Minsky named and formalized "the credit-assignment problem" in 1961 — as a temporal problem about sequential decisions, not network weights

Minsky's "Steps Toward Artificial Intelligence" (1961) contains a dedicated section — III-D, "The Credit-Assignment Problem for Learning Systems" — that coins the problem by name: "If the run is successful, how can we assign credit for the success among the multitude of decisions?" His illustration is sequential and game-shaped: "Suppose that one million decisions are involved in a complex task (such as winning a chess game). Could we assign to each decision element one-millionth of the credit for the completed task?"

The framing matters for how the term is used now. Minsky's problem is temporal — many decisions in sequence, one eventual outcome — the shape reinforcement learning inherited. What backpropagation solves is what the field later called structural credit assignment: distributing a scalar error across the weights of a layered network in one shot (claim-backpropagation-special-case-of-reverse-mode-ad). The 1986 Nature paper solves the structural problem without using Minsky's term (claim-rhw-1986-demonstration-not-invention). Retellings that say "backpropagation solved Minsky's credit-assignment problem" compress two related but distinct problems into one — the same flattening pattern the myth ledger tracks elsewhere. See moc-backpropagation-origins.