AI Finds Hidden Patterns in Stevenson’s Case Archive

By C. Rich

I went into this project with a fairly simple thought. Ian Stevenson and the researchers who followed him accumulated thousands of cases involving children who appeared to remember previous lives. They spent decades interviewing families, checking names and places, comparing statements with dead people, recording unusual behaviors, photographing birthmarks and congenital defects, studying phobias, and trying to determine whether ordinary explanations could account for what they were seeing. Whatever anyone ultimately believes these cases represent, it is an extraordinary archive of human observations. What occurred to me was that Stevenson did almost all of this work before modern artificial intelligence existed, and that means we now have a tool for looking at his archive in a way he simply did not have.

I did not want to ask AI whether reincarnation is real. I am a theoretical philosopher, and that is not the question I am interested in answering here. I wanted to ask what I think is a much more useful question: What happens if we stop treating these cases primarily as individual stories and start treating the entire collection as a pattern-recognition problem? Human beings can notice obvious correlations, and Stevenson certainly did, but nobody can comfortably hold dozens of interacting variables across hundreds or thousands of cases in their head at once. AI can help us do exactly that. It can look across death type, age, culture, family relationship, distance, intermission, phobias, birthmarks, behavior, sex changes, skills, twins, and other variables and ask whether some of them repeatedly travel together while others do not.

We made mistakes getting there. In fact, we wasted a great deal of time at first building methodology around the problem instead of actually doing the pattern recognition. Eventually, I threw that work away and started over. We went back to published cases, published aggregates, actual denominators, cultural comparisons, death categories, birthmarks, phobias, sex-change cases, twin cases, and every usable number we could find. Once we stopped designing the perfect way to study the archive and simply started studying it, the cases began doing something I did not expect. They did not organize themselves into one giant reincarnation pattern. They began separating into different systems.

One of the first patterns involved the amount of time between the death of the alleged previous person and the birth of the child. Stevenson reported a median interval of roughly fifteen months across hundreds of cases, but that overall number hides something much more interesting. When the previous person was a relative or acquaintance, the interval tended to be much shorter. When the previous person was a stranger, the interval could stretch into years or even decades. The same general direction appears in different populations. European family and acquaintance cases tend to have intervals measured in months, while European stranger cases extend into years. American stranger cases can extend into decades. Tribal populations, meanwhile, can contain extraordinarily high proportions of family cases.

That immediately changed the way I looked at the famous fifteen-month figure. There may be no particularly meaningful thing called the average reincarnation interval. What we may actually be seeing is a mixture of different populations with different temporal behavior. If one culture contains many family cases and another contains many stranger cases, their average intervals should differ even before we invoke religion, metaphysics, or anything else. Culture, social relationship, and time between the two alleged lives were beginning to interact, and that was exactly the sort of hidden architecture I had hoped AI might help expose.

Then violence began to fall apart as the explanatory variable I initially thought it might be. Stevenson and others understandably separated violent deaths from natural deaths because murder, suicide, drowning, accidents, war deaths, and other traumatic endings appear so often in these cases. But when we started crossing variables instead of looking at them individually, violence itself stopped looking like the most interesting distinction. What began to look more interesting was whether the death was expected.

Someone dying slowly from an illness can anticipate death. Someone planning suicide can anticipate death. A soldier in a war may know that death is possible. Someone killed instantly in an unexpected automobile accident cannot. Suicide became particularly revealing because it behaves strangely if violence is supposed to be the controlling variable. Suicide can be physically violent and psychologically traumatic, yet one published group of solved suicide cases had a median intermission of only about eight weeks, with nine of eleven involving relatives or friends. That looks much more like the short-gap family pattern than the long-gap stranger pattern we were seeing elsewhere.

At the same time, suicide cases can contain the behavioral features associated with traumatic deaths. Children may reportedly fear the weapon or method involved, reenact aspects of the death, or show other specific behaviors. That gave us something far more interesting than a violent-versus-natural division. It suggested two potentially independent axes. One concerns trauma. The other concerns preparation or expectation. Suicide can occupy both simultaneously because a death can be anticipated and traumatic at the same time. Once I saw that possibility, the traditional violent-versus-natural division began to look much too crude for the data.

