Overview

The idea that dreams might carry information from another mind is ancient. What makes the modern version scientific is testing it under controlled conditions. In a dream-telepathy experiment, a sender sits in an isolated room and concentrates on a randomly chosen target image all night. Meanwhile, a receiver sleeps in a separate room. The receiver’s brain activity is monitored so experimenters know exactly when dreaming is happening. At the right moment, the receiver is woken and asked to describe the dream before it fades. Later, independent judges who don’t know which image was the target compare the dream report against a set of candidate images. They try to pick the correct one. If judges pick the actual target more often than chance would predict, that counts as evidence for dream telepathy.³

This design tries to rule out the most obvious alternative: that the receiver simply guessed well, or that normal sensory cues leaked between rooms. Whether it succeeds well enough is the central dispute in the literature.

Why This Is Hard to Study

Dream telepathy sits at the intersection of several practical and research difficulties. Each one alone would be manageable. Together, they make clean evidence hard to get and easy to dispute.

The most basic problem is infrastructure. Monitoring sleep properly requires an EEG setup, trained staff available through the night, and a facility where sender and receiver can be kept genuinely isolated for eight hours. That is expensive and labour-intensive. Most research groups cannot sustain it for long. This is why the Maimonides program remains the largest single body of controlled work. It is also why post-Maimonides studies mostly shifted to home-based protocols without EEG monitoring.¹

A second difficulty is effect size. In science, an effect size is a number that captures how large or consistent a result is, separate from sample size. The dream-telepathy literature shows a small effect. Small effects are real in many fields, but they require large samples to detect reliably. Most dream-telepathy studies used small numbers of participants. That means any single study is underpowered: it might miss a real effect, or it might find a false one by chance. This is why researchers combine studies using a method called meta-analysis, described in the next section.

A third difficulty is state dependence. Dream recall is unreliable. Even with REM monitoring and deliberate awakening, some receivers remember little. Emotional engagement with the target, the relationship between sender and receiver, and the receiver’s general dream-recall ability all appear to influence results.¹ These factors make replication harder. A study that copies the protocol but not the participant characteristics may not copy the effect.

Finally, the judging procedure has its own vulnerabilities. If the target pool is not truly independent (for example, if some images share obvious themes), judges may pick the right image for reasons unrelated to the dream content. And if the person rating the dreams knows anything about the experimental context, subtle bias can inflate hit rates without any deliberate deception.¹

The Maimonides Program

The Maimonides Dream Laboratory was directed by psychiatrist Montague Ullman and psychologist Stanley Krippner. It ran controlled dream-telepathy studies from 1962 to 1978. It remains the most sustained and carefully developed research program in this area. Ullman, Krippner, and Alan Vaughan summarized the work in their 1973 book, which has been reissued in multiple editions.³ The program tested both unselected volunteers and participants pre-screened for strong dream recall or apparent psi ability. Studies using pre-selected participants and emotionally vivid, multi-sensory targets produced the largest effects in the dataset.¹

One notable feature of the Maimonides work was the use of art prints as targets. These were chosen for emotional richness rather than neutrality. The theory was that emotionally engaging material might be more likely to influence dream content. Later researchers tested this and found some support for it: emotional targets appeared more likely to produce positive results than neutral ones.¹

Not all Maimonides studies succeeded. Individual study results ranged from clearly positive to clearly negative. The program also attracted at least one direct non-replication attempt. Belvedere and Foulkes in 1971 failed to reproduce the effect using a similar protocol.¹⁷ This matters: a single lab’s positive results, even across many studies, carry less weight than independent replication.

What the Meta-Analyses Found

Any single dream-telepathy experiment gives a small, ambiguous result. So researchers combine many separate studies using a method called meta-analysis. The idea is straightforward: if a real effect exists, combining many studies should make it detectable even when no single study is convincing on its own. Researchers convert each study’s result into a common unit called an effect size. They then average across studies, giving more weight to larger or better-designed studies. This is called statistical weighting. A study with 200 participants counts for more than a study with 15.

