Fifty years of random number generators held up to the mind: what was found, what was learned, and what was built.
A history with the receipts attached: what researchers found when they pointed random number generators at intention, what the critics answered, and the strange objects left behind. Every claim carries its citation, and the failures stand next to the successes. The statistics live on the method page; the theory on the science of consciousness.
A machine whose next bit nothing in physics can predict. The cleanest question ever put to chance.
A random event generator samples a process quantum mechanics declares unknowable in advance: radioactive decay, electrons tunneling through a junction, the vacuum fluctuating. The bits pass a balancing stage that cancels any steady electronic bias, leaving a perfectly fair coin, and because its null behavior is exactly computable, any departure can be priced in odds. The randomness itself is settled engineering: a national laboratory has streamed quantum randomness from vacuum fluctuations since 2012, and NIST signs a pulse of certified entropy every minute. Nobody disputes the coins. The dispute is whether mind ever correlates with how they fall. An instrument of this class runs in your browser, and its code is open.
A Boeing physicist wires four lamps to radioactive decay and asks people to call the next flash. Odds in the first report: two in a billion.
Helmut Schmidt built the first quantum machine for this question in the late 1960s: a counter stopped by strontium-90 decay, lighting one of four lamps for subjects to predict. His first experiment ran 691 hits above chance in 63,066 guesses, odds near two in a billion (Schmidt, 1969); willing a binary generator around a circle of lamps later reached 52.4 percent, odds of ten million to one (Schmidt, 1970a). He shut his cat in a cold shed with a heat lamp switched by the generator: the lamp ran on 51.3 percent of 9,000 trials, while empty-shed nights sat at chance (Schmidt, 1970b). Cockroaches on a shock grid received more shocks, and Schmidt suspected his own dislike of them was the active ingredient: the first warning that the experimenter cannot be cut out of this loop. His last decade made the effect fraud-proof instead of larger, with skeptical outside observers holding duplicate tapes: five such studies combined to z = 3.67, odds near eight thousand to one (Schmidt, 1993). His own verdict: “a weak but real effect.”
Twelve years, 91 ordinary people, two and a half million trials. The lean: about one part in ten thousand.
In 1979 Robert Jahn, dean of engineering at Princeton, opened the PEAR laboratory with Brenda Dunne. Its benchmark defined the discipline this site inherits: 200-bit trials under pre-stated intention, high, low, or baseline, interleaved. Over twelve years, 91 volunteers with no claimed gifts ran 2.5 million trials; the high and low records separated by 0.042 bits per trial in the intended direction, and pooled across all the true-random machines the composite reached seven sigma (Jahn et al., 1997). The same operators aiming at deterministic pseudorandom sources produced nothing. A ten-foot wall of 9,000 balls rattling through 330 pegs shifted the same way (Dunne, Nelson & Jahn, 1988). But a three-laboratory consortium rebuilt the experiment with 227 fresh operators and fell an order of magnitude short of significance (Jahn et al., 2000). The lab closed in 2007. Jahn: “If people don’t believe us after all the results we’ve produced, then they never will.” The physicist Robert Park, for the opposition: an embarrassment to science. The archive is public.
The archive’s fine grain taught stranger things than its headline: the effect had personality.
Read closely, the Princeton data refused to behave like a force. Baselines were too quiet, hugging the mean more tightly than chance allows, as if stillness were also an aim (Dunne & Jahn, 1995). Individuals showed consistent styles the lab called signatures, and a few prolific operators carried much of the effect. Two-thirds of the men separated their records with intention, only a third of the women, whose effects ran larger and stranger (Dunne, 1998); bonded couples produced nearly seven times their solo results (Jahn & Dunne, 2005). Distance did nothing out to nine thousand miles, and efforts hours before or after the machine ran scored as well as efforts in the moment (Dunne & Jahn, 1992). And when the lab quietly swapped in trials of two million bits, the effect inverted, twice (Dobyns et al., 2004). Whatever these archives recorded did not scale with energy, distance, or dose. It tracked persons and meaning, which is either a profound clue or the signature of artifact.
Then they unplugged the generators from the desk and carried them into ceremonies. The claim: rooms have weather.
In the 1990s the Princeton group took portable generators into the field and read the record against time-stamped logbooks. Ten venues, from research meetings to ritual circles, compounded to odds of five thousand to one (Nelson et al., 1996). The follow-up pre-stated the hypothesis: resonant gatherings should move the generators, pragmatic ones should not. The resonant set, rituals, sacred sites, chanting in the chambers of Egyptian temples, deviated at p = 2.2 × 10⁻⁶; conferences and business meetings sat at chance, as predicted (Nelson et al., 1998). At funerals the trace did not lean; its variance shrank, as if the room had gone still. Radin ran the same idea from Las Vegas through the Academy Awards and the O.J. Simpson verdict (Radin, Rebman & Cross, 1996), and a team hauling generators to Burning Man watched the deviation peak at 3.77 standard errors minutes into the Man Burn (Radin et al., 2017). Exploratory by its own admission, and the seed of something far more disciplined.
Sixty-five generators around the world, one pre-registered question, seventeen years. Composite odds: about a trillion to one. What it means is the dispute.
