AI Is a 70-Year-Old Marketing Term

On August 31, 1955, four researchers sent a one-page proposal to the Rockefeller Foundation. John McCarthy at Dartmouth, Marvin Minsky at Harvard, Nathaniel Rochester at IBM, and Claude Shannon at Bell Labs. They wanted funding for a two-month summer workshop. The proposal said the workshop would proceed on the basis of the conjecture 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."

That document is where the phrase "artificial intelligence" first appeared in print.

It was a marketing choice. McCarthy picked the name over alternatives in circulation at the time. "Cybernetics" was one. "Complex information processing" was another. He wanted something that would differentiate the group and pull in recruits. It worked. The 1956 Dartmouth Summer Research Project is treated as the founding event of the field. The name stuck.

The problem is what the name stuck to.

The Umbrella Covers Things That Have Almost Nothing in Common

Since 1955, the term has covered symbolic logic systems, search algorithms, theorem provers, expert systems, planning systems, knowledge graphs, machine learning models, neural networks, and now large language models.

A rule-based expert system from the 1980s and a 2024 transformer model have roughly the same relationship to one another as a typewriter and a smartphone. Both produce text. Almost nothing else is shared. Different math. Different hardware. Different failure modes. Different reasons they break.

When someone says "AI can do X," the useful question is which AI. The one that plays chess by searching billions of board positions? The one that classifies an X-ray? The one that predicts the next word in a sentence? Those are three different technologies wearing the same coat.

Russell and Norvig Tried to Organize the Mess

Stuart Russell and Peter Norvig wrote the standard textbook on this field. The fourth edition came out in 2020. They organize AI along two axes.

One axis runs from thinking to acting. The other runs from human-like to rational. That produces four research traditions: systems that think like humans, systems that think rationally, systems that act like humans, and systems that act rationally.

Most modern AI research has converged on the last one. The design of "rational agents" that take actions to maximize an objective inside an environment.

That definition is broad enough to cover a chess engine, an autonomous driving stack, and a chatbot. Which is useful for academics. It's also a warning sign about how much work the word is doing.

The Field Has Crashed Before

The current excitement isn't the first one.

In 1973, Sir James Lighthill delivered a report to the UK Science Research Council. He'd been asked to evaluate the state of AI research. His conclusion was that the field had failed to deliver on its promises. He pointed to what he called the "combinatorial explosion" problem in search. The math got too big too fast. Real-world problems blew up the algorithms.

The Lighthill report caused major funding cuts in the UK and contributed to a broader contraction in the US. That period got named the first AI winter.

There was a second winter in the late 1980s and early 1990s. Expert systems had been the hot product. Companies installed them. They were brittle. They couldn't handle situations that fell outside their hand-coded rules. The market collapsed.

The third wave, the current one, dates from a specific moment. In 2012, Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton at the University of Toronto entered a model called AlexNet into the ImageNet image classification competition. They cut top-5 error from about 26 percent to about 15 percent in a single year. That was the result that made deep neural networks mainstream. Everything you currently call "AI" descends from that win.

How to Read News About AI

If a news article says "AI" without saying which technique, the article is doing you a disservice.

A decision tree trained on tabular insurance data is AI. A convolutional neural network that segments tumors in MRIs is AI. A transformer-based language model that drafts emails is AI. Those three systems don't share architecture, training method, failure mode, or honesty about what they're doing.

The headlines treat all three the same. The vendors selling them treat all three the same. That's how you end up with a hospital buying a chatbot to read X-rays, or a company shipping a customer service bot that hallucinates refund policies.

The rest of this series takes the umbrella apart. Machine learning. Neural networks. Transformers. Large language models. Each one has its own history, its own math, and its own list of things it can't do. The word "AI" by itself is no longer useful for thinking clearly about any of them.

Part 1 of the AI Foundations series.

Sources

• McCarthy, John, Marvin Minsky, Nathaniel Rochester, and Claude Shannon. "A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence" (1955, reprinted in AI Magazine, 2006) (opens in new tab)
• Russell, Stuart, and Peter Norvig. "Artificial Intelligence: A Modern Approach" (4th edition, Pearson, 2020) (opens in new tab)
• Lighthill, James. "Artificial Intelligence: A General Survey" (1973, UK Science Research Council) (opens in new tab)
• Krizhevsky, Alex, Ilya Sutskever, and Geoffrey Hinton. "ImageNet Classification with Deep Convolutional Neural Networks" (2012, NeurIPS) (opens in new tab)
• Stanford Institute for Human-Centered AI. "2025 AI Index Report" (April 2025) (opens in new tab)