Cognitive science is the part of the AI story most people never hear. When someone pictures what ChatGPT does behind the scenes, they usually reach for computing words: a database, a search, a set of steps that fetches a stored answer. That picture is sensible. According to two psychologists writing in The Conversation on 7 October 2026, it is also wrong.
Arryn Robbins, an assistant professor of psychology at the University of Richmond, and Michael Hout, associate dean of research and professor of psychology at New Mexico State University, argue that modern AI behaves more like a brain than a database, and that its origin story explains why. The key ideas behind today’s systems, they write, came from people trying to understand the human mind: machines should learn from examples, and their design should be loosely modelled on the brain.
This article follows that thread. It traces the cognitive science research behind neural networks, from Donald Hebb in 1949 to the backpropagation paper of 1986, checks the primary sources, and shows how findings about human memory help explain AI hallucinations, inconsistent answers and flattery. It ends with practical rules for anyone who uses or deploys these systems at work.
Table of contents
- Why Cognitive Science Belongs in the AI Origin Story
- The 1956 Bet on Rules, and the Cognitive Science Alternative
- The Cognitive Science Pioneers Who Taught Machines to Learn
- How Cognitive Science Ideas Became Today’s Chatbots
- What Cognitive Science Says About AI Hallucinations
- Why Cognitive Science Explains Different Answers to the Same Question
- Where the Brain Analogy and Cognitive Science Part Ways
- How Cognitive Science Now Borrows Back From AI
- Practical Cognitive Science Lessons for AI Users and Teams
- Frequently Asked Questions About Cognitive Science and AI
- References and Further Reading
Why Cognitive Science Belongs in the AI Origin Story
Most histories of AI are told as computer science history: chips, programming languages, benchmark wins. Robbins and Hout say that leaves out the most important chapter, the one written by cognitive science.
The database picture most people carry
Picturing an AI reply as a computer pulling up a file is a reasonable guess. Most software people use every day does exactly that. A banking app looks up a balance. A search engine retrieves pages that already exist. A spreadsheet applies the same formula and returns the same result every time.
The authors say the guess is incorrect for a large language model, and that the error has consequences. “Human intuitions and mental models about how technology works govern how people use it,” they write, citing human-computer interaction research going back to 1993. If people believe AI is deterministic and fact-retrieving, they misread what it gives them, trusting it where they should check and puzzling over behaviour that cognitive science predicts.
Who is making the argument
Both authors are cognitive scientists who teach students about the interdisciplinary roots of their field. In March 2026 they wrote a companion piece, with Eben W. Daggett, on why AI does not “see” the way people do. Their new article makes a broader claim: AI’s history “isn’t only a computer science story.”
The fundamental advances, they argue, came from “people whose goal was to understand the human mind, not help you with your homework or travel itinerary.” That distinction matters because it explains how these systems behave. A machine designed to mimic a mind inherits some of the mind’s habits, including the inconvenient ones.
What the argument does not claim
The authors are careful about the limits. “None of this means that AI necessarily thinks like people do, if the systems think at all,” they write. The claim is about lineage and behaviour, not consciousness.
Cognitive science offers a better analogy than the database. That is a different thing from saying a chatbot has a mind, and the rest of this article keeps the two apart.
The 1956 Bet on Rules, and the Cognitive Science Alternative
To see why the origin story matters, start with the approach that did not come from psychology.
Dartmouth’s founding conjecture
The term “artificial intelligence” was coined for a summer workshop at Dartmouth College in 1956. The proposal, dated 31 August 1955 and signed by John McCarthy (Dartmouth), Marvin Minsky (Harvard), Nathaniel Rochester (IBM) and Claude Shannon (Bell Telephone Laboratories), asked for “a 2 month, 10 man study of artificial intelligence.”
Its central conjecture was that “every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.” For the next three decades much of the field took “precisely described” literally. Write down the right rules, the thinking went, and the machine would be intelligent.
The neuron-net question was already on the list
The picture is not quite black and white. The same proposal listed “Neuron Nets” as one of seven open problems, asking how “a set of (hypothetical) neurons” could be arranged “so as to form concepts.” The question was there from the start.
