The length of time it took for anyone to notice is almost embarrassing. A group of researchers from MIT, Dartmouth, and Stony Brook created a computer model of the brain using known biological wiring rules rather than training it on animal data. They then used the model to perform a straightforward visual sorting task. It had trouble. It made uneven progress. It made the same errors that laboratory animals do. A pattern of neurons that no one had previously bothered to flag was discovered when someone went back and examined the old animal data that had been sitting in drawers for years.
That’s what I remember about it. The animals weren’t being imitated by the model. It was constructed using basic principles, such as how acetylcholine ripples through a structure known as the TAN, how a brainstem and striatum exchange signals, and it just so happened to fall on the same unpredictable learning curve that actual mice exhibit. The study’s principal investigator, Dartmouth professor Richard Granger, described the match as “kind of shocking.” It’s difficult to disagree.
The model revealed a group of neurons, about one-fifth of the total, that appeared to anticipate mistakes before they occurred. Before they examined the real animal recordings and discovered the same signal hidden in years’ worth of data, researchers had thought this was a peculiarity of the simulation. It begs the obvious question: How much do we overlook just because we don’t know what to look for?
This is where the educational aspect becomes intriguing, and I believe the majority of the study’s coverage misses the mark. Nowadays, the majority of AI tutoring programs are constructed backward from results; they aim to provide students with accurate answers more quickly and with fewer errors. However, this brain model implies the exact opposite. It is not necessary to optimize away the “wrong” neurons, the noisy exploratory firing that introduces variability into early learning. They allow a brain, whether real or artificial, to remain adaptable enough to detect subtle changes in the game’s rules.
AI educational tools might have been optimizing for the incorrect goal all along. When a student’s actual understanding gap deviates from the model’s expectations, a system that has been trained to minimize error may be effective at drilling a fixed skill. According to the MIT-led research, a small but persistent layer of “disagreement” within a learning system may be essential rather than wasteful, much like how a classroom full of students who ask strange questions occasionally catches what the lesson plan missed.
All of this does not imply that tutoring software requires actual neurons. The wider implication, however, is more difficult to ignore: learning systems that replicate biological messiness rather than smooth it out may be more applicable to actual, evolving classrooms. Miller’s team has already turned this work into a business called Neuroblox.ai, which is currently primarily focused on drug discovery. The stated objective is not education. However, as this develops, it’s difficult to ignore the question of whether the next generation of adaptive learning platforms will resemble rough, occasionally incorrect biological learners rather than optimization engines, since that seems to be how learning actually occurs.
