Parapsychology and psi research, oriented

Researchers group the claims under one word, psi. Telepathy is claimed access to another person's thoughts, clairvoyance is claimed perception of an object or event when the senses have nothing to work with, and precognition is claimed anticipation of a future event that no inference could predict. Remote viewing, describing a distant place, is treated as a kind of clairvoyance. Calling a result "anomalous" means no accepted mechanism explains it. It is a description of the gap, not an explanation.

The basic test is forced choice: guess which of a fixed set of targets is correct. With four options, you would succeed 25 percent of the time by guessing alone, so the claim lives entirely in the difference from 25. A 1989 meta-analysis of 309 forced-choice precognition experiments, with more than 50,000 participants, found a highly significant result built on an effect of about 0.02 per trial. Millions of guesses can make a tiny lean look overwhelming.

The common slip is to treat significance as size. Say you read a report stating "p less than one in a billion." That says chance alone is unlikely to produce the data; it does not say the effect is large, and it does not say what caused it. Later work moved to reaction times and physiological measures, which are harder to fake but offer more analysis choices.

Rule: before judging any psi result, find the chance rate, then ask how far above it the result sits and how many trials it took.

The society was founded under Professor Henry Sidgwick, and an American society followed in 1884. Its writers used telepathy, from Greek words for "distance" and "feeling," as a provisional name for the claim under test. Frank Podmore's 1894 survey of their evidence opens not with marvels but with sources of error: fraud by signaling, hyperesthesia (unusually keen senses, such as recognizing cards by touch after practice), muscle-reading through contact, and shared guessing habits.

That last one is the case you are most likely to miss. An astronomer found two star surveyors agreeing on magnitudes far more often than chance predicted. It looked like thought-transference until analysis showed both preferred the same round numbers. No deception was involved, which is the point: many leaks need no liar.

The society's largest project, a census of hallucinations, asked 17,000 adults about waking hallucinations. It counted about 32 well-attested cases of a recognized apparition within twelve hours of that person's death, among an estimated 1,300, roughly 1 in 40 against a chance rate near 1 in 19,000. The estimate leaned on a judgment call: recent experiences were reported far more often than old ones, so the authors multiplied their counts by four to allow for forgetting. Every input to the ratio rested on testimony.

Rule: when an argument from odds depends on corrected counts, ask how much the conclusion moves if the correction is wrong.

Rhine wanted to move extrasensory perception out of the séance room and into the laboratory, with controls and statistics. When critics attacked his arithmetic, the president of the Institute of Mathematical Statistics said in 1937 that the objection could not rest on statistical grounds. Critics turned to methods instead, and what finally turned most scientists away was that other laboratories could not reproduce the results.

The era also left a habit worth recognizing. Rhine read a drop in scoring during a session, the decline effect, as a hidden sign of psi even when the total was at chance. Soal, after failing to replicate Rhine, rescored guesses against the next or previous card and found two "displacement" subjects. Below-chance scoring was counted too. If high, low and declining scores all count, the claim forbids nothing.

The statistics need two numbers. A p-value says how often chance alone would give results this strong; at the usual 0.05 threshold, pure chance passes about once in twenty tests. Effect size says how big the departure is. The statistician Jessica Utts gave an example: 70 boys in 100 births against an expected 51 percent gives p = 0.0001, while 7 in 10 gives p = 0.19. Both are 70 percent, so the small study replicated the effect even though it was not significant.

Rule: judge a replication by whether its effect size matches, not by whether its p-value crosses 0.05, and write down before testing which outcome would count against the claim.

A checklist from the 1995 government-commissioned review of remote viewing captures the modern standard. Nobody who knows the target should contact the participant until the response is secured. Nobody who knows the target, or even whether a session went well, should talk to the judge until judging is done. Targets should come from a tested randomizer. Judging should use an unhandled duplicate of each target, so fingerprints or creases cannot mark the real one. Stopping rules, reasons for dropping data and the analysis should all be fixed in advance.

Early remote-viewing work shows why. Sites were used once each, and some transcripts mentioned an earlier session's target, so a judge could match later sessions by elimination with no psi at all.

Deliberate fraud is the other half. The society's early founders saw no ordinary explanation for the Smith-Blackburn telepathy trials until Blackburn confessed and explained the trick. The Soal-Goldney experiments were hailed as decisive for years; critics proposed elaborate conspiracies, but the parapsychologist Betty Markwick eventually showed how Soal had actually cheated, with a method simpler than anyone had guessed.