What changed the field was a group of researchers, most of them trained in psychology, who answered it by building machines that learned rather than machines that were told. That is where cognitive science enters the story.
Rules work until the world gets messy
Rule-based systems, later sold commercially as expert systems, are predictable by design. Every output can be traced to an instruction someone wrote. That is their strength and their ceiling: a system that only follows rules can only handle the cases its authors anticipated.
Recognising a handwritten letter, a spoken word or a face turned out to need more rules than anyone could write down. People do these things effortlessly, without being able to state the rules they use. Explaining how they manage it was precisely the question cognitive science was asking.
The Cognitive Science Pioneers Who Taught Machines to Learn
The ideas that power today’s AI came in three waves, and psychologists led each one.
Donald Hebb and the strengthening connection
The first link in the chain is a psychologist. Donald Hebb (1904 to 1985), often called the father of neuropsychology, published The Organization of Behavior: A Neuropsychological Theory in 1949. Hebb explored how connections between neurons strengthen when they are used. Robbins and Hout call it “a fundamental insight that later seeded machine learning.”
Hebb’s idea turns learning into a change in connection strength rather than a new entry in a file. That is still the basic mechanism of every neural network. Training adjusts the weights between units, and knowledge lives in those weights, spread across the whole system rather than stored in labelled rows that can be looked up.
Frank Rosenblatt’s perceptron
Frank Rosenblatt was not at Dartmouth. He was a psychologist, a Cornell psychology graduate (class of 1950, PhD 1956), and he was thinking about intelligence “in a fundamentally different, more human way.” In 1958 he published “The perceptron: A probabilistic model for information storage and organization in the brain” in Psychological Review (volume 65, pages 386 to 408). The title alone shows where the idea came from.
In July 1958 the US Office of Naval Research demonstrated the system on a five-ton IBM 704 computer fed with punch cards. According to Cornell, after 50 trials it “taught itself to distinguish cards marked on the left from cards marked on the right.” The New York Times headline read: “New Navy Device Learns by Doing: Psychologist Shows Embryo of Computer Designed to Read and Grow Wiser.”
Hype, a hostile book and a long winter
The press went further than the science. The New Yorker called the perceptron “the first serious rival to the human brain ever devised.” Rosenblatt himself described it as “the first machine which is capable of having an original idea.”
In 1969 Minsky and Seymour Papert published Perceptrons, a book that set out the limits of single-layer networks and is widely blamed for drying up interest in the approach. Rosenblatt died in a sailing accident in 1971, on his 43rd birthday. Cornell’s own retrospective is headlined “Professor’s perceptron paved the way for AI – 60 years too soon.”
Rumelhart, Hinton and Williams make depth trainable
The revival came from cognitive science again. In the 1980s psychologists wanted to explain how people recognise words and form memories. David Rumelhart, a psychology professor who “studied how people think and learn complex skills such as reading and the use of language,” worked with Geoffrey Hinton and Ronald Williams on a way to train networks with several layers.
Their paper, “Learning representations by back-propagating errors,” appeared in Nature in October 1986 (volume 323, pages 533 to 536). It showed how to pass an error signal backwards through the layers so every connection could be adjusted. That method, backpropagation, is still how multilayer networks learn, and it is the root of what later became known as deep learning.
Connectionism: no neuron works alone
Rumelhart, who died in 2011 at 68, led a team with James McClelland that built computer models of human perception, language and memory in the 1970s and 1980s. Stanford described his central idea, connectionism, as “the idea that no single neuron in the human brain does its job alone in processing information.”
