Black-box method
A way of studying a system entirely from the outside — by watching what its inputs do to its outputs — without ever opening it to inspect the mechanism within. The wager is that one can learn a great deal about a sealed thing this way, and that the same discipline applies whether the closed system is a circuit, a brain, or a modern machine-learning model.
The method of studying a system by what its inputs do to its outputs, without opening it to inspect the mechanism inside. Stated in general form by W. Ross Ashby's An Introduction to Cybernetics (1956), which traces the problem to wartime electrical engineering — an engineer handed a sealed unit with input and output terminals, asked to deduce its contents — and argues that the same discipline applies to any closed system, including the brain.
Etymology§
The term black box comes from electronics, where it named a circuit treated only by its terminal behaviour. Ashby took the engineering problem and made it a general epistemology; the phrase later acquired a second, secrecy-laden sense in Pasquale's The Black Box Society.
The black-box method is the neutral starting point of this topic. Ashby applies it in the same passages to a sealed circuit and to a patient with brain damage, treating the two as instances of one situation rather than as opposites — the symmetry the modern argument either accepts or contests.
Run forward, the method is what deep neural networks present by default: a mapping from input to output whose internal conversion is not legible from the outside. Run as a test, it is what Jonas and Kording turn on a microprocessor whose wiring is fully known, to ask whether the analyst's tools can recover a mechanism they cannot see. Mechanistic interpretability is the wager that the box need not stay closed; Dennett's intentional stance is the argument that for many purposes it need never be opened.