The PEAR Random Event Generator Programme

The laboratory's core experiment was undramatic by design. An operator sat in front of a machine producing a stream of random binary digits from electronic noise, stated in advance whether they intended to push the output high or low, and then tried. A run consisted of thousands of bits, a session of many runs, and the database of many thousands of sessions across nearly three decades. Nothing visible happened during any of it. The claim was that when everything was summed, the totals in the intended direction exceeded the totals in the baseline condition by an amount that was tiny in each run and, in aggregate, statistically significant. Robert Jahn characterised the magnitude as being on the order of one part in ten thousand.

The methodological bet behind this design is worth stating clearly, because it is the source of both the programme's appeal and its trouble: if a real effect is very small, the only way to see it is to collect enormous quantities of data, but enormous quantities of data are also the condition under which very small biases in equipment, in analysis and in stopping rules become visible as apparent effects. The laboratory also ran a second line of work on remote perception, closer to the SRI protocol than to the generators, which is less often discussed and was reported to smaller effect.

Two pieces of evidence closed the case for most readers outside the laboratory. The first was the PortREG consortium experiment conducted around 2000, in which PEAR joined research groups at Giessen and at Freiburg to run one agreed protocol on PEAR's own hardware and software at all three sites. The point of the design was to remove the objection that only one laboratory could produce the effect. The outcome was trends in the expected direction at each site, none of them statistically significant, and a combined result that did not support the original claim.

Jahn and Dunne argued at length that the shared protocol had constrained the operators in ways the original work had not, which is a defensible reading and also the reading that makes the hypothesis unfalsifiable. The second was the 2006 meta-analysis by Bösch, Steinkamp and Boller in Psychological Bulletin, covering the whole random generator literature. It found a small significant overall effect and, alongside it, a strong inverse relation between effect size and sample size, meaning the largest studies came closest to chance. That asymmetry is the standard fingerprint of publication bias, and the authors concluded the literature was better explained by selective reporting. Proponents contest the correction, not the asymmetry.