The birthmark and phobia cases then gave us another surprise. It would be easy to assume that a child born with a congenital mark corresponding to a previous person’s wound and a child terrified of whatever supposedly killed that person are displaying two versions of the same phenomenon. The available numbers suggest something more complicated. Among forty-seven drowning cases, thirty children reportedly had a fear of water, about 64 percent. Among thirty snakebite cases, thirteen reportedly had a fear of snakes, about 43 percent. In those same thirty snakebite cases, ten reportedly had a birthmark associated with the bite or its treatment, about 33 percent.

That comparison matters because we are looking inside the same death category rather than comparing unrelated collections. Phobias and marks are both elevated, but they are not interchangeable. Some cases have one, some have the other, some have both, and some have neither. Whatever ultimately explains that pattern would therefore have to explain why one alleged previous event appears associated with two partially independent channels of expression. One is behavioral and affective. The other is somatic. They may share an upstream association with trauma without being manifestations of the same downstream process.

The marks themselves then divided into categories that I had not appreciated when we started. Some are reported as corresponding to wounds. Someone is shot, stabbed, burned, bitten, or otherwise injured, and a later child is reportedly born with a mark or defect in a corresponding location. But other cases involve intentional or experimental marking traditions in which a dying person predicts, requests, or is deliberately marked with the expectation that something corresponding will appear on a later child. Those are fundamentally different claims. One concerns trauma and morphology. The other introduces intention. Calling both of them simply “birthmark cases” throws away exactly the distinction that pattern recognition should preserve.

The extremely short death-to-birth cases created another problem for any simple explanation. In one published collection of thirty-eight cases selected for intervals shorter than nine months, congenital anomalies were heavily represented. That sample is selection-biased and cannot be treated as representative of Stevenson’s entire archive, but the internal pattern still raised a fascinating developmental question. If there were any real relationship between a previous death and fetal development, timing should matter because a fetus with seven months of development remaining is biologically different from one only hours from birth. Stevenson himself recognized that some extremely short-gap cases created conceptual problems and sometimes discussed them in terms closer to possession or displacement because the fetus already existed when the previous person died. That suggests the extreme tail may not simply be a more intense version of the ordinary cases. It may represent a different class altogether.

The sex-change cases added another layer. Their frequency varies enormously between cultures, which means cultural influence cannot simply be waved away. Some populations report many opposite-sex previous-life cases, while others report almost none. Yet culture does not erase the internal behavioral association. A statistical study from the University of Virginia group found that children reporting an opposite-sex previous life had dramatically increased odds of gender-nonconforming behavior, with a reported odds ratio of 66.5. Myanmar itself independently showed much higher odds of gender nonconformity than other countries. In other words, culture and the reported sex of the previous person both appear to load onto the behavioral outcome.

There is also a striking directional asymmetry. Girls reporting male previous lives substantially outnumber boys reporting female previous lives. A large Burmese sample produced roughly a seventy-thirty split in that direction, and the twin data we examined showed the same general asymmetry. This means a simple cultural explanation has work to do, because it must explain not merely why cultures differ in how often these reports occur but why the reports have internal behavioral relationships and directional asymmetries. A reincarnation explanation has work to do as well, because it must explain why culture can change the frequency so dramatically. Neither side gets to erase the inconvenient half of the pattern.

The twin cases taught us something different, and perhaps even more valuable. Some of them are fascinating because the alleged previous personalities died together or were closely connected. Among thirty-four twin cases for which the timing of the previous deaths was known, twenty-one, about 62 percent, reportedly died on the same occasion or from the same cause. That is interesting because it suggests paired death may be associated with paired birth in this particular subset, but it is also clearly not an absolute rule.

Then we encountered a claim that sounded spectacular. Stevenson reported that when dominance could be determined between the alleged previous personalities and between the twins, the dominant previous person became the dominant twin every time. At first glance, that sounds like one of the strongest patterns in the archive. When we looked harder, it became one of the weakest. The problem was hidden inside the phrase describing cases in which dominance was known or could be “surmised.” Dominance is a subjective classification. The denominator was not clearly fixed; ambiguous cases could disappear from the comparison, and the researcher evaluating the proposed identity correspondence was also evaluating an inherently interpretive personality characteristic. We could not find an independent analysis specifically auditing that perfect dominance claim.

