Future dreams of electric sheep: Case study of a possibly precognitive lucid dreamer with AI scoring

Abstract

Precognitive dreaming is dreaming about seemingly unpredictable future events. It has been most convincingly replicated in two case studies using a single skilled precognitive dreamer (Maimonides studies by Krippner et al., 1971, 1972). Instead of repeating these original studies with another skilled precognitive dreamer, here we set out to determine whether an individual with another unusual dreaming skill – that of entering a lucid dream state almost at will and sketching images seen in that state upon awakening – could become a precognitive dreamer with practice. After two promising 5-trial practice pilot studies, we pre-registered a formal experiment with two sets of 5 trials. Each trial consisted of five steps: 1) the dreamer had a lucid dream and recorded the contents of that dream in a transcript, 2) the dreamer emailed the transcript to a skeptical target-selector, 3) the target-selector used a random number generator to select a target from among 478 pre-pooled online targets, 4) the target-selector sent the URL for the target and the dream transcript to both the dreamer and an analyst, 5) the analyst stored the date, dream transcript, and target together in a database. After both sets of 5 trials were complete, each set of transcripts was judged for similarity against the set of 5 targets. We used three methods of judging – 1) a pre-registered but flawed judging method using two skilled human judges, 2) an exploratory method drawing on unskilled human judges, and 3) an exploratory method comparing judging performance across five different text embedding models within large language AI models. The skilled human judges matched three dreams to their correct targets, the unskilled human judges only matched one dream to its correct target. The three top-performing embedding models were able to match 5 of the 10 dream transcripts to their correct targets. Thus, human-judging methods, if taken at face value, offered slim evidence for precognition in this lucid dreamer. However, AI-judged methods offered clear evidence for precognition in the same dreamer, but a confirmatory experiment is required before drawing firm conclusions. Further, several of the accurate transcript/target pair matches made by the top-performing text embedding models matched those of the skilled human judges, suggesting that the AI method captured human sensibilities and expanded on them. The differences in accuracy among the embedding models have implications for the selection of AI models for future free-response experiments and can begin to give shape to a future of AI participation in screening, training, performance, and analysis in multiple free-response contexts.