New Math Shows Why AI Makes Mistakes It Can't See Coming

Two researchers, Mahdi Mohammadigohari and Nicole Mücke, published a new mathematical paper on how AI models actually learn. Their target is a quiet habit built into nearly every modern AI system: it picks out useful patterns from its training data, then uses those exact same examples to build its final predictions. Imagine a student who writes the practice questions, studies them, and then grades their own exam. They'll look brilliant — but you've learned nothing about how they'd handle a test they've never seen. That gap is the subject of this paper.

Most existing math for AI pretends the chosen patterns were handed over in advance. That assumption makes the analysis cleaner, but it erases the cost of choosing them. The authors fix this by deriving exact formulas that account for the selection step. Their key measurement, which they call "activation mass," is essentially how strongly the AI's internal switches fire on average. They show that the true difficulty of learning depends on both that firing strength and a separate penalty for the freedom to choose, which they call the selection gap.

They also introduce something they dub "Brownian heads" — a class of prediction layers borrowed from a branch of probability theory that describes random, jittery motion. For a general reader, the useful takeaway isn't the machinery. It's the finding: two models can have identical firing strength yet genuinely different difficulty, purely because of how much freedom they had to choose. Real experiments in the paper confirm the math and demonstrate that gap.

The honest caveat: this is pure theory, with no product, app, or immediate change to anything you use. It's the kind of work that shows up years later inside better tools for deciding how much data a model needs and how much to trust its results. It won't affect your day tomorrow — but it makes 'the AI said so' a slightly less mysterious claim.