AI Mistakes Mirror Human Brain Condition
AI language errors resemble brain activity in aphasia, offering clues for diagnosing the disorder.
Complete the form below to unlock access to ALL audio articles.
Artificial intelligence (AI) tools such as chatbots and agents based on large language models (LLMs) are increasingly integrated into daily life. While these models, including ChatGPT and Llama, produce responses that appear fluent, they can also generate convincing but incorrect information. Researchers at the University of Tokyo have identified similarities between these AI errors and a human language disorder called aphasia, where individuals speak fluently but sometimes produce unclear or meaningless statements. This connection may offer new approaches for diagnosing aphasia and provide insights for improving AI systems.
Aphasia
A neurological disorder that impairs language abilities, often caused by brain injury or stroke. People with aphasia may have difficulty understanding or producing speech despite being able to speak fluently in some cases.Large language model (LLM)
A type of AI trained on large datasets of text to generate human-like language. Examples include ChatGPT and Llama.AI fluency and the risk of misinformation
As the use of AI-generated text grows, ensuring the accuracy and coherence of these tools has become critical. Many popular LLM-based systems produce highly articulate responses. However, these responses can contain fabricated or misleading content. Users unfamiliar with a topic may mistakenly trust inaccurate outputs due to the confident style of the AI-generated text.
The research team, led by Professor Takamitsu Watanabe at the International Research Center for Neurointelligence at the University of Tokyo, observed that this phenomenon resembles the speech patterns seen in Wernicke’s aphasia. People with this condition often speak fluently but their statements may lack clear meaning. This observation motivated the researchers to investigate whether the internal processes of LLMs share characteristics with brain activity in aphasia.
Wernicke’s aphasia
A form of aphasia characterized by fluent but often nonsensical speech and difficulties understanding language.Comparing brain activity and AI model data
To explore this idea, the researchers applied energy landscape analysis, a technique originally developed in physics to visualize energy states in magnetic metals. This method has recently been adapted for neuroscience to study brain activity. The team analyzed resting brain activity patterns from individuals with different types of aphasia and compared these to internal signal patterns from several publicly available LLMs.
Energy landscape analysis
A method used to visualize and analyze the possible states of a complex system by representing them as positions on a surface, where the shape of the surface influences system behavior.The analysis revealed notable similarities between the ways information is processed within LLMs and the brain signals observed in aphasia. Specifically, the movement and manipulation of signals inside AI models resembled brain activity in people affected by Wernicke’s aphasia.
Professor Watanabe explained that energy landscapes can be visualized as a surface with a ball that moves according to the curves of that surface. When the curves are steep, the ball settles into a stable position. If the curves are shallow, the ball moves more erratically. In this analogy, the ball represents the brain state in aphasia or the ongoing signal pattern in an LLM.
Implications for neuroscience and AI development
This research offers potential new directions for classifying and monitoring aphasia based on internal brain activity rather than relying solely on external symptoms. For AI development, understanding these parallels could inform diagnostic tools that identify limitations within AI models and guide improvements in their architecture.
“We’re not saying chatbots have brain damage. But they may be locked into a kind of rigid internal pattern that limits how flexibly they can draw on stored knowledge, just like in receptive aphasia. Whether future models can overcome this limitation remains to be seen, but understanding these internal parallels may be the first step toward smarter, more trustworthy AI too,” said Watanabe.
Despite the similarities, the researchers caution against equating AI systems with human brain damage. They suggest that LLMs may operate within rigid internal patterns that limit their ability to flexibly use stored knowledge, akin to receptive aphasia.
Reference: Watanabe T, Inoue K, Kuniyoshi Y, Nakajima K, Aihara K. Comparison of large language model with aphasia. Adv Sci. 2025;n/a(n/a):2414016. doi: 10.1002/advs.202414016
This article has been republished from the following materials. Note: material may have been edited for length and content. For further information, please contact the cited source. Our press release publishing policy can be accessed here.
This content includes text that has been generated with the assistance of AI. Technology Networks' AI policy can be found here.
