Were Alien Abductions Invented Under Hypnosis?
Does hypnotic regression tell a different abduction story than conscious recall? The false-memory account predicts it should — longer, stranger, more uniform. We tested that against 196 tellings of 158 cases — a telling is one account of one event as it reached the record, so a case told by two witnesses or on two programmes counts twice — the largest historical archive we could assemble. It isn’t there.
Question: Do hypnotically-recovered abduction accounts differ from consciously-recalled ones — richer, more exotic, more templated? Sample and method: 196 labelled abduction tellings — 102 consciously recalled, 94 recovered under hypnotic regression — across 158 cases, ~232 named witnesses, 22 countries and the 1910s–2020s. Re-run 22 Sep 2026 on the current archive; the July figures held.
If hypnosis writes the abduction story, the seams should show
In the 1980s and 90s most famous abduction accounts were recovered under hypnosis, and the standard sceptical reading is that hypnosis creates the narrative — the hypnotist expects the script, the subject obliges, the culture supplies the details. If that’s right, hypnosis-derived accounts should be measurably longer, stranger and more uniform than accounts people simply remembered awake.
Nobody, as far as we know, had measured that against a large historical archive. We did — comparing two routes to the record across 196 tellings, holding the era fixed and letting the route vary. The predictions and the tests were written down before the data was run (that is what “pre-registered” means), so nothing here was chosen after seeing the answer.
For the technically minded: era-stratified permutation tests (10,000 draws), case-clustered bootstrap CIs (2,000 draws), Benjamini–Hochberg FDR across the element screen, seed fixed. In plain English, below.
This does not mean the abductions happened, and it does not prove hypnosis is reliable — memory research is clear that confident false detail can be implanted. It means something narrower: the signature that hypnosis-constructed narratives were predicted to leave is not in the data.
Selection into hypnosis — investigators historically chose whom to regress, and they chose the strange cases — is the one confound — a hidden factor that could produce the result on its own — that can’t be removed. It’s stated wherever it bites (§08).
84 tellings arrive via both routes and are analysed separately. The corpus spans 158 distinct cases, ~232 named witnesses, 22 countries and the 1910s–2020s — independent people who overwhelmingly never met.
Regression peaks in the 1970s, the era of Budd Hopkins and John Mack, the investigators who made regression the standard tool; conscious accounts run later. That is why every test below shuffles labels within an era — a raw comparison would be one decade against another in disguise.
To ask whether regression tellings really differ, we shuffled the “conscious” and “hypnosis” labels across the same tellings 10,000 times and checked how often pure chance produces a gap as big as the real one. Shuffles stay within the same era, so it’s never 1960s-vs-1990s in disguise.
Famous cases get retold — Betty and Barney Hill appear many times. Our uncertainty ranges resample by case, not by telling, so one celebrated story can never masquerade as twenty independent data points.
We compared 37 story ingredients one by one. Test that many things and a few will look “significant” by luck alone — so every per-ingredient result pays a statistical tax before we believe it. Only 5 survive it.
A p-value is how often pure chance alone would produce a gap this big — under 0.05 is conventionally “significant”. A 95% CI (confidence interval) is the range the true effect plausibly sits in; if it crosses the null line — the line where “no difference” sits — chance can’t be ruled out. A ratio of 1.0× means the two routes are equal; 0.8× means regression tellings have 80% as much. pp is percentage points — the plain gap between two percentages. Raw is the simple comparison of the two groups; era-adjusted is the same comparison made decade by decade and then combined, so the hypnosis-heavy 1970s can’t tip it. The outcomes are numbered O1–O7 to match the technical dossier.
The inflation that never came
If hypnosis inflated the account, both effects would sit to the right of the null. Instead each estimate lands left — regression tellings carry marginally fewer story elements (0.833× as many, raw; 0.809× once each era is compared with itself) and are no more exotic (−3.2 pp). That leaner-story gap sits exactly at the conventional threshold (p = 0.050) and runs the opposite way to the prediction. The toggle switches which estimate is drawn — raw or era-adjusted — not the test behind it.
Experience elements per telling
Exotic share of the account
The abduction script itself — being taken aboard, procedures, telepathy, implants, hybrids, what the beings look like.
External sighting detail — lights, craft, witnesses, physical traces. Government/coverup content and missing time are excluded by design — missing time is the reason people were sent to a hypnotist, so counting it would bake the confound into the outcome.
The same story, told through two doors
Every telling is broken into its story ingredients — the 37 common enough to test are grouped here by the seven beats of the abduction script. Each row is one element: the grey dot is how often conscious accounts include it, the blue dot regression accounts — the bar between them is the gap. Read top to bottom and the two routes trace the same profile. 5 elements survive the false-alarm tax (★): 4 lean conscious (bedroom visitation, fear response, telepathy and paralysis) — against the stereotype — and the one that leans regression is medical procedure, the examination beat. Click any highlighted element to see the real cases behind it.
Telepathy
Telepathic contact — far commoner in conscious accounts, and the gap survives the false-alarm tax.
