A Forced-Choice Precognition Experiment with Selected Cohorts

Bancel, P. A., Boban, J., Bensahra, A., & Varvoglis, M. (2025). A Forced-Choice Precognition Experiment with Selected Cohorts. Journal of Anomalous Experience and Cognition, 5(1), 14–46. https://doi.org/10.31156/jaex.26394

AI Assessment

A pre-registered precognition study whose confirmatory hypothesis was null; the eye-catching results are exploratory, and the strongest of them is circular. Two cohorts (experienced meditators and an interest-only group) ran a binary forced-choice precognition task on a new at-home platform. The pre-registered confirmatory test (an increase in session-to-session variance) failed for both cohorts, and the hit rate stayed at chance. The headline-grabbing variance signal comes from non-registered recruitment (“tryout”) data that was also used to design the protocol, and the authors conclude the formal studies found no direct evidence for psi.

Provenance

Source. Peer-reviewed open-access article (CC-BY): Journal of Anomalous Experience and Cognition, 5(1), pp. 14–46, 2025. Institut Métapsychique International, Paris. Supported by the Bial Foundation (grant 2020/356).

Study type. Pre-registered binary forced-choice precognition experiment on an at-home internet platform, comparing two selected cohorts; the designated confirmatory outcome was session-variance, not hit rate. Chance baseline is 50%.

Funding. Institut Métapsychique International (logistical and financial); Bial Foundation grant 2020/356.

Data availability. Session counts, trial totals, and Monte-Carlo iteration counts are reported in the article; the study was pre-registered.

Source basis. Figures confirmed against the published JAEX article (pp. 14–46), Method and Results sections.

What the paper reports

The study had two aims: to test a new collaborative platform (Psi@Home) for running at-home psi experiments, and to look for precognition in two cohorts – experienced meditators and an “Open” cohort selected only for their interest. Each of 80 formal sessions per cohort comprised 20 binary forced-choice trials (50% chance). The single pre-registered confirmatory hypothesis was an increase in the variance of session hit rates; it was not confirmed for either cohort, and the hit rate did not differ from chance.¹

The authors’ own conclusion: “Two pre-registered cohort studies found no direct evidence for a psi effect.” The significant findings are exploratory – chiefly a higher session-variance for the meditator cohort (p = .032) and a very large variance in non-registered recruitment data (p = .000031) whose hit rate nonetheless stayed at chance.

How it was run

• Two cohorts were recruited in 2022: experienced meditators and an “Open” cohort selected only for interest. 47 cohort members contributed 80 formal sessions per cohort; each session was 20 binary forced-choice trials.
• The task ran in a custom at-home application (the “Selfield”), described by the authors as a binary test of precognition: the participant predicts the future outcome of a randomly generated binary event, giving a 50% chance baseline.
• The platform (Psi@Home) handled randomization, scoring, and secure data return (AWS servers, per-user passwords, hidden computer-serial-ID checks, and a software checksum to detect tampering).
• The pre-registered confirmatory hypothesis was an increase in session-variance relative to mean chance expectation (MCE). Exploratory hypotheses were a positive hit-rate bias and a higher session-variance for the meditator cohort. A discard rule (sessions with fewer than 17 trials excluded) was pre-specified.
• Variance hypotheses were tested with chi-square statistics against MCE using large Monte-Carlo simulations (40k to 2M iterations); hit-rate hypotheses used one-tailed exact binomial tests.
• “Tryout” sessions (completed by candidates during recruitment) were analysed separately as non-registered exploratory data; a 50-session tryout result was what the authors used to set the formal sample size (80 per cohort) and to designate variance as the confirmatory measure.

Results, as reported

The single pre-specified confirmatory outcome (a session-variance increase) was null for both cohorts; the hit rate was at chance throughout. The significant variance results are exploratory (the cohort comparison) or non-registered (the tryout data). The paper’s abstract and results text number the exploratory hypotheses differently; the descriptions here follow the content rather than the labels.

Eleven-dimension audit

Pre-registration

A genuine strength: the study was pre-registered, with a single designated confirmatory hypothesis (session-variance increase), pre-specified exploratory hypotheses, fixed sample sizes (80 sessions per cohort), and a discard rule. The important caveat is circular provenance – the confirmatory hypothesis was chosen because a 50-session block of recruitment (tryout) data had shown a strong variance effect; that pre-registered confirmatory test then failed in the formal study.

Randomization

Each binary outcome is set by a pseudo-random process that is re-seeded every trial from the millisecond timing of two participant keystrokes, and is scored automatically against the prediction. The design is a true precognition design (the target does not exist when the choice is made). The authors flag, as an explicit limitation, the absence of an independent, automated calibration of this random process.

