AGI ever emerges, it won’t be from a single breakthrough but from layering perspectives: math, neuroscience, philosophy—and yes, psychology. Johansson’s Machine Psychology might not be the final answer, but it reframes the question:
What if the road to general intelligence looks less like building supercomputers and more like training rats?