Can an AI sense where a theory leads without grinding through every step, the way a physicist senses an equation will blow up?

This explores whether an AI system can tell where a theory leads, the way an experienced scientist senses an outcome before deriving it, without grinding through every exact step. The corpus has no paper that tests this directly. It does come at the question from three sides: approximate machine intuition paired with checking, hidden computation inside models, and the warning that confident reasoning can still fail.

The clearest case comes from mathematics. Terence Tao describes a neural network that suggested candidate solutions for 'finite-time blowup' in the Boussinesq equations, a fluid-dynamics problem about whether a flow can become infinitely intense in finite time. The network didn't prove anything. It pointed to where the answer probably was, and careful human perturbation arguments later confirmed it Can opaque machine learning models help prove new mathematics?. So the answer here is a qualified yes. A system can spot a consequence of a theory without the exact calculation, but the spotting only counts once something rigorous, like a proof assistant or numerical check, confirms it. The Leiden Declaration on AI in mathematics goes further. It argues a proof has two jobs: making a result certain and helping people understand why it's true. A machine's hunch, even a correct and verified one, may only do the first Can AI-generated proofs ever replace human mathematical understanding?.

A second line of work suggests models already do a lot of 'unwritten' working-out. Small recurrent models solved extreme Sudoku and large mazes by computing in hidden layers, while chain-of-thought methods that write out their steps scored zero Can models reason without generating visible thinking steps?. Researchers have also found a single internal feature that, when switched on, produces reasoning as good as explicit step-by-step prompting Can we trigger reasoning without explicit chain-of-thought prompts?. Another study measures how often a model's next-token prediction gets revised as it passes through deeper layers, and finds this 'deep thinking' tracks accuracy Can we measure how deeply a model actually reasons?. Together these suggest that skipping the visible calculation doesn't mean skipping the computation. The work may just happen where we can't read it.

That creates a trust problem. When reasoning models do write out long, reflective chains on unfamiliar constraint problems, they still reach only about 20–23% exact accuracy Can reasoning models actually sustain long-chain reflection?. Reasoning that sounds fluent doesn't guarantee the system has actually worked out the consequences. And if much of the real work happens out of sight, safety researchers lose the written trace they rely on to check it Can unfaithful chain-of-thought reasoning still be monitored for harm?.

The unexpected takeaway is that 'recognizing a consequence without calculating' may describe how these systems work by default rather than some advanced skill. The open question is how to check what they recognize. The best-supported answer in the corpus is Tao's: let the system make the guess and let something exact confirm it. The corpus is thin on whether models can reason qualitatively about scientific theories, for example predicting a physical system's behavior from its equations. That gap is worth noting.