A driver is traveling at night in fog. The headlights pick out only the nearest meter of road, and a map marking every turn is useless because there are no landmarks to see. That is roughly how Maxwell's demon feels in a noisy world: it tries to extract work from thermal motion, but its knowledge of particle trajectories drowns in interference. The authors of a new study have shown that under such conditions, detailed path statistics not only fail to help but actually become worse than a simple glance at the current moment.

The experiment is built on a microcantilever, a tiny plate that oscillates at 1100 Hz. Its position is measured to a precision of 1 pm, but noise is deliberately added, as if part of the windshield were smeared with mud. A control force switches depending on the sign of the measured coordinate, and a little work can be squeezed out of thermal fluctuations. The more precisely the demon knows where the plate is, the more work it extracts.

When measurements are nearly ideal, the tightest bound turns out to be unavailable information. But add noise, and this bound quickly weakens: the detailed history of the path becomes unreliable. Markovian mutual information, which has no memory of the past at all, holds up better. Over a wide range of noise, it is precisely this bound that predicts how much work can be extracted, and it accurately captures the moment when the engine stops being an engine.

The fog thickens, and the driver stops looking at the map, which requires knowing every turn. He watches only what is visible right in front of the hood. So it is with the demon in a noisy world: the simple instantaneous link between what it sees and what it does proves more reliable than the most detailed memory of the road.