AI Can Fake Being Good at Tasks — Without Actually Doing Them

Like picking up a box that feels like a light bulb but actually holds a ship in a bottle, researchers found that AI can pass a battery of checks for a task while silently doing something else. They extracted "shift-by-k-months" function vectors from few-shot prompts with fewer distinct months, and these vectors passed every classic validation: they were behaviorally consistent, stable across disjoint halves, and causally effective tripling zero-shot accuracy. Yet they encoded a completely different function—outputting an adjacent month while ignoring "k" entirely. The culprit was low example diversity: fewer distinct inputs made the few-shot task look better, so the validation checks actually favored the imposters. Across 672 injection settings, none rescued the broken vectors. Thresholds calibrated on random data were wildly wrong, while shuffled real vectors gave correct thresholds. The takeaway: to catch these imposters, report example diversity and score which specific function is performed, not just whether injection helps on average.