A high-powered, transparent replication of Bem's precognition experiment_

The scale

This project set out to test the most discussed precognition claim of recent years: Bem's finding that people can guess, above chance, which side of a screen a future image will appear on. It ran three pre-registered replication studies with 26,483 participants and 420,472 critical trials.

The controls were unusually strong. Data were deposited in real time to a version-controlled repository so the collection could not be quietly revised, and the protocol was designed jointly by sceptics and proponents, Bem included. A random event generator was substituted for participants as a check on the process itself.

The result

Study 1 did not replicate the effect, and pointed weakly the other way, at 49.5 per cent. Study 2 was built to test that reversed result properly, and it reproduced it, at 49.65 per cent (p = .013). Study 3 then failed to reproduce Study 2, landing at 50.07 per cent, indistinguishable from chance.

So the one anomaly the project confirmed did not itself survive the next replication.

The reading

By ordinary scientific standards, an effect that appears once under confirmation and then vanishes on independent repetition is more plausibly a quirk of the method than a signal that defies physics. The authors reach that conclusion themselves, and treat the project mainly as a demonstration of how to test the reliability of a process, not just a hypothesis.

Why it matters to us

This is the most rigorous test of the specific Bem claim yet run, and it is close to the platform we are building: one instrument, uniform collection, pre-registered analysis, transparent data, and controls that test the process as well as the person.

The lesson we take from it is not that precognition is impossible. It is that a signal only counts once it survives an independent replication, and that most do not. That is the bar we are building to clear, not to dodge.