topicactive
- 18sources
- 1article
- 10concepts
- 28people
Black box
The problem of a system known only from the outside — its inputs and outputs visible, the conversion between them closed to inspection — and the argument over whether the demand to open it should fall on machines, on minds, or on both. The method is old: W. Ross Ashby's 1956 cybernetics gave the black box its general statement and applied it in the same paragraphs to a sealed circuit and to a damaged brain. The modern machine form runs through the interpretability literature — the case for replacing opaque models (Rudin), the charge that "interpretability" is itself underspecified and that humans are no more transparent (Lipton), the post-hoc explanation methods (Ribeiro's LIME; Wachter's counterfactuals), and the mechanistic programme that claims to reverse-engineer a network feature by feature (Olah's circuits; Anthropic's Golden Gate Claude). The human form is the confabulation literature — Nisbett and Wilson on the absence of introspective access, Gazzaniga's left-hemisphere interpreter, Haidt's post-hoc moral reasoning. The two halves meet where a model's stated reasoning turns out not to be its real reason (Turpin), and where standard analysis methods fail on a chip whose wiring is fully known (Jonas and Kording). The stakes are drawn by Pasquale on the politics of algorithmic secrecy and by Dennett, whose intentional stance asks why a machine should owe a legibility that no mind has ever provided.
Algorithmic opacity · Black-box method · Confabulation · Counterfactual explanation