In May 2024, a team of researchers from the University of California, San Diego, reported that in a series of Turing tests, people could not distinguish the large language model GPT-4 from a human being. The year before, a Stanford-based group of researchers had GPT-4 take the bar exam – and the artificial intelligence passed. GPT-4 has passed medical board residency exams, written articles online, and written useful code. The list goes on. So, the question becomes: Are AI models processing natural language in a way similar to the human brain, or are they processing language in entirely different ways?
“That’s a concrete question that we want to answer,” said Uri Hasson, professor of psychology and neuroscience at Princeton University. “It’s really important to know.”
As a neuroscientist, Hasson is working on opening up the “black box” of the human brain – the black box being a reference to any systems that operate in ways that aren’t visible to or understood by us. The same metaphor of the black box can also apply to the inner workings of large language models. If these increasingly sophisticated language models like ChatGPT indeed operate similarly to the human brain, then studying and understanding them will help researchers reach their goal of understanding human language processing as well.
“When we open the black box of the model and the black box of the brain, we see a lot of similarities and a lot of differences,” said Hasson. His research indicates both systems rely on similar computational principles, though the way these principles are implemented within each black box is drastically different.
A similar objective
On Oct. 22, Hasson gave a seminar at the Center for Statistics and Machine Learning as a part of the center’s ongoing Lunchtime Faculty Seminar series. In his talk, Hasson expounded upon some of the key similarities between large language models and the human brain.
Machine learning models such as ChatGPT learn to generate text by predicting what the next word in a given sequence will be. Hasson's lab found that similar to language models, the human brain is also engaged in constantly predicting what the next word will be when listening to another person talk. “This indicates that the human brain is engaged in optimization of an objective which is very similar to the objective we optimize when we train language models,” explained Hasson.
At the same time, there are also key differences between AI and people. A large language model learns by “reading” information taken from around the entirety of the internet – billions of words. A singular person, Hasson pointed out, could never read all books and posts on the internet.
For a more realistic comparison between the model and humans, Hasson is a part of a research team which feeds a neural network with human-centered naturalistic data. “We are conducting a project, called the First 1,000 Days Project, that monitors how babies learn to see and speak by recording their lives at home for 12 hours a day over 1,000 days, from birth until they are three years old,” said Hasson. The researchers aim to train models that will learn to speak like a baby using each baby's inputs rather than data retrieved from the internet.
Hasson said that the field is reaching a point where researchers are finally starting to gain a mechanistic understanding of how the brain develops and learns language in real-world contexts. “Only a few years ago, I wasn’t sure we would ever be able to understand how the brain processes language during our lifetimes,” he said. “Now I’m thinking we will have real progress, which is exciting.”