The Global Consciousness Project grew from two funerals: twelve independent datasets deviated together while Princess Diana was buried, and stayed silent for Mother Teresa a week later. From 1998 Roger Nelson ran a permanent network, one 200-bit trial per generator per second, every event’s analysis window registered before the data arrived. The formal series closed in 2015 at 500 events: composite Z = 7.31, odds near a trillion to one (Nelson & Bancel, 2011). The texture is honest: most single events, the Paris attacks among them, showed nothing. The claim lives entirely in the accumulation, exactly where a real but tiny effect would have to live.
September 11 is its most contested day: the formal test returned p = 0.028, re-analysts found the result vanished if the window moved, and NASA’s Jeffrey Scargle judged the whole subject “still completely in the exploratory phase”, all published side by side (Nelson, 2002; May & Spottiswoode, 2001; Scargle, 2002). The deepest turn came from inside: Peter Bancel, the project’s own analyst, concluded after seventeen years that the anomaly is real but follows the experimenters’ goals rather than any field (Bancel, 2017), and Nelson’s reply ran beside it. Either reading is an anomaly of mind; the argument is about whose. The network lives on as GCP 2.0, hosted in ordinary homes.
Eighty broods of chicks and a robot that wandered. The claim: the chicks pulled it home.
In the 1980s the French physician René Peoc’h let newly hatched chicks imprint on a tychoscope, a small robot that moves in random steps. Caged where they could see it, their robot spent two and a half times as long on their side as it did for controls (Peoc’h, 1988). In the famous variant the room was dark and a lit candle rode the robot: in 71 percent of eighty trials it kept the light near the chicks, while with an empty cage its path stayed random (Peoc’h, 1995); moving the generator 23 kilometres away, he reported, changed nothing (Peoc’h, 2001). The counter-record is instructive: skeptics simulated similar traces from wall effects alone, Peoc’h declined outside observers, and no independent laboratory has replicated it. Princeton’s own generator-driven robot, tested as the lab closed, gave one series with the aim and one against (Jahn et al., 2007). A tantalizing, unfinished corner of the record.
Bits recorded to tape a day earlier, seen by no one. The effort still found them.
Schmidt’s deepest experiment barely fits in the language. He recorded a generator’s output to two identical cassettes, observed by no one, one locked in a safe. Days later a subject listened and tried to will the clicks toward one ear: across 832 blocks, 54.6 percent leaned the willed way, and the safe copies matched, while unexposed controls stayed at chance (Schmidt, 1976). Stranger still: targets played back four times leaned further than targets played once, and if anyone observed the bits before the attempt, the effect vanished (Schmidt, 1985). Schmidt did not read this as changing the past but as quantum measurement unfinished: the listener is the recording’s first observer. “It appears unreasonable,” he wrote, “to talk about physical reality before observation.” The theory this touches is laid out in the science of consciousness.
Every study anyone could find, weighed together, twice. The two verdicts do not agree.
In 1989 Radin and Nelson weighed 597 experiments by 68 investigators in Foundations of Physics: controls at chance, experiments not, by about three parts in ten thousand per bit, with quality predicting nothing; cancelling it would take a drawer of 54,000 unpublished nulls (Radin & Nelson, 1989). In 2006 a second team re-weighed 380 studies in Psychological Bulletin: significant but very small, wildly heterogeneous, and “publication bias appears to be the easiest and most encompassing explanation” (Bösch et al., 2006). The reply ran in the same issue: the critics’ fixed-bias assumption contradicts the data and guarantees their heterogeneity, three enormous studies dominated their model, and a survey found about 59 unreported studies, not the 1,500 required (Radin et al., 2006). The rejoinder conceded rare ground: “we ourselves remain undecided.” Both sides ended by demanding the same thing, registration and permanent records. A Munich laboratory now runs every study into an open registry with no file drawer: 34 studies, 40,000 participants, effects that appear and fade under preregistration (Maier & Dechamps, 2018). The looking is finally honest, and this site’s record-keeping is the same discipline.
The archive kept spilling into objects: lamps, text messages, art, an app with ten million users.
Princeton’s younger generation founded Psyleron and sold the instrument itself: a laboratory-grade generator for anyone’s desk, a lamp whose color drifted on quantum tunneling noise, and SyncTXT, which texted you whenever your stream ran strange. The lab’s ArtREG had two pictures fight for the same pixels and is remembered for its honesty: no correlation with intention, though the few images that ran beyond chance all carried archetypal charge (Jahn et al., 2000). The strangest descendant is Randonautica, launched in 2020: state an intention, draw quantum randomness, walk to the coordinates. Ten million users in five months, one suitcase of human remains found by teenagers on a Seattle beach, a coincidence by every account, and a physicist’s verdict on the record: “people using big science words to sound magical.” Meanwhile the patent office holds granted claims on “devices responsive to influences of mind” (Wilber, 2016), a bet on file that the effect is real enough to engineer.
Fifty years of measurement, three readings of the ledger, and no verdict.
The generators are sound, the archives are public, and the composite deviations are in the records. Three readings stand: that mind faintly touches chance; that the anomalies are real but follow the experimenters rather than any field; that selection and publication habits made the whole ledger. Each has serious people behind it. This site is a wager on how it gets settled: pre-declared, permanent, public measurement, run by people who wanted to see for themselves. The instrument is in your browser now. The record never resets.