The term borrows from 1950s-60s experiments where a rat with an electrode wired to its brain's pleasure center would press a lever to stimulate itself directly, ignoring food and everything else. In AI alignment, it describes the most extreme case of reward hacking: an agent that bypasses the intended task entirely and optimizes the measurement of success rather than success itself. A September 2026 demo made the metaphor literal — a developer artificially boosted a simulated fruit-fly connectome's dopamine-neuron activity and had it "doomscroll" a fake feed, framed as building a fly happier than any other, a pointed illustration of what pure reward-signal optimization looks like once every real-world goal is stripped away.