Evelina Fedorenko found that how people process sentences mirrors what happens inside AI language models.
The episode opens with a framing paradox: AI is being used to model human behavior, but we cannot verify its accuracy because we do not yet fully understand our own minds. Host Dina Temple-Ruston visits the lab of Dr. Evelina Fedorenko at MIT's McGovern Institute for Brain Research, where two human brains in jars sit on a shelf above her desk. Fedorenko is introduced as a neuroscientist focused on how the brain processes language and thought, two things most people assume are the same. For centuries, philosophers from Plato to Noam Chomsky argued that thought is formed by language, not merely communicated through it. Fedorenko was skeptical of this assumption and designed fMRI experiments to test it. She identified which brain regions activate during language tasks, then had subjects perform non-language tasks like math, and found the language regions went essentially silent. After hundreds of MRIs over 15 years, she concluded that language and thought rely on distinct neural networks. This discovery also suggested the brain's language network could potentially be mapped and repaired for people with language processing disorders. A persistent bottleneck in this research was the inability to run controlled experiments on living human brains — unlike colleagues studying vision or motor control, Fedorenko could not use animal models because animals do not use human language. The emergence of large language models changed her research trajectory. She fed language tasks to both LLMs and human volunteers and found that the internal representations of sentences were strikingly similar across both systems. She concluded that both the human brain and LLMs operate as predictive systems, constantly anticipating what comes next. This discovery effectively gave neuroscientists a new laboratory subject: a digital model they could pause, probe, alter, and dissect in ways impossible with human subjects. Postdoc Andrea De Varda explains the concept of digital dissection, noting that individual digital neurons can be isolated and deleted to test specific cognitive hypotheses. If removing a component changes behavior, the hypothesis gains support; if nothing changes, it must be revised. The research then expanded from language to reasoning. A new class of large reasoning models, trained on math and logic problems, shows its work step by step, generating 264 tokens to solve a simple subtraction problem and far more for complex equations. Fedorenko's team found that problems humans solve quickly, these models also solve quickly, and harder problems slow both down similarly. This parallel raises the possibility that large reasoning models could serve as experimental proxies for studying human reasoning, learning, and intelligence. The episode closes by noting that the same AI systems illuminating the brain are also being deployed commercially to anticipate human desires, decisions, and behaviors, suggesting the gap between human and machine cognition may be smaller than previously imagined.
