New Ruler Measures How Much AI Thinks Without Showing Its Work

Right now, when an AI solves a problem, it usually shows its work — writing out steps in plain text before giving an answer. That running commentary, called chain-of-thought, is how researchers, auditors, and safety teams spot mistakes, bias, or outright deception. It's the closest thing we have to a window into an AI's mind.

But newer AI designs are starting to reason silently instead — doing the thinking inside hidden mathematical states rather than in words anyone can read. If that happens, the window closes. You'd get an answer with a confident explanation attached, and no way to check whether the explanation matches what the model actually did.

This paper proposes a ruler for measuring that. It's called 'opaque serial depth' — roughly, how many steps of hidden thinking a model can chain together before it has to write something down. Think of a student doing long division entirely in their head versus on paper. For a standard AI model, researchers estimate that number is roughly equal to the model's number of layers — potentially dozens of invisible steps. The paper's main contribution is defining what counts as a genuine 'readable checkpoint': a spot where the model outputs actual text, code, or similar — and hasn't been rewired to disguise hidden thinking as readable words. Normal training techniques like prompting and reinforcement learning still qualify.

The catch: this is a measuring proposal, not a fix. Nothing here stops a company from building a model that thinks in secret. It just gives regulators and researchers a shared number to compare, and a way to ask, publicly, how opaque any given AI really is.