History of Artificial Intelligence

Learning outcomes

After studying this lesson, you will be able to:

• explain the mathematical and computational foundations of AI;
• distinguish early precursors from the formal emergence of the field;
• compare symbolic, expert-system, statistical, and neural approaches;
• explain why AI winters occurred;
• connect major breakthroughs to data, algorithms, and computing.

1. Before AI had a name: logic and computation

Mathematical logic made it possible to describe inference through formal rules; computability theory clarified the limits of algorithmic procedures; and electronics enabled general-purpose programmable machines. In 1943, Warren McCulloch and Walter Pitts published a mathematical model of an artificial neuron. It was an abstract logical model, not a copy of the brain: it showed how networks of simple elements could implement logical operations.

Cybernetics, information theory, computing, and learning research developed in parallel. AI emerged from the convergence of several scientific lines, not from a single inventor.

2. 1950: Turing’s question

In Computing Machinery and Intelligence, Alan Turing reframed “Can machines think?” through the imitation game. In its popular form, the Turing Test asks whether an interlocutor can reliably distinguish a machine from a human in text dialogue. It is an operational test of a particular behaviour, not a universal measure of intelligence or a test of consciousness.

A system may sustain a convincing conversation without demonstrating human understanding, intentions, or subjective experience. This distinction remains essential when evaluating modern language models.

3. 1955–1956: Dartmouth and the name Artificial Intelligence

In August 1955, John McCarthy, Marvin Minsky, Nathan Rochester, and Claude Shannon proposed a research project at Dartmouth College for the summer of 1956. The proposal used the phrase artificial intelligence and listed language, abstraction, problem solving, and self-improving programs among its goals.

The Dartmouth workshop became a symbolic milestone in the formation of AI as a distinct research field. Important work on logic, neuron models, and computation already existed; the participants did not share one unified theory.

4. The 1950s–1960s: symbolic AI

Early programs operated on symbols, concepts, logical expressions, and rules. Logic Theorist, by Newell, Simon, and Shaw, demonstrated automated mathematical theorem proving. General Problem Solver explored problem solving through goals and operators. These systems showed that parts of reasoning could be represented as search through possible steps.

In the mid-1960s, Joseph Weizenbaum created ELIZA, a dialogue program based on patterns and phrase transformations. User reactions showed how readily people attribute understanding to a system with a relatively simple mechanism.

5. The 1970s–1990s: limitations and AI winters

Demonstrations often worked in narrow settings. Real-world use exposed combinatorial explosion, limited compute, language ambiguity, brittle rules, and the difficulty of formalizing knowledge. Promises of imminent general intelligence did not match the results.

AI winter refers to periods of reduced funding and interest after expectations exceeded practical results. Commonly discussed downturns occurred in the 1970s and again in the late 1980s to early 1990s. Research did not stop completely; priorities and funding sources changed.

Expert systems

During the 1970s and 1980s, expert systems encoded specialist knowledge as facts and rules, then used an inference engine to apply them to cases. They were useful in bounded domains but expensive to maintain: rules had to be elicited, reconciled, updated, and tested. Outdated knowledge could produce incorrect conclusions.

6. The 1980s–2000s: statistical machine learning

Rather than manually writing every rule, researchers increasingly used models whose parameters were estimated from examples. Probabilistic methods, statistical pattern recognition, decision trees, support vector machines, and ensembles developed. Developers specified the task, features, and objective; parameters were estimated from data.

Domain expertise remained essential: data quality had to be checked, target leakage prevented, and performance evaluated on independent samples.

7. 1997–2016: specialized victories and deep learning

In 1997, IBM’s Deep Blue defeated world chess champion Garry Kasparov. This was a specialized achievement in search and position evaluation, not proof of general intelligence. In 2012, AlexNet, a deep convolutional network, achieved a landmark result on ImageNet. Its success combined architecture, labelled data, GPUs, and effective training. In 2016, AlphaGo defeated Lee Sedol by combining neural networks, search, and reinforcement learning.

8. 2017–2020s: Transformers and generative AI

In 2017, Attention Is All You Need introduced the Transformer architecture based on attention. It became foundational for many modern language models. Large-scale pretraining, instruction tuning, and multimodal models followed.

The public launch of ChatGPT in November 2022 made conversational access to a large language model a mass phenomenon. Image, audio, and video generation, search tools, and multi-step workflows then developed rapidly. Generative systems can still make mistakes and confidently produce false information.

9. Why history moves in waves

1. A new idea proposes a way to represent knowledge or train a model.
2. A prototype works on a controlled task.
3. Scaling exposes limitations in data, compute, and methods.
4. Expectations are revised; funding and interest may decline.
5. New data, computing resources, and algorithms make the task practical again.

Earlier approaches do not disappear: rules, search, statistics, and neural networks are used in parallel. Real systems are often hybrid.

10. Practical task