Paranormal claims are a useful place to learn how any surprising claim gets tested, because every ordinary source of error shows up at once. Carry four questions through every chapter: what exactly is claimed, what ordinary route could produce the result, how was that route closed, and did anyone independent get it again? The chapters move from claims and witnesses, through coincidence, readings and deception, to controls, statistics and the research programs where those tools were applied. You learn how a test works and what it reported, not a doctrine about what is possible. Reported experiences are treated as real; the question is always the explanation offered for them. Orientation for study, not medical, psychological, legal or financial advice.
Each chapter opens with the short version. Tap one to read the detail.
What a claim must say before it can be tested
~2 min
A claim becomes testable when it names, in advance, a result that would count against it. The person asserting it supplies the evidence, and "unexplained" points to no particular cause.
Split every report in two: the experience and the explanation. Someone who sees a figure in a doorway is describing something real to them, and your test aims only at the cause offered for it, never at that person's honesty.
A claim becomes testable once it says, ahead of time, which result would count against it. That demand, from Popper, is why a forecast on record can earn credit while a reading fitted to events afterward cannot. In a four-choice picture test, "sensing a hidden image" means picking the real target from among three decoys, so pure guessing scores 25 percent. Being testable does not make a claim scientific or true: philosophers note that many claims called pseudoscience were tested and failed.
The party asserting the claim supplies the support. "Nobody has disproved it" adds nothing, and "unexplained" only describes current knowledge. So clear chance, misperception, memory error, instrument fault and trickery first. The evidence you need scales with what a result would overturn, the weighing behind Hume's maxim on miracles. A search that finds nothing counts only if it could have detected the effect.
The trap is the escape clause. Say a claimant scores at chance, then explains that a skeptic in the room blocked the ability. That rescue restores the fit, but a claim that survives every outcome predicts nothing. The opposite slip turns "no controlled study has replicated this" into "this is impossible," which goes beyond the evidence.
Rule: before you look at a result, ask which outcome would count against the claim and who owes the evidence for it.
The witness as an instrument
~2 min
An honest witness reports what they noticed, which can differ from what happened. Attention, viewing conditions, expectation and question wording all shape the report.
A vivid story has no denominator. It reached you because it was striking, so a stack of remarkable cases was selected for being remarkable.
Attention decides what registers: many viewers of one magic routine missed a coin crossing openly between napkins in about half a second. Pattern detection fires on noise, so a face in a stain, called pareidolia, is a real experience and evidence about vision. Expectation pushes the same way: told that most noisy displays held a human figure, people said they saw one more often.
Picture a hiker who reports a huge light darting across the night sky. A fixed dim point in total darkness can seem to wander, the autokinetic effect, and a light overhead offers few distance cues, which can leave size, distance and speed uncertain together. The hiker can be reporting faithfully while the motion still needs independent support.
"What color was the craft's beam?" assumes a craft and a beam, while "Tell me what you saw" leaves the content to the witness.
The trap is treating honesty or expertise as a measure of accuracy. Tests of staged slate-writing, written up by 1900, found sincere sitters describing marvels that never occurred. Nobody had to lie for the record to be wrong.
Rule: ask what the witness could have seen from there, what they were expecting, and how the question was worded, then weigh the report by those answers.
Memory rebuilds, and suggestion can supply parts
~2 min
Recall reassembles an event and fills the gaps, so later information can enter a memory. Hypnosis can raise how much people report without raising accuracy.
Remembering works less like playback than like rebuilding. Details you meet afterward, from a conversation or a news story, can join your memory of the event, which is the misinformation effect. You can also keep the content and lose its origin, filing something imagined as something seen. Studies have produced full recollections of events that never occurred, such as getting lost in a shopping mall, in people who were not lying. Observers shown recordings of true and implanted memories could not reliably sort them.
Researchers still argue over whether hypnosis is a special state or ordinary social behavior. The memory finding is firmer: people can report more because their standard for "remembered" has loosened, and the extra material is no more accurate.