That became one of my favorite lessons from the entire exercise because it taught me to become more suspicious of perfect patterns, not less. Thirty out of forty-seven tells me something. Thirteen out of thirty tells me something. Ten out of thirty tells me something. An odds ratio with a confidence interval tells me something. Messy numbers with real denominators can be investigated. “Every time” becomes much less impressive when I cannot tell exactly what qualified for inclusion or how many ambiguous cases were removed before anyone started counting.

We found another example of this problem when looking at cultural sex-change rates. Stevenson investigated seven Kutchin cases and found six sex-change cases, an astonishing 86 percent. A larger historical sample from the same population contained twenty-two sex-change cases among forty-four, or 50 percent. The phenomenon did not disappear, but the dramatic appearance of near universality did. That is exactly what a pattern-recognition project should discover. Sometimes the important finding is not that a pattern survives. Sometimes it is discovering that a pattern shrinks when the denominator becomes real.

By the end of this work, I had stopped seeing Stevenson’s archive as one enormous pile of evidence for or against reincarnation. I began seeing something considerably more interesting. The archive appears to contain several partially independent systems. Social relationship is associated with intermission length. Culture affects the composition of family and stranger cases and strongly affects some reported behaviors. Trauma is associated with phobias and congenital correspondences, but those two outcomes do not simply collapse into one another. Expectedness or preparation for death may operate on a different axis from trauma, with suicide providing an especially interesting intersection of the two. Morphological cases divide between apparent wound-correspondence and intentional marking. Sex-change cases contain a powerful behavioral association while also being heavily culture-dependent. Twin cases may contain their own paired-death structure, while simultaneously demonstrating how an apparently flawless rule can emerge from subjective categories and uncertain denominators.

That is where I think we may have done something new. I am not claiming that nobody before us noticed any of these individual relationships. Stevenson noticed many of them, and researchers after him have statistically examined several. The difference is that I approached the archive as a pattern-recognition problem and deliberately allowed variables that had traditionally been discussed under one umbrella to separate from one another. Instead of asking which fact supports reincarnation, I asked which variables travel together, which ones do not, which relationships reproduce across cultures, which disappear when the sample changes, and which apparently unrelated variables suddenly intersect when the data are viewed from another angle.

That changes the nature of the question for me. I am no longer looking at thousands of strange stories and asking whether I believe them. I am looking at a system with variables, dependencies, exceptions, subpopulations, possible latent factors, and competing causal structures. Some variables move together. Others stubbornly refuse to. Some relationships appear across very different populations. Others change dramatically with culture. Some spectacular claims weaken when their denominator is examined, while some much less dramatic relationships keep appearing every time the data are divided another way.

This is where modern artificial intelligence can contribute something Stevenson could not have had when he began collecting these cases. Stevenson had the archive. He and his colleagues spent lifetimes building it. What we now have is a new instrument capable of looking across that archive and asking questions that become almost impossible when thousands of cases have to be processed one human mind at a time.

I do not think the responsible next move is to declare that we have found reincarnation, Host-Jumping, cultural mythology, anomalous information transfer, developmental psychology, or any other final explanation. As a theoretical philosopher, I am interested in the possibility space before I am interested in closing it. The patterns should be allowed to come first, because once we force them into an explanation too early, we risk seeing only what the explanation taught us to look for.

The world already has thousands of stories about children who supposedly remember previous lives. Adding another story to that pile would accomplish very little. What may actually be new is the possibility that underneath those stories is a discoverable architecture, one that nobody could see clearly when the cases were examined primarily one at a time. If that architecture continues to survive as more of Stevenson’s archive is brought into the analysis, then the important discovery may not initially be what these children are remembering or whether anyone has traveled from one life to another. The discovery may be that the strange reports Stevenson spent his life collecting are not distributed randomly at all, but are organized by a set of relationships and constraints that any serious explanation, conventional or extraordinary, will eventually have to reproduce.

Charles Richard Walker

Functional Consciousness Project
Zenodo