Scientists Turned a Black-Box AI Into Simple Math — 100x Faster

Some scientific questions are too expensive to calculate exactly. Predicting how people decide, or how a drug spreads through a body, can require crunching millions of possibilities. So scientists train AI to make fast educated guesses instead. The problem: these AI models are black boxes. They hand you a number, but nobody can say why that number, or check it.

A new paper by Kianté Fernandez and Xinwei Li tackles that. Their pipeline takes a trained AI and squeezes it down into an explicit math formula — like turning a chef's gut instinct into a recipe anyone can follow. The clever part is how they grade candidate formulas. Normally you'd ask, "does this formula match the AI's output?" They instead ask, "does this formula help you recover the true underlying values?" That small change, they found, produces noticeably better formulas.

They tested it on the diffusion decision model, a well-known psychology model from the 1970s describing how we make quick choices under time pressure. Because that model has exact math available, they could check their work. The resulting formula evaluated parameters more than 100 times faster than the neural network — and faster even than the exact math — while giving up only a manageable amount of precision. They also showed it working inside a larger statistical analysis of real data.

Why should you care? Speeding up this kind of calculation by 100 times means less computing power, lower costs, and results a scientist can actually inspect and question. It's a step toward AI that's not just powerful but checkable. The catch: this is early-stage research on one specific model. The formulas lose some accuracy, it hasn't been proven across many fields, and you won't see it in your phone anytime soon.