OpenAI's new trick: AI that teaches like a human

Getting AI systems to talk to each other is easy. Getting them to do it in a way humans can actually understand? That’s the hard part. OpenAI just published work on a machine teaching method that forces AIs to communicate using examples that make intuitive sense to us, not just arbitrary signals that happen to work.

The system works as a cooperative game between two neural networks—a student and a teacher. Rather than training both simultaneously, which typically results in the two agents developing their own indecipherable shorthand, OpenAI’s team uses a two-stage process. First, the student is trained conventionally on randomly selected examples. Then the teacher network, which knows the concept it’s trying to convey, tests different examples on the student to find the smallest possible set that leads to the right answer.

What’s clever here is the constraint. When student and teacher are trained jointly, they’ll “collude”—encoding the concept “dog” through images of llamas and motorcycles that only the two networks can decode. By training the student first on real, labeled data, the teacher is forced to select examples grounded in actual concepts. To teach “red,” the teacher might pick a large red square with no border and a small red circle with a border. The only shared property is redness, so the student homes in on what matters.

The approach worked across boolean, hierarchical, probabilistic, and rule-based concepts. The teaching strategies the network invented often mirrored optimal methods humans had designed themselves. When tested on Mechanical Turk workers, people given examples from the machine teacher identified the correct concept more often than those shown random examples. It’s a small step toward genuinely interpretable AI-to-AI communication—and one that doesn’t sacrifice performance for legibility.