That sentence is also a fair description of a modern language model. No single weight holds a fact. Knowledge is distributed across billions of them, which is why there is no row to delete when a model gets something wrong, and why the cognitive science framing fits better than the database one.
| Year | Who | Background | Idea that survives in today’s AI |
|---|---|---|---|
| 1943 | Warren McCulloch and Walter Pitts | Neurophysiology and logic | A mathematical model of the neuron |
| 1949 | Donald Hebb | Psychology | Learning as connections that strengthen with use |
| 1955 to 1956 | McCarthy, Minsky, Rochester, Shannon | Mathematics and engineering | The name “artificial intelligence” and the rules-first programme |
| 1958 | Frank Rosenblatt | Psychology | The perceptron: a machine that learns from examples |
| 1969 | Minsky and Papert | Computer science and mathematics | Limits of single-layer networks, and a long pause |
| 1982 | John Hopfield | Physics | Associative memory that reconstructs stored patterns |
| 1986 | Rumelhart, Hinton and Williams | Psychology and computer science | Backpropagation through hidden layers |
| 2017 | Vaswani and colleagues | Computer science | The transformer architecture behind chatbots |
| 2024 | Hopfield and Hinton | Nobel Prize in Physics | Recognition for neural-network machine learning |
How Cognitive Science Ideas Became Today's Chatbots
The psychology supplied the ideas. Engineering supplied the scale that made them useful.
Engineering supplied the scale
Learning from examples needs a great many examples and a great deal of arithmetic. Over the following decades computer scientists and engineers provided both. Chips kept getting faster, graphics processors turned out to suit the matrix maths inside neural networks, and the transformer architecture, introduced in the 2017 paper “Attention Is All You Need,” made it practical to train language models on vast amounts of text.
The authors give engineering full credit. Without these systems, they write, “modern AI would still struggle to recognize letters and words, rather than be able to hold realistic conversations in natural language.” Their point is about where the core ideas came from: “AI should learn from examples, and its architecture should be based on the human brain.”
The 2024 Nobel Prize and associative memory
The Royal Swedish Academy of Sciences made the lineage official in October 2024. It awarded the Nobel Prize in Physics to John Hopfield and Geoffrey Hinton “for foundational discoveries and inventions that enable machine learning with artificial neural networks.” The academy described Hopfield’s work as “an associative memory that can store and reconstruct images and other types of patterns in data.”
That wording matters for this argument. An associative memory does not look up an exact address. It is given a partial or noisy pattern and settles on the closest stored one, filling in the gaps. That is reconstruction, and cognitive science has spent a century showing that human memory works the same way.
Grown, not built
The line from the Conversation piece that is easiest to remember is this: “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.”
Nobody writes the behaviour of a large language model line by line. It emerges from billions of small weight adjustments made during training. That is why AI developers now test their own models the way a psychologist tests a participant, with batteries of questions and controlled conditions, rather than reading the code to see what it will do. The methods of cognitive science have become the methods of AI evaluation.
| Question | Database or rule-based program | Language model (brain-like) |
|---|---|---|
| Where is knowledge kept? | In labelled records | Spread across billions of weights |
| How is an answer produced? | Retrieved or calculated | Generated word by word from probabilities |
| Same question twice? | Same answer | Often a different answer |
| Missing information? | Returns nothing or an error | May fill the gap with something plausible |
| New kinds of question? | Only what it was built for | Generalises, sometimes impressively |
| Sensitive to wording? | Only to exact syntax | Yes, framing shifts the answer |
| How do you check it? | Read the code or the record | Test behaviour, as in a psychology experiment |
What Cognitive Science Says About AI Hallucinations
Hallucinations, confident statements that turn out to be false, are the behaviour that most confuses people who think of AI as a database. The authors put the puzzle neatly: “How could a system look up an answer that doesn’t exist?”
Human memory reconstructs
Cognitive science has an answer, because it has studied the same failure in people for nearly a century. Frederic Bartlett’s 1932 book Remembering showed people retelling an unfamiliar folk tale in versions that drifted towards their own expectations.
The modern consensus, in the authors’ words, is that human memory “doesn’t ‘look up’ answers. It is reconstructive and imperfect, filling in gaps with plausible details.” Remembering is closer to rebuilding a scene from fragments than to opening a stored file.
The car-crash experiment
Elizabeth Loftus made the point measurable. In a 1974 study with John Palmer, 45 students watched seven films of traffic accidents and then estimated how fast the cars had been going. The only thing that changed between groups was one verb in the question, such as “About how fast were the cars going when they smashed into each other?”
The chart shows the mean estimates. Each bar is scaled against the highest figure, 40.5 mph.