Two meta-analyses have examined the dream-ESP literature. Both report a positive overall effect. The question is how robust that effect is once modern statistical tools are applied.¹² A reanalysis published in 2018 applied stricter weighting. It found the effect was substantially smaller than either meta-analysis reported. The effect disappeared once corrections for publication bias were included.⁶

Publication bias is a specific concern here. It refers to the tendency for positive results to get published while negative results go unreported. If many small studies with positive results are published and many small studies with null results sit in file drawers, a meta-analysis will overestimate the true effect. The reanalysis found that large-sample studies produced near-zero effects, while small-sample studies produced the largest effects. That pattern is a warning sign for publication bias.⁶

judges could correctly identify target materials more often than would be expected by chance using dream mentation¹

The Storm et al. (2017) meta-analysis also found that effect sizes declined significantly over the 51-year period of the literature. Crucially, this decline was not explained by improvements in study quality over time. Better-designed studies did not produce smaller effects than worse-designed ones.² A decline that is unrelated to quality improvements is hard to interpret. It could reflect regression to the mean in an early literature, or it could reflect something about the phenomenon itself.

The dream-telepathy literature does not exist in isolation. Three bodies of mainstream science inform how to read its evidence base: the cognitive psychology of what dreams normally contain, the statistical framework that the judging procedure assumes, and the baseline rates at which people can reliably recall dreams. None of these fields treat dream telepathy as a hypothesis worth testing in their own right, but each constrains the interpretation of dream-ESP results.

The cognitive psychology of dream content, developed across Foulkes and others, finds that dreams systematically reflect waking concerns, recent experiences, and stable individual differences in imagery and memory. Dream content is patterned rather than random, which means any random target image has some non-zero baseline similarity to typical dream content. The Maimonides protocol’s use of emotionally vivid art prints addresses this by selecting targets that diverge sharply from mundane dream themes, but the underlying issue (that dreams are not blank slates) remains relevant when interpreting hit rates.⁸

The blind-judging procedure used in Maimonides and post-Maimonides studies is, in statistical terms, a free-response signal-detection task. Signal-detection theory, formalized in psychophysics by Green and Swets in the 1960s, treats every judging decision as a noisy classification between signal (the actual target) and noise (the decoy pool). The framework predicts that hit rates depend not just on the underlying signal strength but on the size and similarity of the target pool, the judge’s decision criterion, and the granularity of the rating scale. Translating dream-ESP outcomes into SDT terms is rare in the parapsychology literature, but doing so would let researchers compare effect sizes across protocols on a common metric.⁹

Finally, dream-recall reliability sets a hard ceiling on what any dream-ESP study can detect. Dream-recall research outside parapsychology, summarized by Schredl and collaborators, finds that recall frequency varies systematically with age, gender, sleep duration, and attitude toward dreams. Test-retest reliability for self-reported dream recall is high (around r = 0.85), but absolute recall rates depend heavily on the awakening procedure. The Maimonides protocol’s REM-triggered awakening produced recall on 75 to 80 percent of attempts; home-based studies without EEG monitoring rarely match that. Any difference in effect size between lab and home protocols is at least partly explained by this recall-rate difference, independent of any psi signal.¹⁰

Individual studies in the reference library

The studies listed below are the individual dream-telepathy results currently held in the ESP-Nexus reference library, underneath the pooled figures discussed above. What share of the published literature on dream telepathy they represent has not been measured, so the table summarizes what the library holds rather than counting what has been published. The studies also report different kinds of number (direct-hit counts, odds against chance, analysis-of-variance statistics, t and z values, hit rates), and those cannot be added together into one bottom-line figure.

Both directions are present in the table. Within the Maimonides series, the strongest results came from the single-subject studies: the two Erwin series, in which correct transcript-target pairings were rated significantly higher than incorrect ones in the first (p < 0.01)¹² and the Latin-square analysis of the second gave F = 6.43 (p < 0.001),¹² and the Van de Castle series, in which the outside judge scored five direct hits in eight nights.³ Pooled across the two screening studies, the subjects’ own rankings produced nineteen hits and five misses in twenty-four rankings.³ The null results sit beside them: the second screening study, the Posin study and the Grayeb study were all non-significant,³ the Belvedere and Foulkes non-replication discussed above did not reproduce the effect,⁷ and in the one held study built to isolate the sender’s contribution, hit rates were above chance in both conditions but the planned similarity analysis was not significant and sender presence made no difference.⁴

The meta-analyses discussed above also report telepathy-specific figures. In the Storm and colleagues dataset, the twenty-five telepathy studies combined to a mean effect size of 0.22 with a Stouffer Z of 3.74 (p = 9.20 x 10^-5), and no significant differences were found between the telepathy, clairvoyance and precognition modes.² Sherwood and Roe report that the most successful post-Maimonides dream ESP studies were the two telepathy experiments conducted with Sweeney as receiver (r = 0.58 and 0.80), and note that some of the most successful post-Maimonides studies were conducted by particular groups of researchers.¹ Roe and colleagues state their own conclusion directly: their sender comparison ended by “failing to support the proposal that senders play an active role in dream ESP success.”⁴ The Edinburgh consensus-vote dream studies of the 1990s, sometimes discussed alongside this literature, are classified by Sherwood and Roe as clairvoyance studies rather than telepathy studies, and are therefore not listed in the table above.¹

Skeptical Critiques and Discussion

Critique 1: The positive meta-analytic effect is driven by small studies and publication bias, not a genuine psi signal.