Say someone woke terrified years ago and, under hypnosis, now describes a room and faces in rich detail. That detail could come from memory, imagination or earlier reading, and hypnosis cannot sort them. Past-life regression has the same problem, with cryptomnesia, forgotten material resurfacing, as one ordinary source.
The trap is reading vividness as proof of origin. Flashbulb memories of shocking news drift in their details while confidence stays high. Whether trauma can be repressed and recovered is itself disputed, and a dispute about method is no verdict on any one person's memory.
Rule: find the earliest record made before retelling or suggestion, and weigh every later detail against it.
Coincidence, base rates and remembered hits
~2 min
With enough chances, striking matches are expected somewhere, and matches on common details prove little. People tend to keep the hits and lose the misses, which is ordinary human reasoning.
An event with million-to-one odds, given a hundred million chances, should turn up about a hundred times. This law of truly large numbers changes your question from "how unlikely?" to "how many chances were there?" The birthday problem shows it in miniature: any pair can match, and 23 people already form 253 pairs, which gives slightly better than even odds of a shared birthday.
A reader's line about a relative with heart trouble sounds pointed, yet heart disease is the leading cause of death in the US, so it often lands by chance. Ignoring how common a detail is goes by the name base-rate neglect. Randomness also looks streakier than people expect, so a cluster needs no special cause of its own.
Say you dream about an old friend and she phones the next day. The fair calculation covers every night, every person you might dream about and every call that might follow, plus the dreams followed by nothing, which nobody retells. Francis Bacon complained long ago that hits on a prophecy get marked and misses do not.
Confirmation bias, favoring what supports your belief, adds to the filtering. In one study, paranormal believers showed stronger illusions of a link between unrelated events only where no real relationship existed, and they detected real relationships as well as anyone else.
The trap is computing the odds of this exact event after you have already seen it, when the fair figure is the chance of some striking match of that general kind.
Rule: before asking how unlikely a match is, count its chances, check how common the detail is, and go looking for the misses.
Readings: why they feel right and how to score them
~2 min
General statements feel personal, and a reader can learn more from your reactions. Astrology and other readings are checkable the same way: hide the real target among decoys and compare with chance.
A Barnum statement sounds aimed at you and fits almost anyone: "You keep some doubts behind a capable front." In a classic classroom demonstration, each student scores a seemingly personal profile, then learns that the whole class got the same text. One modern repeat still drew ratings near 80 out of 100.
Such lines belong with cold reading, a set of techniques that make a stranger feel understood. Many general guesses make a match likely, and matches get noticed. Fishing draws the detail out of the sitter's own answers. A two-sided line pairing a trait with its opposite is hard to get wrong. Facts gathered beforehand and presented as intuition are a different route, since nothing is being read from the sitter at all.
The control removes contact and adds a decoy. Suppose a reader works for two absent sitters through a stand-in, and each sitter then scores two anonymized readings, unaware of which was meant for them. Real insight should push the intended reading's score higher. One such study reported success for some mediums, and the wider field has no consensus.
Astrology makes checkable claims, so it faces the same design: can a chart be matched to its owner's profile above chance, with nobody knowing the pairing? Philosophers cite a double-blind test in Nature among the studies behind their statement that astrology has been tested and refuted.
The trap is judging a reading by how it feels, or rejecting it because no mechanism is known. A sense of fit measures your reaction, and a missing mechanism is a separate question from whether predictions succeed.
Rule: score a reading against a decoy the sitter cannot tell apart, and let the hit rate decide.
Deception calls for its own expertise
~2 min
Scientific training prepares you to read nature, which never tries to fool you. A protocol that may face deliberate cheating needs conjuring knowledge and control of every interval.
Misdirection steers your attention so the key move passes unnoticed, even in full view. Audiences then remember the effect and invent a method for it.
