Discovery Lab Answers

Autoganzfeld vs. earlier manual ganzfeld

Coverage first. This comparison is drawn from the ESP-Nexus library’s holdings, which include 57 of the 78 studies in the standard ganzfeld meta-analytic database plus the published abstract of one more. The 16 studies still missing skew toward null results, so any summary drawn from these holdings — including this one — runs somewhat more positive than the full published database.

The two databases

The autoganzfeld program at Psychophysical Research Laboratories (PRL) was a deliberate methodological response to critiques of the earlier manual ganzfeld — target selection, stimulus presentation, blind judging, and data recording were placed under computer control to close the routes for sensory leakage and recording error that Ray Hyman had documented in the pre-1986 studies. The automated testing system was designed and built by Rick E. Berger; Charles Honorton led the PRL research program. You can see the site’s treatment of the system on the Automated ganzfeld methodology and system design page.

How the numbers line up

The headline point is that the automated PRL series produced hit rates in the same neighborhood as the earlier manual meta-analyses, not a collapse under tighter control — but the individual studies and later replications vary widely, and several are null.

Bem & Honorton’s autoganzfeld pooled hit rate (~0.32) sits close to the earlier manual database’s frequently-cited ~1/3 accuracy figure that Bem himself summarized, and Williams’s post-communiqué collection of 59 independent studies landed at ~0.31 — the argument being that the stricter automated protocol did not erase the effect and that more than half of fifteen contributing labs produced hit rates within the PRL range.

The nulls and the disagreement

The evidence set is not uniform, and honesty about that matters:

• Milton (1999) analyzed 30 new ganzfeld studies (1987–1997) and found essentially no effect (ES 0.013, p = .24) — a null result on the post-Bem/Honorton generation. The same paper’s extended cumulation to March 1999, once Dalton (1997) was included, reached ES 0.038 (p = .011).
• Goulding, Westerlund, Parker & Wackermann (2004), the first fully digital autoganzfeld with real-time external judging, returned a 23% direct-hit rate (p = .386) — below the 25% chance expectation and non-significant. That same group’s earlier manual ganzfeld work in Göteborg had produced a 36% hit rate across five studies (p = .0012), so within a single research group the automated real-time-judged version underperformed their own manual results.
• Across the full structured set, reported directions disagree (some positive, some null or below chance), between-study heterogeneity is high (I² ≥ 50% in the pooled analyses), and four result rows report a null or below-chance effect.

On the Milton & Wiseman (1999) row: those figures (ES 0.53) are the original study’s, relayed inside Tressoldi’s (2010) later analysis, not Tressoldi’s own finding — the original Milton & Wiseman meta-analysis of 30 studies is generally cited as the prominent non-significant replication of the ganzfeld, and the relayed figure should be read against that context, not as an endorsed effect.

Skeptical critiques

What critics argue. Hyman (1985) argued the earlier manual ganzfeld database was compromised by methodological flaws — sensory leakage, inadequate randomization, optional stopping, and selective reporting — such that its pooled excess over chance could not be taken as evidence of psi. He and Honorton later co-authored a joint communiqué (1986) specifying stricter guidelines, and the autoganzfeld program was explicitly built to meet them.

What the experimental data show. The PRL autoganzfeld, run under those guidelines, produced a pooled hit rate near 0.32 with the full database (including pilot series) reported to prevent file-drawer inflation. Bem & Honorton reported that inclusion of two incomplete studies did not create an optional-stopping problem (overall z moving from 2.89 to 2.76, p = .003). But Milton’s (1999) later 30-study cumulation and Goulding et al.’s (2004) digital replication show that the effect did not appear uniformly in subsequent automated work.

Analysis. The dispute turned on whether automation removed the flaws Hyman named. The PRL series and Williams’s post-communiqué collection report above-chance hit rates under computer control; Milton (1999) reports a null across 30 new studies, and Goulding et al. (2004), using the same protocol family but with real-time external judging, report a below-chance rate — and notably a lower rate than that same group’s own earlier manual ganzfeld. Independent replication of the PRL-magnitude effect exists in some labs and not others.

For the site’s synthesis, see Autoganzfeld and Ganzfeld.