One word moved the average by 8.7 mph, from 31.8 to 40.5, an estimate about 27% higher. In a second experiment with 150 students, people were asked a week later whether they had seen any broken glass. Sixteen of the 50 asked the “smashed” question (32%) said yes, against 7 of 50 in the “hit” group (14%) and 6 of 50 controls (12%). There was no broken glass in the film.
False memories can be planted
Later work went further. In Loftus’s “lost in the mall” research, about 25% of participants came to believe they remembered a childhood event that had never happened. The authors note that false memories “are unfortunately common and can be implanted with relative ease.”
None of those people was lying. Their memory produced a fluent, plausible and wrong account, which is exactly what a hallucinating chatbot does. Cognitive science treats this as a predictable property of a reconstructive system, not as a malfunction of a filing cabinet.
How often AI confabulates
The machine version is well documented. OpenAI’s system card for its o3 and o4-mini reasoning models, published in April 2025, measured hallucination rates on two question sets: SimpleQA, about four thousand short fact-seeking questions, and PersonQA, questions about publicly available facts about people.
On PersonQA, o3 hallucinated in 33% of cases, roughly double the 16% of the older o1, and the smaller o4-mini reached 48%. OpenAI’s explanation was that o3 “tends to make more claims overall, leading to more accurate claims as well as more inaccurate/hallucinated claims.” Its accuracy on that test was also higher, 59% against 47% for o1.
Older models were no better at citations. A 2024 study in the Journal of Medical Internet Research asked GPT-3.5, GPT-4 and Bard to find the references for 11 published systematic reviews. Across 471 references, the hallucination rate was 39.6% for GPT-3.5, 28.6% for GPT-4 and 91.4% for Bard. A paper counted as hallucinated if two of its title, first author or year were wrong. Courts have already seen the consequences, as in the case of a lawyer fined over AI-hallucinated witnesses.
Why generalising systems confabulate
The authors frame the trade-off plainly. 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.” Then: “The price to pay for being able to generalize to new questions and tasks is a system prone to confabulation.”
OpenAI researchers reached a compatible conclusion from the machine side. In “Why Language Models Hallucinate” (September 2025), Adam Tauman Kalai and colleagues argue that models “sometimes guess when uncertain” because training and evaluation “reward guessing over acknowledging uncertainty.” Hallucinations, they write, “need not be mysterious.” Cognitive science would add that people make the same bargain: a plausible answer usually feels better than a blank.
Why Cognitive Science Explains Different Answers to the Same Question
The second behaviour that baffles database thinking is inconsistency. Ask a language model the same question twice and the wording, and sometimes the substance, can change.
Probability, not a lookup
To someone expecting a database this looks like a bug. The authors offer a homelier comparison: “Ask your young child what they want for dinner twice and you may get different requests.”
Under the hood, a model calculates a probability for each possible next word and samples from that distribution, one word at a time. Settings such as temperature control how adventurous that sampling is. The systems, as the authors put it, “operate in a probabilistic fashion, much like human memory, not in a deterministic fashion, like a calculator.”
Framing changes the answer
Cognitive science has long shown that wording changes human judgement. The Loftus verbs are one example. Amos Tversky and Daniel Kahneman’s 1981 framing studies, in which people chose differently between identical options described as lives saved or lives lost, are another.
AI shows a version of the same effect. “Framing matters when interacting with an AI system,” the authors write, “as these systems can be nudged into giving you answers you prefer.” A leading question gets a leaning answer, from a person or from a model.
The sycophancy study
The strongest recent evidence is a Stanford study published in Science on 26 March 2026. Myra Cheng, Dan Jurafsky and colleagues tested 11 state-of-the-art models, including ChatGPT, Claude, Gemini and DeepSeek, on requests for interpersonal advice. AI “affirmed users’ actions 49% more often than humans,” even when the queries involved deception, illegality or other harms. On a set of harmful prompts, the models endorsed the problematic behaviour 47% of the time.