A Michigan housewife’s car was moved for miles while others slept; months later she was drawn barefoot from her home to meet a note-taking being.
Three silvery-suited beings communicated with the paralysed 20-year-old.
A patrolman met a hovering saucer; car, radio and flashlight failed. Under hypnosis he recalled being taken aboard.
A truck driver changing a tyre was illuminated and paralysed; three blonde beings took a blood sample.
Regression walks fewer beats, not more
Folklorist Thomas Bullard showed abduction stories follow a canonical seven-beat sequence. If regression pushed witnesses further into the script, its bars would run longer. They don’t — the conscious arm covers more beats (3.72 vs 3.07 of 7), with the biggest deficits at capture and communication.
−0.64 (CI −1.35 to +0.13, p=0.054)
Grey = conscious, blue = regression, as a share of tellings that touch each beat; the right-hand figure is the gap in percentage points, and ✓ marks a gap chance would rarely produce (p under 0.05). 3 beats reach significance — capture and communication favour the conscious arm; procedures is the one that favours regression.
A cross-route pair shares 87% of what a same-route pair does
If regression wrote its own narrative, two regression tellings should overlap far more than a regression–conscious pair. They barely differ. Overlap here simply means the fraction of story ingredients two tellings share (the “Jaccard” score — 0 is nothing in common, 1 is identical).
of a same-route pair’s overlap is shared by a cross-route pair. One story, thinly accented — same-route minus cross-route overlap = +0.0195, p<0.001.
The one hint in the sceptics’ favour — regression summaries read slightly more alike as prose. A language model scores every pair of summaries from 0 to 1 by similarity of meaning; the figure is the average for each route, and higher means more uniform.
Raw Δ+0.049 (p=0.001) clears the stricter p<0.01 bar we set for this outcome in advance. On 30-word-truncated summaries — the same test with length taken out — the gap shrinks to Δ+0.017 (p=0.182) — so much of it was length, not route. Treat convergence as a hint, not a finding.
Three natural experiments
Three cases appear in the corpus via both routes — the closest this data comes to a controlled comparison, because the underlying event is held fixed while the memory route varies. What matters isn’t the raw overlap (diluted by brief tellings) but what kind of element each route adds.
Herbert Schirmer abduction
Pascagoula abduction
Travis Walton abduction
The null holds in every slice
The same two questions asked 8 different ways: one vote per case instead of per telling; only cases our archive grades L3+ (corroborated by three or more independent sources) and only L0 (a single source); dated cases only; the 84 both-route tellings counted as regression, then as conscious; and a wider pool that adds contact and close-encounter reports. Every point hugs the null line — where “no difference” sits; marker size scales with sample size. Below 1.0× = regression tells a leaner story. Every cell sits at or near the null — the largest, 1.029×, comes from counting both-route tellings as regression.
The primary row re-runs the §02 test inside the robustness grid with 10,000 fresh shuffles; it reproduces §02's p (0.050).
What this study cannot rule out
Precision = correct labels ÷ labels that could be judged; 19% were genuinely ambiguous. Self-hypnosis and conscious-fragment-then-regression cases blur the arms — which pushes them toward each other. Mislabelling of that kind can only hide a real difference, never invent one — so if anything the “no difference” result is understated, not inflated. A full transcript audit remains a follow-up.
Selection into hypnosis
THE BIG ONEInvestigators of the Hopkins and Mack era chose whom to regress — and chose the strange cases. That selection should have pushed the regression group toward MORE strangeness. It shows less or equal, so this hidden factor works against our result rather than for it.
Archival visibility
CONTROLLED-FORFamous, media-covered cases are over-represented in any archive. The one-case-one-vote rule (§01) stops any single famous case from counting as many independent points, and the L3+ and L0 rows in §07 show the result isn’t carried by either the well-corroborated or the single-source end.
Extraction
MEASUREDStory elements are pulled out of narrated recordings by a language model, one remove from the witness. The label audit measures this error (16 tellings hand-graded: 86% conscious precision, 100% regression precision, 19% genuinely ambiguous); it does not erase it.
Hypnotically-recovered tellings are not richer (0.833× as many story elements; 0.809× era-adjusted, p=0.050 — the lean runs the other way), not more exotic (−3.2 percentage points), and touch fewer beats of the seven-beat script (3.07 vs 3.72, p=0.054). The two routes are overwhelmingly telling one story — a cross-route pair shares 87% of a same-route pair’s ingredients.
Where individual elements genuinely differ, the direction runs against the stereotype: bedroom visitation, fear response, telepathy and paralysis all skew conscious; regression’s one lean is procedural — the medical-examination beat, exactly where a session probing “what happened inside?” would dwell. The corpus can’t certify exact parity, but the inflation the false-memory account predicts is absent everywhere the sample was large enough to show it — and the story was already fully formed in conscious-recall accounts decades before regression became standard practice.
RE-RUN 2026-09-22 ON THE CURRENT ARCHIVE · EVERY FIGURE ON THIS PAGE IS READ FROM THAT RUN’S EXPORT · DOSSIER: dossiers/recall-method.md
