Ray Hyman’s 1985 critical appraisal of the ganzfeld database identified four categories of methodological concern: (1) inadequate randomization, targets selected by human shuffling or pseudo-random methods susceptible to pattern exploitation; (2) sensory leakage, experimenters who knew the target could inadvertently cue receivers through handling, timing, or verbal behavior; (3) multiple-analysis effects, researchers testing many dependent variables and reporting only significant ones; and (4) selective reporting, studies with null results remaining unpublished. The autoganzfeld system addressed each: a noise-based hardware RNG eliminated predictable target sequences (mitigating randomization artifact); computer-controlled target delivery with the experimenter remaining target-blind during the session eliminated direct cueing (eliminating the primary sensory-leakage pathway); pre-specified primary and secondary hypotheses with exploratory analyses labeled as such addressed multiple-comparison inflation; and inclusion of every session since 1983 inauguration, including pilots, addressed the file-drawer problem.²¹

The autoganzfeld’s target library consisted of 40 judging pools, 20 dynamic (film clips) and 20 static (art prints, photographs, magazine advertisements), recorded on four VHS-format videocassettes, with five dynamic and five static pools per tape alternating in sequence. A signal recorded on one audio track of each cassette allowed the computer to access and cue specific targets precisely. The hardware RNG, a noise-based entropy source, selected the target pool and specific target for each session; the selection was logged to disk before the session began, creating a time-stamped audit trail that could not be altered without leaving a detectable gap. The receiver’s judging was also computer-guided: after the mentation period, the receiver viewed all four pool members on the monitor and entered rankings directly, with the computer recording responses without experimenter transcription, eliminating transcription error and experimenter-recording bias as artifact pathways.¹

The autoganzfeld system’s use of computer-controlled VCR playback made it the first ganzfeld protocol capable of presenting dynamic targets, film clips with both visual and auditory content, as a standard experimental condition rather than an ad hoc variation. Berger and Honorton hypothesized that dynamic targets would yield stronger hit rates than static targets because they depict familiar, emotionally engaging situations of the kind associated with spontaneous psi reports, engage the sender’s attention more continuously, and provide information across both visual and semantic channels simultaneously.¹ This dynamic-vs.-static comparison was pre-specified as a secondary hypothesis in the autoganzfeld series, not a post-hoc observation, and was tested via the same pre-specified binomial framework as the primary hit-rate hypothesis.²

A central problem Berger identified in the RNG-psi literature of the early 1980s was the absence of standardization: basic terminology varied across investigators, hardware RNG characteristics were frequently undescribed in ways that prevented cross-laboratory comparison, and subject recruitment methods were inconsistently reported. PsiLab II addressed this by embedding the experimental design, data-collection protocol, and motivational features within the computer program itself, making the protocol identical across sites regardless of experimenter. The starter kit included a hardware RNG board (plugging into an Apple II peripheral expansion slot), machine-language drivers, statistical analysis software, and ready-to-run experiments. The RNG output was subjected to periodic statistical tests under control (no-observer) conditions to verify adequate randomness before experimental data were collected, partially addressing the artifact that a biased RNG could produce apparent psi effects without any anomalous cognition.³

The primary hypothesis, that subjects would identify the actual target at above-chance rates, was tested via exact binomial probability with p = .25 (four-alternative forced choice), q = .75, alpha = .05 one-tailed. Across 358 sessions, 123 direct hits (34.4%) yielded z = 3.90, exact binomial p = .000048.² Two secondary hypotheses were pre-specified: (a) dynamic targets would yield a significantly stronger hit rate than static targets, and (b) success rate would be significantly higher when sender and receiver were friends or acquaintances rather than strangers. All remaining analyses were explicitly labeled exploratory. The database included every session since the 1983 inauguration, including pilot series, to prevent selective-reporting (file-drawer) inflation of the effect estimate. The effect size and direction were later compared against the earlier pre-autoganzfeld meta-analytic database to assess whether the stricter protocol produced a different result from the looser earlier studies.⁵

Two “First-Timers” series (FT1 and FT2) examined whether inexperienced ganzfeld participants could produce significant outcomes on the autoganzfeld system, a question with direct implications for replicability, since most potential replicating labs would not have access to pre-screened experienced participants. FT1 (N = 50 novice participants, one session each) produced nonsignificant results on both direct-hit and sum-of-ranks indices (p[direct hits] = .38, p[sum of ranks] = .26). FT2 (23 of 50 sessions completed at time of reporting) showed both indices approaching significance (p[direct hits] = .04, p[sum of ranks] = .046).⁴ The Series Manager program specified the number of subjects, sessions per participant, and other experimental parameters in advance, providing a fixed-N stopping rule for the completed FT1 series. The FT2 interim result is exploratory given the incomplete series at time of reporting.

A specific skeptical concern about the autoganzfeld program, raised after the main results were published, was whether the target-pool distributions themselves were uniform, or whether certain targets appeared more frequently than others in ways that could inflate hit rates through response-bias (participants preferring certain target types regardless of psi). Bierman, Broughton, and Berger (1998) conducted a forensic audit of the PRL autoganzfeld target and target-set distributions, examining the randomization integrity of the actual session records. This audit directly addressed the artifact of non-uniform target sampling: if some targets were selected more often by the RNG than others, and participants had a systematic preference for those targets, apparent hits could accumulate without anomalous cognition. The audit’s findings on distribution uniformity were reported in the Journal of Parapsychology. This represents a case of the system’s own architect participating in a post-hoc adversarial check of the randomization infrastructure he designed, partially addressing but not fully eliminating the concern, since the audit examined distributional properties rather than the full sequence of RNG outputs.