Skeptic source: Howard (2018) argues that when studies are weighted by sample size, the overall effect shrinks to near zero. It becomes non-significant after publication bias correction.⁶

Response: Storm et al. (2017) report a highly significant combined effect across 50 studies. Bayesian analysis also excluded zero. They note that quality improvements over time did not account for the decline in effect size. They interpret this as a genuine feature of the data rather than a flaw in how the studies were run.²

Analysis.

Critique 2: The Maimonides results have not been independently replicated under comparably rigorous conditions.

Skeptic source: Belvedere and Foulkes (1971) ran a direct non-replication of a Maimonides study and did not reproduce the effect. Critics including Hyman have raised concerns about judging procedures and target-pool independence in some of the original studies.¹⁷

Response: Sherwood and Roe (2003) acknowledge the non-replication but note that post-Maimonides studies as a group still show a small positive effect (r = 0.14). Procedural differences (no REM monitoring, home-based protocols, neutral targets) likely reduced effect sizes in later work rather than reflecting a true absence of the phenomenon.¹

Analysis.

Critique 3: Judging bias and target-pool problems could inflate hit rates without any psi involved.

Skeptic source: If target images in a pool share thematic features, or if judges have any contextual knowledge of the experiment, ratings can be skewed toward the actual target for non-psi reasons. Hyman and others have raised this concern about specific Maimonides studies.¹

Response: The Maimonides protocol used outside judges who were blind to the target identity. Later studies moved toward consensus judging, which Sherwood and Roe (2003) found to outperform individual judging.¹ Whether all studies in the meta-analytic corpus met adequate blinding standards is not uniformly documented. This remains a legitimate unresolved concern.

Analysis.

Critique 4: The decline in effect size over 51 years suggests the early results were inflated, not that a real effect is fading.

Skeptic source: Howard (2018) notes that the inverse relationship between sample size and effect size is the standard signature of publication bias. The decline over time fits regression to the mean in an early literature dominated by small, enthusiastic research groups.⁶

Response: Storm et al. (2017) tested whether quality improvements explained the decline and found they did not (r = .08, p = .600, meaning essentially no correlation). They interpret the decline as a real feature of the data. Possible explanations include changes in participant selection, target materials, or experimental context over time, rather than simple publication bias.²

Analysis.

Open Questions: What Would Settle This

The dream-telepathy literature is in an unusual position. The positive meta-analytic signal is statistically significant by conventional standards. But a credible reanalysis argues it is an artifact of how studies are weighted and of publication bias. Several specific studies could move the debate forward.

The most direct test would be a large, preregistered, adversarially designed replication of the Maimonides protocol. It would need full EEG-based REM monitoring. It would need a target pool verified for independence by researchers from both pro-psi and skeptical camps. Judging would need to be conducted by blind outside raters with no knowledge of the experimental context. The sample would need to be large enough to detect an effect of r = 0.15 with adequate statistical power. The study would need to be registered before data collection, with the primary hypothesis and analysis plan locked in advance. A null result from such a study would substantially reduce confidence in the phenomenon. A positive result would substantially increase it.

A second valuable study would directly test the target-emotionality hypothesis. Sherwood and Roe (2003) found that emotional targets appeared more likely to produce positive results.¹ A preregistered experiment randomly assigning receivers to emotional versus neutral target conditions, with all other variables held constant, would test whether this pattern is real or just a post-hoc observation in existing data.

A third open question is whether the sender plays any causal role. The meta-analytic data do not clearly distinguish dream telepathy from dream clairvoyance. The receiver might be responding to the target image directly rather than to the sender’s mental state.¹ A design that includes a no-sender clairvoyance condition alongside a sender condition, with targets matched for emotional content, would test this directly. Roe (2007) examined the role of sender and experimenter in dream-ESP research and identified this as an unresolved question in the literature.⁴

Finally, the publication bias question could be partially addressed by a prospective registry for dream-ESP studies. Researchers would log their planned studies before data collection. If null results from registered studies are then published alongside positive ones, future meta-analyses would have a more complete picture of the true effect size distribution.