The trap is reasoning from your own imagination, and it catches skeptics and believers alike. "I cannot see how this could be faked" describes the observer, not the experiment.

Rule: judge a protocol by whether it would still work if a participant were cheating and the experimenter were hoping, not by whether anyone has spotted a flaw yet.

Half ping-pong balls over the eyes and white noise in headphones create a uniform "whole field" while the participant speaks every image that comes. Afterward they rank four pictures or clips, one of them the target. A 1985 meta-analysis of 28 studies reported 38 percent hits. Its critic Ray Hyman and its author Charles Honorton then wrote a joint communiqué setting stricter standards, and a computer-controlled version, the autoganzfeld, reported 106 hits in 329 sessions, 32.2 percent. Even then Hyman found that hits came only on a target's second or later use, a pattern a randomization flaw would produce.

Meta-analysis pools studies, and every pooling choice you make can move the answer. A 1999 analysis of 30 later studies found no significant effect; a reanalysis of the same trials counted 327 hits in 1,198 sessions, about 27 percent, which was significant. Publication bias is real here too: one parapsychology unit reported 70 percent of its significant studies but only 15 percent of the others.

A 2016 simulation added questionable practices at realistic rates, such as restarting after a bad start or folding good pilots into the main study. It could not erase the excess, leaving about 2 percentage points, but detecting that reliably needs about 700 participants, more than any single ganzfeld study had run. A 2024 registered meta-analysis estimated an effect size near 0.08.

Rule: ask what each pooled figure includes, how it was weighted, and how big a single study would need to be to test it.

In 1972 two laser physicists at the Stanford Research Institute (SRI), Russell Targ and Harold Puthoff, drew CIA interest with informal demonstrations. A CIA physicist who managed the early work, Kenneth Kress, later described results that were inconsistent across subjects and repeats. In one case a viewer, given only coordinates, drew a large crane at a Soviet test site that analysts judged accurate, yet most of his other details were wrong. Accurate and false details came mixed, with no way to tell them apart in advance.

Later work ran under a Defense Intelligence Agency program called Star Gate. In 1995 the CIA asked the American Institutes for Research to evaluate it, and the review paired a statistician sympathetic to psi, Jessica Utts, with the skeptical psychologist Ray Hyman. Judges ranked five photographs against each description, so chance gave an average rank of three. Utts reported small effect sizes, about 0.21 for SRI and 0.23 for the later contractor.

If you read only one finding from the review, read this one. Both reviewers agreed the effect was statistically significant and that methods had improved. They disagreed about cause. The same viewers, the same target set and a single judge, the principal investigator, were used throughout, so no one could say where the effect came from. Operationally, reports were vague, sometimes edited to fit known background cues, and never used to guide an operation.

Rule: keep two questions apart, whether a result beats chance and what produced it, and ask which parts of a study were varied and which were held fixed.

Bem took well-known psychology effects and reversed their order. In his best-known task, a participant picks one of two on-screen curtains; only afterward does the computer randomly decide which hides an erotic picture. Chance is 50 percent. Because the target does not exist when the choice is made, no sensory leak is possible, which is the design's strength. What remains open is everything else: the randomizer, the analysis and how many variants were run.

A 2016 meta-analysis by Bem and colleagues pooled 90 experiments from 33 laboratories and reported an effect of 0.09, or 0.06 for independent replications alone. Quick-response tasks did better than slower ones, but that split was made after the fact.

Critics found specific weaknesses. Bem reported one-tailed tests throughout. In his recall study, misspelled words were corrected by someone who could see whether each was a practice or a control word, a path for unconscious bias. Treating the stimuli as a random sample, as one statistical team urged, weakened a reanalyzed result. Three preregistered exact replications of the recall study, with 150 participants and blind coding of every unclear word, found nothing: combined p = 0.83.

The trap is counting papers. When you see ninety studies, they can look decisive, but their size, flexibility and filtering matter more than their number.

Rule: weight each study by whether its analysis was fixed in advance and how large it was, not by how many studies agree.

The Transparent Psi Project started from the most common objection to any result: the other side never agreed to the method. A panel of 29 experts, 15 who accepted Bem's psi reading and 14 who did not, approved the protocol in advance. Data went straight into a public third-party repository as it was collected, software ran on a central server so no site could alter it, sessions were video-recorded and outside auditors checked the work. Stopping points, and what each outcome would mean, were fixed beforehand.