Education is weak protection, because nature does not set out to deceive you. A classroom study led educated adults, including scientists, to credit ordinary tricks to psychic power, and introducing the performer as a psychic instead of a conjuror drew more of those explanations. So a team testing anyone who might cheat adds an experienced conjuror.
Take billet reading. Sitters write names on slips and fold them, and the performer seems to read them unopened. A 1920 account has one medium swapping the slips for blanks, then reading the originals behind his hands while relighting a cigar. The fix is plain once you know the method: the slips stay with the people who wrote them.
In 1888 a Fox sister admitted fraud in print and then made the rapping sounds on stage, which carries more weight than a bare statement, since the method could be seen.
The trap is assuming that a careful observer would have noticed. Stage mediums beat committees that bound them, and witnesses to one conjurer's writing trick insisted they never lost sight of his thumb. Calling something a trick entertains, while calling the same trick real makes a factual claim you can test.
Rule: assume a skilled performer is present, ask what they could do at each step, and leave no unwatched moment.
Controls, blinding and closed channels
~2 min
A comparison group, random assignment and blinding each block a different error. Information can also leak by channels nobody meant to open, so a protocol must name who knows what and when.
A comparison group shows what would have happened anyway, since people improve through placebo, natural change or practice. Random assignment keeps the groups from differing in ways you never measured. A blinded participant cannot let expectations color the report, and a blinded researcher cannot let the condition color handling and scoring.
The classic case is Clever Hans, the counting horse. A commission ruled out deliberate tricks. Pfungst then found the real signal: a small lean forward set the tapping going and a tiny upward head jerk ended it. Large blinders stopped the success, which showed the cues were visual. The questioners were honest, and thousands of spectators had missed it. Leakage is any ordinary route by which the answer reaches the responder, with or without intent.
So your protocol specifies who chooses the target, by what method and at what moment. In ganzfeld work, the picture-guessing design covered later, an automatic random procedure usually picks the target, and the assistant handling the participant must stay ignorant of it until the ratings are finished.
The trap comes in two forms. One is calling a setup secure because nobody meant to cue anyone. The other is the objection invented after an unwelcome result, as when Hans's owner blamed failed blind trials on the horse's moods.
Rule: list every ordinary route by which the answer could reach the responder, close each one before the first trial, and write down who knows what at each step.
What a significant result can and cannot say
~2 min
A p-value says how unusual the data are if chance alone is operating, not whether a hypothesis is true. Flexible analysis and unpublished nulls inflate a literature, so plans get fixed in advance.
A p-value starts by assuming chance alone is at work, then reports how surprising the data would be if so. It cannot give the odds that chance produced the result, and it cannot say which assumption failed. With enough trials a tiny deviation clears the bar, so check the effect's size as well: three points above chance is three points.
Flexibility can turn noise into findings without anyone cheating. Suppose a lab checks 20 independent outcomes at the usual 5 percent threshold. Even with no real effect, the chance of at least one "significant" result is about 64 percent. Peeking and stopping once results look good raises false positives too, and presenting a surprise as a prediction, called HARKing (hypothesizing after the results are known), hides what happened.
Nulls left in the drawer make the published record overstate an effect, and pooling studies adds precision while inheriting every bias in the pool.
Preregistration files predictions and analysis before data exist. A parapsychology registry shows why it matters: the studies registered in advance came up significant far less often than unregistered ones from the same years. A registered report goes further, with the journal reviewing the plan and promising publication whatever the outcome.
The trap is reading "significant" as "chance is ruled out and the effect is real." The calculation began by assuming chance, and a broken assumption can produce the same small number.
Rule: ask what was fixed in advance, what went unreported, how large the effect is, and whether anyone independent got it again.
Four research programs and where each dispute sits
~2 min
Ganzfeld, remote viewing, micro-psychokinesis and precognition each drew positive reports and sustained criticism. Read each by its chance baseline, inclusion rules and independent replication.
