According to Martin Schrimpf, artificial intelligence is about understanding how human intelligence functions rather than merely automating tasks.
Schrimpf’s goal is to use artificial intelligence (AI), specifically neural networks that mimic the interactions of real neurons, to build a digital twin of the human brain. His method is methodical: evaluate how the human brain reacts to language or visual tasks, contrast it with AI’s performance on the same tasks, and adjust the AI accordingly. This feedback loop eventually produces AI that functions more and more like the human brain.
Schrimpf started Brain-Score, an open-source platform that assesses how well AI models match data from the human brain, in order to scale this process. This platform has been used to test thousands of models against almost 100 neural and behavioral datasets since 2017.
Schrimpf initially wanted to work in technology, but he discovered that neuroscience offered him a greater sense of purpose. He studied in Germany and co-founded startups before relocating to MIT to pursue a Ph.D. in brain and cognitive sciences. He now manages the NeuroAI Lab at ETH Lausanne, back in Europe.
He co-authored a study in 2023 that showed how artificial intelligence (AI)-generated sentences could predictably increase or decrease brain activity. This was the first noninvasive way to affect high-level brain functions. This creates the possibility of using customized brain stimuli to treat diseases like dyslexia or depression.
Why Pay Attention to Language and Vision?
Due to the ease of presenting stimuli on screens, Schrimpf started with visual processing because it is a well-mapped area in neuroscience. He went on to examine whether models developed for sensory systems would also work for language, and they did. It turns out that human language systems encode information similarly to how vision does, indicating that different functions share cognitive structures.
Recognizing and Complementing the Brain
Because every component of AI models is accessible, even the most complex ones can be broken down and comprehended. Researchers can verify which models most closely match reality by using Brain-Score to compare model outputs with human brain responses. Many models are approaching human-like performance in language and vision tasks.
Schrimpf claims that despite these networks’ shortcomings, their value is underestimated. They can mimic and even reflect the behavior of the real brain because of their structure, which is similar to that of neurons.
Controversies and Critics
According to some neuroscientists, Schrimpf’s research ignores psychological information. He acknowledges that his method essentially bets on outcomes rather than comprehension by eschewing classical neuroscience in favor of data-heavy, model-driven techniques. He thinks both conventional and AI-driven neuroscience can coexist, despite their differing advantages.
The Digital Brain Dream
A functional digital replica of the human brain is Schrimpf’s ultimate objective. Brain-Score monitors your progress in that direction. He believes it can be accomplished in a few decades.
Once attained, such a model could be used to create interventions, such as fonts that are accessible to people with dyslexia, or to customize depression treatments. By first simulating the effects in a digital twin, even invasive treatments like brain stimulation could be optimized.
Brain Influence’s Ethical Challenges
There are significant ethical issues with using AI to modify brain activity. Although it is possible to influence perception, decision-making is still out of reach — at least for the time being.
Prior to these tools becoming products, Schrimpf stresses the significance of working with ethicists and developing policy frameworks. Many capabilities currently exist without legal oversight, which is dangerous.
He is concerned that innovation frequently outpaces regulation. The need for responsible development and foresight grows as AI’s capabilities increase.
Does Human Intelligence Only Exist in Computation?
Despite being entirely computational, many AI models are able to forecast how the human brain will react. According to Schrimpf, this implies that intelligence is a process that can arise in other systems and is not just a biological phenomenon. Even though modern AI lacks biochemistry, it still exhibits traits that are similar to those of human cognition.
Schrimpf views this as an opportunity to learn more about what makes us human, rather than as a threat. According to him, our life experiences—rather than our neurons—are what define us as individuals.