With 2,115 participants and 37,836 trials, the hit rate was 49.89 percent, against Bem's 53.07 and chance at 50. An equivalence test showed the true rate was below 51 percent, which is how a study shows absence rather than just failing to find something. Individual scores also matched pure guessing, with no consistent high or low scorers.

Micro-psychokinesis, the claimed influence of intention on a quantum random number generator, followed a similar path. A 2006 meta-analysis found a tiny, inconsistent effect, and a 2000 replication of a benchmark study, run with the original laboratory, failed. In 2018 a team whose own earlier study had looked positive ran 12,571 people online and found evidence favoring no effect about ten to one.

If you take one habit from this chapter, make it spotting the rescue: a pattern noticed after a null result is a new hypothesis, never a save of the old one.

Rule: one large, preregistered, jointly designed test outweighs many small favorable ones.

A Bayes factor of 3 means the data fit "there is an effect" three times better than "there is none." What you should believe afterward is your own starting odds multiplied by that factor, so the same data move different readers different distances. In one worked example, someone who gives psi a 20 percent chance holds prior odds of 0.25; after a factor of 3, they reach 0.75 and still lean against it. Someone who starts at one in a million needs a factor in the millions. That arithmetic is the honest form of "extraordinary claims require extraordinary evidence."

Proponents accept a higher bar because psi has no mechanism consistent with current physics and biology, but argue that evidence should still be examined, noting that aspirin was used long before anyone explained it. The 2016 precognition meta-analysis reported a Bayes factor of about 5 billion with an effect of only 0.09. A very large data set can do that for a tiny departure. Under the model used, the factor says only that the pooled data depart from pure chance, not whether bias, filtering or a shared flaw explains the difference. Results also depend on the alternative hypothesis chosen, which is why Bem's critics and defenders reached opposite Bayesian verdicts on the same data.

Ioannidis showed that whether a significant finding is true depends on prior odds, power and bias. Small studies, small effects and flexible analyses all lower it.

Rule: read any Bayes factor alongside the effect size, the alternative it assumed and the ways other than psi to depart from chance.

Heuristics are quick mental shortcuts that usually work and sometimes mislead, and everyone uses them. Studies of paranormal belief find that the reasoning skill most closely tied to belief is judging whether a sequence is random. One interpretation is a readiness to see coincidences as meaningfully connected, so chance events feel like signals. Believers have also made more conjunction errors, rating two events together as likelier than one of them alone.

Consider a case you may know. You dream about an old friend, and the next day the friend calls. The story is vivid and detailed, and detail makes it feel more probable, though each added detail makes the conjunction less likely, not more. Restated neutrally, it is one of many daily thoughts matching one of many daily events.

Fairness matters here. One study found that people who report more unusual perceptual experiences tend to hold stronger beliefs, and that belief seemed to give those experiences a framework. The trait measured belongs to a set of personality traits found across the general population; it is not a diagnosis. The experiences are real experiences, and the question is the explanation. The authors also note their data were cross-sectional, so they show which things go together, not which causes which. Belief types differ too: weaker statistical reasoning tracked traditional beliefs more than New Age ones.

Rule: before calling a coincidence unlikely, count how many chances there were for some match to happen.

Utts argued that similar effects across laboratories, experimenters and countries could not all come from one flaw, and that fraud would need many people cooperating. Hyman replied that each generation's best evidence, from early telepathy trials through Rhine's cards and Soal's experiments, was later discarded, while Newton's prism experiment still works. Matching effect sizes, he added, are only departures from a chance model, which many causes can produce. He and the parapsychologist John Palmer argued the field needs a positive theory that says when psi should and should not appear. The 1995 panel's wording is the careful summary: a significant anomaly, with no demonstration that a paranormal ability causes it.

Hyman also specified what would change his mind: the effect holding with new participants, new targets, several independent judges and independent laboratories. Stating such a standard in advance binds both sides.

You can apply the same tests to the next headline. Was the hypothesis, sample size and analysis registered before data collection? Would critics have approved the design? Could outsiders see the data as it arrived? How large is the effect against the chance rate, and how many trials support it? Was the procedure documented well enough for someone outside the room to confirm it?

Suppose you read a claim of 52 percent hits on a two-choice task from 60 volunteers. Without registration, that is a lead, not a finding.

Rule: decide what result would move you, and how far, before you read the study.