In a ganzfeld session, a receiver rests in an even, unpatterned sensory field and describes passing images while a sender elsewhere views a randomly chosen picture or clip. The receiver then picks the target from four candidates, so chance is 25 percent, and pooled databases report about 31. A 2024 meta-analysis, written by people behind earlier positive reviews, found a small effect. An earlier simulation by other researchers concluded that questionable practices had likely inflated significance, while leaving an effect unexplained. There is no consensus.
Remote viewing asks someone to describe a distant target, with judges rating the match, and critics objected that those judges were tied to the labs. After declassification of the US programs in the mid-1990s, two outside reviewers disagreed. Utts judged remote viewing empirically validated, and Hyman did not consider it scientifically established. Both accepted consistent effect sizes in the later studies, so the dispute centered on what to conclude.
Micro-psychokinesis research asks whether intention can tilt a physical random generator from 50 percent. One 2006 meta-analysis of 380 studies reported a very small and uneven effect that weakened as studies grew, and researchers disagree about what that means. A multi-site attempt to repeat one benchmark study, adding over 2 million trials, failed to confirm it.
Precognition tests record your response first and let a computer choose the target afterward. Daryl Bem's 2011 experiments reported time-reversed influence. Three preregistered replications of one experiment found no significant effect, and a later multi-laboratory replication of another landed near chance. Bem and colleagues pooled 90 experiments in defense of the claim, which remains disputed.
The trap is taking one headline from either side. A pooled result above chance does not establish a mechanism, and a failed replication does not establish impossibility.
Rule: for any program, find the chance baseline, the inclusion rules and the independent repeats, and report a live disagreement as a disagreement.
Movement you do not feel yourself making
~2 min
Expectation can drive small muscle movements with no felt intention, so a rod or pointer seems pushed from outside. Blind tests and message-passing tests show whose knowledge is producing the result.
Expecting an outcome can set off small movements with no sense of willing them, so a rod or pointer feels pushed from elsewhere. William Carpenter applied this ideomotor account to dowsing rods and pendulums in 1852, and it now covers Ouija boards as well. In 1853 Faraday put rollers between sitters' hands and a table to expose sideways pushing, and a table that had turned then stayed put.
Dowsing persuades in the open because most land has ground water beneath it, so in many areas a well finds some, and dry holes get excused. A US Geological Survey leaflet concludes from controlled work that more testing would not repay its cost.
Facilitated communication asks the same question with higher stakes. A helper gives physical support at the hand or arm while someone spells out words. In a message-passing test, you show an object to the communicator alone, then check whose knowledge appears in the typing. Users have largely failed to convey what the facilitator lacked, and more than 19 rigorous experiments over about 25 years traced the messages to the facilitator. Proponents point to naturalistic studies and want new research. Critics count renamed forms such as the rapid prompting method as one technique, so the same test applies.
The trap is reading a helper's sincerity as validity. These movements require no intent to deceive, so an honest person can produce them. And the test concerns whose words these are. It says nothing about the communicator's intelligence or worth.
Rule: hide the answer from whoever is in contact with the device or the hand, then see whether the result survives.
Haunted rooms and night visitors
~2 min
Field readings and recorded voices need controls before they mean anything, and natural explanations need evidence too. Sleep paralysis shows how an ordinary mechanism can produce a frightening visitor.
Suppose your team's meter jumps near a staircase and a recording seems to say a name. An electromagnetic field (EMF) meter reads a local field that fades fast with distance, so shifting it two or three feet can change the number a lot, and no reading names a cause. Electronic voice phenomena (EVP) are apparent voices on ordinary recordings, and priming matters: over a third of people in one study heard voices in random noise after videos about EVP. So give several listeners the clip with no hint and compare transcripts, then repeat your readings in a nearby control house.
Environmental accounts need the same scrutiny. A 1998 proposal said sound near 19 hertz could shake the eyeball and create ghostly sights, but later measurements at a tested venue did not match. A 1921 report traced a couple's phantom footsteps to carbon monoxide from a bad furnace, which puts a safety check first, though one review found it documented in just a couple of cases.
