When people think about what ChatGPT or other large language models are doing behind the scenes, it’s common to reach for the language of computing. You might assume there is a database, or a logical series of steps being executed, in order to provide a response to your query.
Picturing AI responses as akin to your computer pulling up a file is a sensible line of thinking. But it’s incorrect. Understanding the central influence of cognitive science in AI’s evolution helps explain why.
AI’s history isn’t only a computer science story. Fundamental advances that gave rise to today’s AI were brought about by people whose goal was to understand the human mind, not help you with your homework or travel itinerary. This distinction is important, because it helps better explain how these systems “behave” the way they do.
We are cognitive scientists who teach students about the interdisciplinary roots of our field, including its deep and often overlooked ties to computer science and AI. There’s a shared history that often gets lost in conversations about what AI is doing.
AI’s lineage is often overlooked
The term “artificial intelligence” was coined in 1956, at a Dartmouth College workshop. At the time, people thought that a machine could be made intelligent if you could write down the right set of rules for it to follow.
Frederic Lewis/Archive Photos via Getty Images
Rather than build a machine that followed rules, however, psychologist Frank Rosenblatt built one in 1958 – the Perceptron – that learned from examples. Rosenblatt hadn’t been at the Dartmouth workshop, but he was thinking about how to make a machine intelligent in a fundamentally different, more human way. His key insight was to build an artificial neural network: a computer algorithm modeled loosely on the architecture of the brain.
This approach built on the work of psychologists like Donald Hebb who were studying brains and behavior. Specifically, Hebb explored how neural connections strengthen when used, a fundamental insight that later seeded machine learning.
Further advances came in the 1980s, as psychologists sought to explain how people perform complex cognitive tasks, such as recognizing words and forming memories. Cognitive and computer scientists together – David Rumelhart, Geoffrey Hinton and Ronald Williams – figured out how to train artificial neural networks with multiple layers, an idea that later became known as “deep learning.” These multilayered neural networks were capable of more complex tasks and generalizing what they “learned” to apply to new examples.
In the following decades, computer scientists and engineers enabled computing capabilities to be scaled up, developing graphics chips, the transformer architecture and more. Without these systems, modern AI would still struggle to recognize letters and words, rather than be able to hold realistic conversations in natural language.
The fundamental insights, however, came from the study of the mind: AI should learn from examples, and its architecture should be based on the human brain.
These foundational ideas matter for reasons of practicality. Modern AI isn’t born from rules – it’s grown from examples. And things that are grown are not as predictable as systems based on rules.
Its origin helps explain its ‘behavior’
Human intuitions and mental models about how technology works govern how people use it. The idea of AI as software – deterministic, fact-retrieving – causes people to misinterpret the outputs they provide.
AI “hallucinations” are common; chatbots often state falsehoods. Thinking of AI as a database makes this behavior appear confusing. How could a system look up an answer that doesn’t exist?
But human memory doesn’t “look up” answers. It is reconstructive and imperfect, filling in gaps with plausible details. Thanks to pioneering work by psychologist Elizabeth Loftus, researchers know that false memories are unfortunately common and can be implanted with relative ease.
A system that “looks something up” gives the same answer every time, but it is therefore limited in responding to a finite set of questions. Human minds don’t work that way, and neither does AI. The price to pay for being able to generalize to new questions and tasks is a system prone to confabulation.
Thinking of AI as like a primitive brain also helps explain additional puzzling behaviors, such as getting different answers to the same question. Ask your young child what they want for dinner twice and you may get different requests. Framing matters when interacting with an AI system too, as these systems can be nudged into giving you answers you prefer. They operate in a probabilistic fashion, much like human memory, not in a deterministic fashion, like a calculator.
None of this means that AI necessarily thinks like people do, if the systems think at all. Humans and AI commonly “perceive” things differently; computer vision can spot subtle abnormalities that human screeners might miss, but might also misinterpret something, such as a defaced road sign.
Geraint Rowland Photography/Moment via Getty Images
Why those origins still matter
If modern AI is unpredictable, how can people come to understand it?
The answer is by understanding where this technology came from. Recognizing that AI systems are more akin to brains than databases can help people build more accurate mental models of what AI can do and can lead to better ways of understanding how these systems work.
- Users overtrust the wrong things. With a significant proportion of ChatGPT users indicating that they use ChatGPT like a search engine, it seems that people are treating outputs as retrieved fact rather than fluent, confident guesses. This misunderstanding can cause users to skip the verification step they’d likely apply to a person and potentially make decisions, including financial ones, based on confabulation.
- Institutions grapple with differing mental models. Faculty remain split between banning AI tools outright and building them into coursework, treating AI as producing answers to be checked or trusted like collaborators. Divisions are causing confusion and sowing rifts among (and within) educators and their students.
- Computer science is rediscovering cognitive science’s questions. Today’s systems are opaque even to their builders. One approach to understanding them, probing what goes in and what comes out, resembles behaviorism, a once-dominant field of psychology that studied only observable behavior. Researchers are now moving toward interpretability: opening up the models to see what’s happening inside – closer in spirit to neuroscience – and working on problems cognitive scientists have chewed on for decades.
The exchange runs both ways. Just as computer science draws on cognitive science’s questions and methods, cognitive science draws on the accessibility of AI models themselves, using them to test and refine theories of the mind. Computer science and cognitive science split off from one project decades ago: understanding how intelligence works. The more these fields keep trading questions and tools, the better placed humanity will be.
Leave a Reply