Dr Ariel Goldstein is a computational cognitive neuroscientist at the Center of Human Inspired AI (CHIA) at University of Cambridge. His research investigates how human brains and artificial neural networks process language, particularly during natural activities such as listening to stories, conversations and lectures.
By combining brain recordings with computational language models, he asks whether biological and artificial systems develop similar ways of representing sounds, words, context and meaning, and what both the similarities and differences can teach us about language, intelligence and the human mind.
What is the main aim of your research, and what key questions does it address?
The broad aim of my research is to understand the computational principles that allow language to become meaningful.
When we listen to someone speak, our brain does much more than process one word at a time. It continuously combines the current word with what came before, our knowledge of the world, and our expectations about what might come next. This allows us to move rapidly from sounds to words, from words to ideas, and from individual sentences to a broader understanding of a conversation or story.
Modern language models face a related challenge: they must also use context to interpret language and predict what is likely to come next. My research compares patterns of activity in the human brain with the internal representations that emerge in these artificial systems.
Some of the questions we ask are: How does linguistic context accumulate over time? How do representations change as information moves from sound to language and meaning? Is language represented only in specialised brain regions, or is it distributed across several interacting systems? And when human brains and artificial models appear similar, does this reflect a genuinely shared computational principle, or simply the fact that both are responding to the same language?
What does your research involve on a daily basis?
My daily work combines neuroscience, computer science and psychology.
On some days, I analyse brain data recorded while people listen to or produce natural speech, such as stories, conversations or university lectures. We align each word with the corresponding brain activity and then compare these neural responses with the representations generated by language models processing the same material.
A complementary part of my research works in the opposite direction: I study how modern AI systems achieve their remarkable capabilities and use what we learn from them to formulate new hypotheses about the human mind and brain. These hypotheses can then be tested directly using human behaviour and neural recordings. In this sense, AI is not only something we try to understand through neuroscience; it is also becoming a new conceptual and computational tool for understanding human cognition.
More broadly, I use deep-learning frameworks to revisit classical questions in psychology and neuroscience. For example, I am interested in what kinds of computations the brain performs, how information is represented and transformed across different stages of processing, and how systems such as reward and motivation might be understood through the lens of modern machine learning.
What outcomes or contributions do you hope your research will lead to?
I hope this research will help establish a more unified scientific framework for studying biological and artificial intelligence.
Artificial neural networks can serve as explicit computational hypotheses about how information might be represented and transformed. Unlike the human brain, we can inspect these models in detail, alter their components and test how those alterations affect their behaviour. At the same time, the human brain provides an important reference point for evaluating what artificial systems learn and how their internal organisation differs from human cognition.
In the longer term, this dialogue between neuroscience and AI could lead to better theories of language and cognition. I also hope the work will help create a shared language between disciplines that have often studied intelligence separately.
What aspect of your work is most exciting to you right now?
There are three directions that I find particularly exciting at the moment.
The first is understanding how thoughts are translated into words. We have remarkably little conscious access to this process: when we speak, we usually know what we want to say, and then the words simply seem to emerge. The computations that transform an internal thought into a sequence of words have therefore remained something of a mystery. With new neural recordings and computational models, we are beginning to observe this process in much greater detail and ask what happens in the brain in the seconds before a word is spoken.
The second is the idea of artificial motivation. As AI systems become increasingly capable and autonomous, I am interested not only in what they can do, but in what drives their behaviour. Can concepts such as goals, rewards, preferences and motivation—concepts traditionally studied in psychology and neuroscience—help us understand artificial systems? And conversely, can studying how motivation emerges in artificial systems teach us something new about motivation in humans?
The third is understanding collaboration across humans and machines. I am interested in what makes human–human, human–AI and eventually AI–AI collaboration successful. Intelligence is often studied at the level of an individual, but many of our most important achievements arise from groups of minds working together. Understanding when different individuals or systems complement one another—and when they interfere with one another—could help us develop a broader theory of collective intelligence.
How does the symposium theme, “Language and communication across boundaries” relate to your research?
My research crosses several kinds of boundaries.
The most obvious is the boundary between biological and artificial intelligence. By presenting the same language to humans and computational models, we can ask whether very different systems arrive at related ways of representing meaning.
The work also crosses disciplinary boundaries, bringing together neuroscience, psychology, linguistics and computer science. Each field describes language at a different level, and the challenge is to connect these levels rather than treating them as separate explanations.
Language itself is also a mechanism for crossing boundaries. It allows private thoughts and experiences in one mind to become communicable to another. Meaning must travel from a speaker’s intentions, through sounds or text, into the changing neural activity of a listener. Artificial systems now participate in this process as well: they learn from language produced by many people and cultures, and their outputs increasingly influence human communication in return.
My research asks what happens to information as it moves across these boundaries—between words and meaning, between one mind and another, and between human and artificial systems.
Is there anything else you would like to add?
I think it is important to avoid two extremes when discussing language and AI.
One is to assume that fluent behaviour automatically means that an artificial system thinks, understands or experiences language exactly as a human does. The other is to dismiss similarities between artificial systems and the brain as superficial simply because the systems were built differently.
My aim is to take an empirical position between these extremes: to identify precisely which representations and computational principles are shared, which remain distinct, and what those findings allow us to conclude about language and intelligence.