Sleep paralysis offers a clearer mechanism. The muscle atonia of REM (rapid eye movement) sleep persists into waking, leaving you aware and unable to move. Hallucinations can include a felt intruder, chest pressure and floating, and estimates of ever having an episode range from about 20 to 40 percent. Local belief supplies the story, from a demon to an alien abduction.
The trap is letting a natural account win by default, or letting an overlap settle every case. Abduction reports share features with these episodes, and each account still needs its own evidence.
Rule: compare with a control location, keep listeners blind, and hold your favorite ordinary explanation to the same standard of evidence.
Health claims: why people improve anyway
~2 min
Symptoms peak, fluctuate and fade, and expectation can help, so improvement after a remedy says little alone. Testimonials cannot rule those causes out.
Say someone tries a remedy on the worst day of a flare, and within a week they feel better. Three ordinary things could produce that. Regression to the mean is the drift of an extreme measurement back toward average, and bigger extremes drift further. Many conditions vary or resolve on their own, so timing hands credit to whatever came last. A placebo effect is a benefit produced by expecting help, and it can follow any treatment.
Only a comparison separates these. People given a placebo or sham, measured the same way, show how much change would have come anyway. Testimonials cannot do that job, because even genuine stories leave those causes open, and advertisers can pick their best results. Federal Trade Commission (FTC) guidance wants competent and reliable scientific evidence behind health claims, generally randomized, controlled human clinical testing for a health benefit. Surveys and anecdotes never qualify, and that includes a practitioner's observations of patients.
Check what was measured, too. A surrogate endpoint stands in for the outcome that matters, as blood pressure does for stroke, and the Food and Drug Administration counts one as validated only after extensive evidence. FTC guidance also asks for outcomes set in advance, all of them reported.
The trap is treating "it can't hurt to try" as a free test. The FTC warns that an unsupported treatment claim might lead someone to skip a more effective option, so the cost includes whatever the remedy displaced. None of this is advice about any treatment; it is how to read the evidence offered for one.
Rule: before crediting a remedy, ask what the same people would have done untreated, and who was measured against whom.
Running a fair test and reporting what it showed
~2 min
Settle the protocol and the success criterion in writing before the first trial, then say exactly what the result covers. A failed test speaks about one claim on one occasion.
When believers and skeptics agree on conditions, controls and which outcomes count for whom, neither side can easily walk away from the result. A pilot run settles procedure and does not count. Suppose a claimant accepts 20 trials with four options each. Chance predicts 5 hits, and you might agree that 11 or more counts as success, since guessing alone gets there about 0.4 percent of the time.
Then report only what your test covers. A failed run concerns one claimant, one protocol and one occasion, so it counts against a claim without refuting it everywhere. An unclaimed cash prize shows only that no one passed.
Reports and footprints exist for yeti and bigfoot, yet no body or recent fossil has been authenticated, and science withholds a verdict on such an animal until evidence is examined. A 2014 survey sequenced 30 hair samples and found known mammals, though its polar bear match for two Himalayan hairs was later questioned. A 2017 study placed its yeti samples with local bears, apart from one dog tooth. Those results describe those samples.
The trap is letting a skeptical conclusion become contempt. Skepticism accepts only claims that survive rigorous tests and keeps the question open, while cynicism closes it in advance. Some paid claims reach grieving or ill people, which calls for care in how you speak about them.
Rule: write down what would count before you start, and afterward claim only what that test on that occasion can support.
Written by Keentune. We are not affiliated with or endorsed by the organizations whose documentation informs this guide, and any linked sources belong to their respective owners.
All exam, test, and product names and trademarks are the property of their respective owners and are used here for identification and reference only. Keentune is independent study practice — not affiliated with, authorized, or endorsed by any of these organizations.