Sensory leakage

Because the target is a future random event, there is no contemporaneous target to leak. The at-home, unsupervised format removes a live experimenter but raises data-integrity questions, which the authors address with platform security (encrypted servers, per-user passwords, machine-ID checks, and a code checksum) while acknowledging that “security measures are never inviolable.”

Blinding

Scoring is an objective, automated hit/miss against a target generated after the prediction, with no human judge and no experimenter in the room. Blinding in the conventional sense is moot; the trade-off is the loss of direct supervision inherent to at-home testing.

Optional stopping

The formal study had a pre-registered fixed N (80 sessions per cohort) and a pre-specified discard rule, which binds stopping for the confirmatory test. The tryout data, by contrast, were analysed at accumulating points (10, then 50, then 90 sessions), an interim-look pattern that warrants caution in interpreting their significance.

Outcome measure

Unusually, the pre-registered confirmatory measure was session-to-session variance rather than the hit rate. This choice was driven by the tryout data, in which variance, not the mean, was elevated. The hit rate (the more familiar measure) was relegated to an exploratory test and stayed at chance.

Effect size

The confirmatory test produced no effect (both cohorts null). The significant exploratory and tryout results are variance effects (the tryout variance is roughly four standard deviations above expectation), while the hit-rate estimates all hover near 50% (ranging 48.8% to 51.5%) with none significantly above chance, so there is no mean-level effect to size.

Multiple comparisons

One confirmatory test was pre-registered, which is disciplined. The significant outcomes, however, come from a cluster of exploratory and non-registered analyses: a between-cohort variance comparison and tryout analyses examined at accumulating sample sizes. The authors do apply a Bonferroni correction to the tryout variance (a factor of 14, across the hit-rate and variance tests on seven datasets), giving an adjusted p = .00043 that remains highly significant; even so, these are exploratory, interim-look results rather than the pre-registered confirmatory test. The authors also report a robustness check that materially qualifies this signal: the tryout variance rests on a handful of sessions. Removing the 6 most extreme of the 90 tryout sessions raises the variance p from .000031 to .032 (χ²(84) = 106.7), and dropping 2 further sessions raises it to p = .12, which the authors call a nearly complete attenuation of the variance anomaly arising from a few extreme hitting and missing sessions.

Internal replication

The two formal cohorts are effectively two attempts, and both were confirmatory-null. The sharpest internal tension is that the tryout data (the same participants and protocol, collected during recruitment) showed a large variance effect that the formal study did not reproduce – a non-replication the authors attribute to differing participant attitudes across the two collection periods.

External replication

The work sits within the contested forced-choice precognition literature, where positive “feeling the future” claims (Bem, 2011)² meet sustained statistical and methodological critique (Rouder and Morey, 2011),³ and a cumulative meta-analysis reports only a small overall effect.⁴

Transparency

High. The study is pre-registered and reports its confirmatory and exploratory structure, full statistics with Monte-Carlo iteration counts, the discard rule, platform-security details, and a frank conclusion that the formal studies “found no direct evidence for a psi effect.” The authors openly flag the tryout-to-formal discrepancy and offer a psi-hitting/psi-missing interpretation rather than burying it. Their own stated limitations include the macOS-only platform (which slowed recruitment), the security exposure of installed software to experimenter fraud (mitigated by encryption, key-coding, and duplicate databases, but in their words never inviolable), and the absence of an independent, automated calibration of the trial random process.

The adversarial record

• The result that matters is the pre-registered confirmatory test, and it was null for both cohorts; the hit rate never left chance (48.8% and 51.5%). By the study’s own pre-registered standard, this is a non-detection.
• The striking numbers are exploratory or non-registered. The strongest, the tryout variance (p = .000031; Bonferroni-adjusted p = .00043), is circular: that same recruitment data is what the authors used to choose variance as the confirmatory measure and to set the sample size, and it was examined at accumulating sample sizes. A confirmatory test motivated by a dataset cannot also be confirmed by it. On the authors’ own trimming analysis it is also fragile, nearly vanishing once a few extreme sessions are removed.
• Reading against the study: the only signal is in variance, never in the hit rate, which the authors interpret as a psi-hitting/psi-missing mixture – an account that can absorb a null mean and is hard to falsify. A skeptic reads the excess variance as ordinary heterogeneity in unsupervised at-home data rather than precognition. The platform is novel and well-secured, but the precognition claim is, on the pre-registered evidence, unsupported.