In an audience question-and-answer episode of Moonshots with Peter Diamandis, a listener named Nathan said he had a question closer to psychology. AI systems keep collecting data about all of us, he said. Would that data eventually become detailed enough to show people what is going on in their own minds, including brain patterns and psychological states they could never identify on their own?
The episode was recorded on September 18, 2026, and published on September 22. Host Peter Diamandis, founder of XPRIZE, answered first. He said a large number of companies are working on brain-computer interfaces, meaning devices that record from or stimulate the brain directly. He described a conversation, following an introduction from someone he called Ray, with a company that is placing "the equivalent of nanobots" throughout the brain to read and write neural activity. He said this would give "incredible fidelity." He did not name the company.
He then passed the question to Alexander Wissner-Gross, a computer scientist and founder of Reified. Wissner-Gross gave a different kind of answer, one that needs no brain implant.
The argument: brain states leak into training data
"I think large language models, foundation models trained off of human behavior on the internet are already a weak form of human mind uploading," Wissner-Gross said. A foundation model is a large, general-purpose AI system trained on huge amounts of data, much of it text people have written online.
His reasoning runs in three steps. First, what people write and do online is shaped by what happens inside their heads. In his words, internal brain state is "quite leaky into the training data of internet behavior." Second, training squeezes that data into a model. He said the behavior "in turn is then compressed into the foundation models." Third, the models therefore carry some trace of human mental processes: "the foundation models do, yes, I think reflect internal brain state." He expects people to see this "with the benefit of hindsight."
His "exhibit A" was what he described as work from Jean-Rémi King's group at Meta, done a few years earlier. According to Wissner-Gross, it showed that even the hidden activations of GPT-2, an earlier-generation language model, were "linearly correlated with fMRI voxels in the human brain." Hidden activations are the numbers a model computes internally as it processes each word. fMRI is a brain-scanning method that tracks changes in blood flow linked to neural activity. A voxel is one small three-dimensional unit of the resulting image.
What the 2022 study measured
The study that most closely matches that description is a paper by Charlotte Caucheteux, Alexandre Gramfort and Jean-Rémi King, published on September 29, 2022. The researchers used fMRI recordings from 101 people who listened to spoken stories: seven narratives totaling 70 minutes of distinct audio. That produced 237 usable pairings of one participant and one story.
The researchers fed the same stories to GPT-2 and recorded its internal activations. They then fit linear models, a simple kind of mathematical mapping, to predict each person's brain responses from GPT-2's activations. They tested those predictions on held-out data that the mapping had not been trained on. The correlation between predicted and actual brain activity became a "brain score."
The main finding links the model to understanding. After each story, participants answered questions about it. People whose brain responses GPT-2 predicted better tended to understand the stories better. The correlation between brain scores and comprehension scores was r = 0.50. On that scale, 1 would be a perfect match and 0 no relationship.
The paper also showed where the match was strongest. Representations that took the surrounding context into account helped explain brain regions involved in higher-level language processing. So did the model's longer-range "attention," the mechanism that lets it relate a word to earlier words. Early auditory regions, which handle sound more directly, lined up better with shallow representations that looked only a short distance back. Shortening the attention span in the model's early layers sometimes improved the match.
What the study leaves open
The authors point out the limits. The findings are correlational: they show that model activity and brain activity move together, not that one explains the other. Comprehension was measured with questionnaires about the stories. Attention, prior knowledge and working memory could all affect both brain responses and understanding, and none of them was experimentally controlled.
The match was also partial. Brain data is noisy, so researchers estimate a "noise ceiling": the most that any model could be expected to predict. GPT-2's brain scores reached at most 32 percent of that ceiling. Much of the brain activity that could in principle be explained was not.
Three separate claims
The exchange touches three ideas, and the evidence treats each one differently.
The first is shared representations. A language model trained on human text builds internal patterns that partly resemble how human brains process the same language. The 2022 study supports a version of this, for GPT-2 and spoken stories, within the limits described above.
The second is inference about an individual, which is what Nathan asked about: whether data collected about one person could reveal that person's hidden psychological patterns. The study did not test this. It predicted how people's brains responded to stories they were hearing. It did not try to recover their private thoughts or use their personal data or online history.
The third is mind uploading, meaning a copy of a specific person's mind. Wissner-Gross qualified his claim with "weak." He described the models as reflecting human brain states compressed from a vast amount of internet behavior. The study that matches his example shows a measurable but partial overlap between one model and the brains of people listening to stories. It does not show whether that overlap could ever capture a particular individual, which was the heart of Nathan's question. His answer in the episode did not address that step.