Free-response ESP tests ask participants to describe targets such as images, videos or locations, allowing richer impressions than forced-choice tests. This article outlines common judging and statistical methods, identifies desirable and problematic designs, and explains why randomisation, sensory-cue controls and preregistered confirmatory procedures matter.
- Free-response methods use open-ended impressions about hidden targets, most often in ganzfeld and remote-viewing studies.
- Judging may rely on direct hits, ranks or ratings, each with different statistical and practical implications.
- Designs with constrained target selection can introduce sensory-cue risks and require appropriate analysis methods.
In free-response ESP research, participants attempt to obtain impressions about a completely unknown target that is typically a picture, video, or location. This allows a much greater range of ESP impressions than forced-choice experiments in which the targets are limited to a few choices known to the participant. Free-response targets can be less artificial and more engaging. Ganzfeld and remote viewing are the two most common types of free-response ESP experiments.
Although commonly used research methods for published free-response ESP experiments are well developed and not controversial, particularly for ganzfeld research, some variations are problematic. Most of these variations were recognized as undesirable in the 1970s1Burdick & Kelly (1977); Kennedy (1979a, 1979b)., but these older technical writings are largely unknown to later researchers and have not been easily accessible. Thus, past problematic practices sometimes reappear today.
The primary disadvantage of free-response experiments is the amount of time and effort that is needed for each trial. One free-response trial may take thirty minutes to an hour to complete, while a forced-choice trial may be completed in a few seconds. In the past, studies collecting several hundred free-response trials were not practical. However, online research is overcoming that limitation.
Most recent free-response experiments have been for clairvoyance, in which the target is selected before the participant makes a response, or precognition, in which the target is selected after the participant makes a response. Early free-response experiments were often for telepathy, in which a person who knew the target attempted to send it mentally to the participant who made the response. Few telepathy experiments have been conducted in recent years because of the additional effort needed to manage two people during the study and the possibility that the sender and receiver could use electronics to communicate and create false results. Such subterfuge is particularly of concern when the participants are college students or persons claiming to have special psychic abilities.
A free-response trial is evaluated by comparing the participant’s response with several possible targets (typically four), one of which is the actual target. These comparisons are needed for statistical evaluation to provide evidence that correspondences between the response and target were not simply random coincidences due to chance. The participant’s response may include writing, diagrams, and/or spoken comments that are audio recorded.
The comparison or judging can be done by the participant or by someone else. Judging by the participant has the advantage that the participant presumably has the best knowledge and understanding of the impressions. In addition, less research effort is needed than finding appropriate independent judges and managing the added judging sessions. However, some researchers believe that judging by the participant may make the participant less focused on the actual target and reduce ESP accuracy. If that is true, the optimal practice is for the participant to see only the actual target and the judging to be done by someone else. The judges must be carefully isolated from any discussions or informal comments about any experimental sessions or results. In all cases, a researcher who knows the selected target for a trial should not be present during the judging. Artificial intelligence (AI) will likely have an increasing role in judging free-response trials.2As in Mossbridge et al. (2025).
The statistical comparison of the response with the possible targets can be based on direct hits, ranks or ratings. Direct hits are based on the one target that has the greatest correspondence with the response. For ranks the targets are ordered (eg, 1,2, 3, and 4) based on the degree of correspondence with the response. In theory, ranks should provide for better analysis because it incorporates near misses. Established statistical analysis methods are available for direct hits (binomial analysis) and ranks (preferential ranking analysis). For ratings, the judge rates the degree of correspondence between the response and each target on a scale such as 1 to 30 or 1 to 100. In theory, this would be better than ranks if psi occurs; however, such ratings do not fit an established statistical model. The ratings can be evaluated with computer simulations as described in the article about stacking effects. The judge is usually instructed not to give the same rating to two different possible targets. Thus, ratings can be reduced to ranks, and ranks can be reduced to direct hits.
The theoretical rationale that ranks are more sensitive than direct hits and ratings are more sensitive than ranks is based on the assumption that psi conforms to the standard reductionistic models of information processing in science and technology. That is a debatable assumption for psi.3Kennedy (2013). Direct hits, ranks, and ratings require increasing analysis effort. The practical value of the increased effort in the context of psi remains uncertain.
Contents
Desirable Experimental Designs
The most common practice is to create a target pool consisting of a number of target sets. Each target set has (typically) four possible targets that were selected to be distinct from one another. For each trial, a target set is selected and the target is randomly selected from it. The target set does not need to be randomly selected, but the target within the set should be randomly selected. After the participant makes a response, the participant is shown the four possible targets in the set in random order and asked to compare the impressions with each of the targets and indicate the outcome (direct hit, ranks, and/or ratings). This is the simplest procedure and also the least susceptible to bias.
If a participant does multiple trials, a separate target set may be selected for each trial, which would prevent the participant from having the same target twice. If the same target set is used for two or more trials, the random target selection should be independent and unconstrained for each trial, which means that the same target could be selected for different trials (in statistical terms, random selection with replacement). The reasons for this are discussed below.