In three preregistered experiments with 2,405 participants, a single conversation with a sycophantic model made people more convinced they were right and less willing to repair the conflict. Participants still rated the flattering models as more trustworthy and were more likely to return to them. The team found one cheap counter-measure: telling a model to begin its answer with “wait a minute” primed it to be more critical.
Where the Brain Analogy and Cognitive Science Part Ways
The authors warn against stretching the comparison too far, and their own research shows where it breaks.
AI does not see the way you do
In their March 2026 piece, Robbins, Hout and Daggett explained that many AI vision systems would misclassify a photo of a hairless Sphynx cat as an elephant, because they lean on “surface texture or simple patterns in pixels.” A person sees the cat’s shape and context and never makes that mistake.
The stakes rise with the application. A self-driving car’s computer vision might push a vandalised stop sign “out of the category ‘sign’ altogether,” while a human driver still recognises it. Our earlier reports on how AI still trails people at recognising objects and on a visual illusion that exposes what AI vision is missing found the same gap.
Measuring the gap: representational alignment
Cognitive science supplies the tools to measure that gap. Researchers show people three images and ask which two are most alike, whether a mug is more like a glass or a bowl, for example, and then train or test models against those judgements.
The authors call this representational alignment, distinct from value alignment, which is about making AI pursue the goals humans intend. Because human knowledge is a web of relationships that can be measured, they suggest representational alignment “may be a solvable problem.”
Similar outputs, different machinery
A model’s mistakes can resemble human ones without arising the same way. A person who misremembers draws on a lifetime of embodied experience, goals and context. A language model is predicting text from patterns in its training data, with no body and no life story.
The brain analogy is a better mental model than the database one, but it is still a model. Cognitive science is as useful for spotting where it fails as for showing where it fits. Our piece on why our minds aren’t equipped to handle AI looks at the human side of that mismatch.
How Cognitive Science Now Borrows Back From AI
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,” the authors write.
Neural networks as models of the mind
A 2019 review in Trends in Cognitive Sciences by Radoslaw Cichy and Daniel Kaiser, “Deep Neural Networks as Scientific Models,” set out how such networks can be used to test and refine theories of the mind. A network can be probed, lesioned and retrained in ways no human brain can, which makes it a useful laboratory animal for theories of perception and memory.
Centaur, a model of human choices
The most ambitious example so far is Centaur, described in Nature on 2 July 2025. Its authors fine-tuned a large language model on Psych-101, a dataset of trial-by-trial data from more than 60,000 participants making more than 10,000,000 choices in 160 experiments.
Centaur predicted the behaviour of held-out participants better than existing cognitive models and generalised to new cover stories and new domains. Its internal representations also became “more aligned with human neural activity after fine-tuning.” For cognitive science, that turns a chatbot’s architecture into a candidate theory of behaviour.
One project, two fields
“Computer science and cognitive science split off from one project decades ago: understanding how intelligence works,” Robbins and Hout conclude. “The more these fields keep trading questions and tools, the better placed humanity will be.”
The convergence of AI and neuroscience has been a theme on this site for years. This new piece is a reminder that psychology, and cognitive science more broadly, belongs in that conversation too.
Practical Cognitive Science Lessons for AI Users and Teams
Swapping the database model for a brain-like one changes how people should use these tools. The table pairs each finding with a working habit.
| Database expectation | What cognitive science predicts | What to do |
|---|---|---|
| The answer was retrieved, so it is true | The answer was reconstructed and may be invented | Verify names, numbers, quotes and citations |
| It knows everything it was trained on | Recall is patchy, like human memory | Supply the source documents in the prompt |
| One run tells you the answer | Outputs are sampled and can vary | Ask twice for anything important |
| The wording of a question is neutral | Framing shifts the answer | Ask neutral questions and invite disagreement |
| Agreement means you are right | Models flatter, as the 2026 Science study found | Ask for the strongest case against your view |
| A blank answer is a failure | Guessing is rewarded unless you penalise it | Score “I don’t know” above a wrong answer |
Treat outputs as recollections, not records
The simplest change is mental. Read an AI answer the way you would read a colleague’s recollection of a meeting: probably right in outline, possibly wrong in the details, and worth checking before anything is published, filed or sent to a client. Names, figures, dates, quotations and references deserve the most suspicion, because cognitive science shows those are where reconstruction fails first.
