IntelligenziaIntelligenzia

AI and Cognitive Science

The connection between cognitive science and artificial intelligence from history to today. How the human mind and machine learning relate to each other.

Artificial intelligence and cognitive science were born at the same moment. In 1956—the same year the cognitive revolution began—AI was established as a discipline at the Dartmouth conference. These two fields have since developed in parallel, feeding each other.

Cognitive science studies natural intelligence. AI aims to build artificial intelligence. Both ask the same fundamental question: what is intelligence and how does it work?

Shared history

1950s: Birth

AI's founders—John McCarthy, Marvin Minsky, Allen Newell, Herbert Simon—were also central figures in the cognitive revolution. They believed that by building programs simulating intelligence, we could better understand how the human mind works.

Newell and Simon developed Logic Theorist (1956), a program that proved mathematical theorems. This was the first program to do something that looked like intelligent problem-solving.

1960s–70s: Symbolic AI

Early AI was based on symbolic information processing: explicit rules, logical operations, data structures. The idea was that intelligence is fundamentally symbol manipulation—and brains are computer-like symbol processors.

This approach produced expert systems that modeled human knowledge with rules and facts. They succeeded in narrow domains but failed at broad, everyday tasks.

1980s: Connectionism

The 1980s saw an alternative view emerge: perhaps intelligence isn't based on symbolic computation but on neural networks. Rumelhart and McClelland introduced connectionism, which modeled cognition as parallel, distributed processes.

Neural networks learned from examples and didn't require explicit rules. This more closely resembled how children learn.

2010s: Deep learning revolution

2012 marked a turning point. Deep learning—multi-layered neural networks—began achieving human-level performance in image recognition, speech recognition, and language processing.

This didn't happen through cognitive science inspiration but through massive data and computing power. Still, deep learning's success rekindled questions about what human and machine intelligence have in common.

2020s: Large language models

GPT, Claude, and other large language models have revolutionized AI and raised cognitive questions to a new level:

  • Do language models understand language or merely mimic statistical patterns?
  • Do they have something significant from a cognitive science perspective—or are they just very effective pattern matchers?
  • What do they tell us about human language ability?

AI as a tool for cognitive science

AI is more than a research subject—it's a tool for cognitive science research:

Computational models

Cognitive science uses computational models to test theories about how the mind works. If a theory can be formulated as a program, its predictions can be compared to human behavior.

Examples:

  • ACT-R – John Anderson's cognitive architecture modeling memory, learning, and problem-solving
  • Bayesian models – Describe how the mind might reason under uncertainty
  • Reinforcement learning models – Explain how we learn from rewards and punishments

Modeling the brain

Neural network models were originally inspired by the brain. Today, neuroscience uses them to model how the brain processes information. For example, in visual perception research, convolutional neural networks have proven a good model for how the brain's visual system works.

Experimental control

AI enables controlled experiments in ways not possible with humans. We can, for example, train a language model in a specific way and study how its "cognition" develops—without ethical constraints.

Cognitive science as a tool for AI

The connection works both ways. Cognitive science offers AI:

Inspiration for architectures

Neural networks were inspired by brain structure. Newer developments like memory systems (in transformers) and attention mechanisms have drawn from cognitive research.

Understanding human cognition

For AI to work with humans, it must understand how humans think. Cognitive science provides models of human decision-making, learning, and communication.

Evaluation criteria

How do we evaluate AI's intelligence? Cognitive science provides frameworks for assessing what AI actually can do—not just whether it passes benchmark tests.

AI and human intelligence: differences and similarities

Learning

Human: Learns from few examples, generalizes flexibly to new situations AI: Typically requires vast amounts of data, can be brittle in new situations

Generalization

Human: Smoothly transfers knowledge between domains AI: Often works well only in its training domain

Physical understanding

Human: Intuitive understanding of physics, social relations, everyday life AI: Struggles with "common sense" reasoning

Energy efficiency

Human: The brain consumes about 20 watts AI: Large models require megawatts for training

Language use

Human: Learns language in natural context, understands reference relations AI: Language models produce fluent text but their "understanding" is controversial

AI in Finland and cognitive science

Finland is a significant player in both AI and cognitive research:

FCAI (Finnish Center for Artificial Intelligence) – A joint research center of Aalto University and University of Helsinki studying AI from a human-centric perspective.

Bayesian cognition – Finnish researchers specialize in the Bayesian approach, which combines cognitive science and machine learning.

Human-AI interaction – Research on how humans and AI systems can collaborate effectively.

Open questions

At the intersection of AI and cognitive science are many open questions:

  • Can a machine understand? – What does understanding even mean?
  • Consciousness and AI – Can a machine be conscious?
  • Limits of learning – Why do humans learn from so little, AI from so much?
  • Ethical AI – How do we build AI that acts according to human values?

Read more