The same basic procedure is used if the judging is done by one or more other persons. However, if a participant does more than one trial, the responses should be checked to verify that they do not mention or give hints about the targets on previous trials. Alternatively, the trials can be judged in the order that they occurred and the judging for a trial cannot be changed after the judge has started the next trial. As noted above, the judges must also be carefully shielded from any comments or information about the outcomes of the experimental sessions. These possibilities for compromising information are eliminated if the participant does the judging at the time a response is made, which has been the most common practice for ganzfeld research.
Primarily of historical interest: If paper targets are used in a telepathy experiment, duplicate copies of the targets should be used for the person sending and for judging to eliminate any cues from the handling of the targets by the sender.
Problematic Experimental Designs
Significant problems can occur when a target set is used on multiple trials and the target selection is constrained to not select the same target twice (random selection without replacement). This constraint may be applied if researchers do not want a participant to get the same target twice or want to balance the use of individual targets. This can also occur if predesignated target sets are not used and targets are randomly selected with four sequential trials being judged in a group as a target set. In these cases, the judging is done by independent judges.
The most intractable problem with this design is the possibility of sensory cues. When a participant does multiple trials with one target set and receives feedback about the actual target on each trial, the response on a trial may contain cues about the targets on previous trials. The target on a previous trial is in the target set for judging, but cannot have been the selected target on a trial if it was the target on a previous trial (unlike with unconstrained randomization above). The probability of correct judging is increased due to the cues rather than due to ESP.
The cues could be blatant, such as mentioning a previous target; or the bias could be more subtle, like avoiding descriptions that are consistent with the targets on previous trials. For example, a participant in a study with this design commented that he worried that some responses were too much like previous targets and would be difficult to judge.4Mossbridge et al. (2025). In the experiment, the participant was shown the selected target on each trial and the judging was done by someone else later. Intentional or unintentional efforts by a participant to help the judging process may introduce bias based on the participant’s knowledge about the targets on previous trials.
The possibility of bias from knowledge about previous targets can be eliminated if the participants are not shown the selected target (no feedback) or if each participant does only one trial with a specific target set and has no information about the outcome of other trials with the target set. The researcher present when the participant makes a response should also have no information about the targets on previous trials.
Proponents of psi may argue that subtle sensory cues are implausible and unproven with this design. However, others may consider that psi is implausible, unproven, and less likely than sensory cues. Sceptics of psi could argue that significant results in these experiments are evidence that such subtle biases occur given that psi is not a reasonable possibility. The more desirable designs described above do not have sensory cues as a possible confounding factor.
Incorrect statistical analysis also tends to occur with this design. The binomial and preferential ranking analyses that are used with the desirable designs above with unconstrained randomization are not appropriate for designs with constrained randomization. The average expected number of hits by chance is the same for both designs but the random variability due to chance can be greater with constrained randomization because the targets are not independent.
Applying the usual binomial and preferential ranking analyses that assume independent targets tends to produce results that are incorrectly more significant (smaller p-values) than with an appropriate analysis.
Using simulations to evaluate the probability of a study outcome as described in the article about stacking effects is appropriate here—and for virtually any other psi experiment that involves randomization. This analysis does not solve the problem of sensory cues, but may be useful when participants do not receive feedback or do only one trial with a target set.
General Methodological Practices
Good research should incorporate lessons from the replication crisis, including distinguishing between exploratory and confirmatory research; preregistering planned studies; software validation, and other quality control measures that prevent and detect unintentional errors and intentional errors (researcher fraud); making the data publicly available; and conducting formal confirmatory research with adequate sample sizes and without exploratory flexibility. The uncertainties and debates about ganzfeld research have focused on these practices more than on the standard design and analysis methods described above.
In evaluating free-response research, studies that were preregistered as confirmatory, used appropriate statistical analysis, and did not have controversial designs with the potential for sensory cues should be distinguished from other studies and provide the primary evidence about an effect.
James E Kennedy
Acknowledgment: Professor Emeritus Caroline Watt kindly reviewed an earlier version of this article and made valuable comments that improved it significantly.
Works Cited
Burdick, D.S., & Kelly, E.F. (1977). Statistical methods in parapsychological research. In Handbook of Parapsychology, ed. by B.B. Wolman, 81-130. New York: Van Nostrand Reinhold.
Kennedy, J.E. (1979a). Methodological problems in free-response ESP experiments. [Full text.] Journal of the American Society for Psychical Research 73, 1-15.
Kennedy, J.E. (1979b). More on methodological problems in free-response psi experiments. [Full text,[ Journal of the American Society for Psychical Research 73, 395-401.
Kennedy, J.E. (2013). Can parapsychology move beyond the controversies of retrospective meta-analysis. [Full text.] Journal of Parapsychology 77, 21-35.
Mossbridge, J., Green, D., French, C.C., Pickering, A., & Abraham, D. (2025). Future dreams of electric sheep: Case study of a possibly precognitive lucid dreamer with AI scoring. [Download PDF.] International Journal of Dream Research 18, 151-68.