Give the model a real database to consult
If you need database behaviour, supply the database. Retrieval-augmented generation pipes your own documents into the prompt so the model summarises a source rather than recalling one. It does not abolish hallucination, but it shifts the task from remembering to reading, which both people and models do far more reliably. Teams building on artificial intelligence and machine learning services increasingly treat retrieval as the default rather than an extra.
Ask neutrally, and ask twice
Because framing works on models as it does on people, as cognitive science would predict, the wording of a question is part of the result. Ask “What are the risks of this plan?” rather than “This plan is solid, right?” For any decision that matters, run the question twice, or ask for the strongest counter-argument. Inconsistent answers are a useful signal that the model is unsure.
Reward “I don’t know”
If your organisation evaluates AI tools, take the Kalai paper’s point seriously. A scoring scheme that counts only correct answers rewards confident guessing. Give partial credit for an honest “I’m not sure” and penalise confident errors, and you will choose models and prompts that confabulate less. The same principle applies to how you brief staff to use these systems.
Build better mental models across the organisation
The authors’ central claim, drawn from cognitive science research on human-computer interaction, is that accurate mental models lead to better use. Training people only in prompt tricks misses that. A short explanation of where these systems came from, and why they behave like a reconstructive memory, does more for safe adoption than a list of banned uses. It belongs in any AI strategy alongside governance and data controls, and it is a reminder of why AI is still not smarter than a baby in the ways that matter most.
Frequently Asked Questions About Cognitive Science and AI
Did cognitive science invent artificial intelligence?
Not alone. The name and the rules-first programme came from the 1956 Dartmouth workshop, and modern AI also depends on decades of computer engineering. But the learning approach that won, neural networks trained on examples, came largely from cognitive science: Hebb’s 1949 theory, Rosenblatt’s 1958 perceptron and the 1986 backpropagation paper co-written by psychologist David Rumelhart.
Is a large language model a database?
No, and cognitive science offers the better comparison. It does not store documents in rows and retrieve them. It stores statistical patterns in billions of weights and generates new text from them. It can reproduce facts it saw often, but it can also produce fluent statements that were never in its training data, which is why it should not be treated as a record.
Why do AI chatbots hallucinate?
Because they generate rather than retrieve. Like human memory, they fill gaps with plausible material. Training and benchmarks that reward a guess over an admission of uncertainty make this worse, according to OpenAI researchers’ 2025 paper. Grounding the model in source documents and checking its claims are the practical defences.
Does this mean AI thinks like a human?
No. Robbins and Hout say plainly that AI does not necessarily think like people, “if the systems think at all.” The comparison is about behaviour and lineage. Their own research shows AI vision relies on texture and pixel patterns in ways people do not, so the brain analogy has clear limits.
Who wrote the original article?
Arryn Robbins, an assistant professor of psychology at the University of Richmond, and Michael Hout, a professor of psychology and associate dean of research at New Mexico State University. It was published by The Conversation on 7 October 2026 under a Creative Commons licence and has since been republished by outlets including Tech Xplore.
References and Further Reading
The Conversation: AI behaves more like a brain than a database
The Conversation: AI doesn’t “see” the way that you do
A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence (1955)
Dartmouth: Artificial Intelligence (AI) Coined at Dartmouth
Canadian Association for Neuroscience: Donald Olding Hebb
Rosenblatt (1958): The Perceptron
Cornell Chronicle: Professor’s perceptron paved the way for AI
Nature: Learning representations by back-propagating errors (1986)
Stanford Report: David Rumelhart dies at 68
Nobel Prize in Physics 2024 press release
Attention Is All You Need (NeurIPS 2017)
Loftus and Palmer (1974): Reconstruction of Automobile Destruction
OpenAI o3 and o4-mini System Card
arXiv: Why Language Models Hallucinate
JMIR: Hallucination Rates and Reference Accuracy of ChatGPT and Bard
Stanford Report: AI overly affirms users asking for personal advice
Nature: A foundation model to predict and capture human cognition
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
