What "AI" actually means

The single most useful thing to understand about "AI" is that it is not one technology. It is a roughly seventy-year-old umbrella term that has covered a long succession of very different approaches to making machines do things that seem to require intelligence. The phrase was coined in 1956 at a summer workshop at Dartmouth College, where the mathematician John McCarthy and a small group of colleagues proposed studying how to make machines "use language, form abstractions and concepts, solve kinds of problems now reserved for humans." That ambition has stayed constant for seven decades, but the methods underneath it have been replaced several times over. When you read that "AI did X" in two different headlines, the word can refer to two systems with almost nothing in common under the hood. This is the root cause of most executive confusion about the field, and it is worth holding onto before anything else.

Today the umbrella covers at least three layers that are commonly all called "AI." The oldest surviving layer is rule-based systems: software where humans wrote out explicit if-then logic, the dominant approach from roughly the 1960s into the 1990s. Your fraud flags, tax software, and many "smart" workflows are still this. The second layer is machine learning, where instead of writing rules a system learns patterns from historical data, which is what most quietly effective enterprise "AI" has been for two decades: the model behind credit scoring, demand forecasting, churn prediction, and recommendation engines. The newest and loudest layer is generative AI, the technology behind tools like ChatGPT that produce fluent text, images, and code on demand. Generative AI is a subset of machine learning, which is a subset of the broader AI field. The chatbot that has dominated headlines since late 2022 is one recent layer inside a much older and wider category, not the whole of it.

Why this matters for an organization is practical, not academic. When a vendor, a board member, or a regulator says "AI," the risks, costs, and capabilities differ enormously by layer, and conflating them leads to bad decisions. A rule-based system does exactly what it was told and fails predictably; a machine-learning forecasting model can quietly degrade as the world changes (the subject of the drift concept) but is otherwise stable and auditable; a generative model can produce confident, fluent output that is simply wrong (the subject of the hallucination concept) and behaves differently each time you run it. Budgeting, procurement, and governance all change depending on which one you actually have. A useful first question in almost any AI conversation is therefore not "is this AI?" but "which kind, and what is it actually doing?"

It also helps to know that the field has been here before. AI has gone through at least two well-documented boom-and-bust cycles, known as "AI winters." In the 1970s, early optimism about machine translation and general problem-solving collided with the messy reality of language and perception, and government funding collapsed. In the late 1980s a commercial wave built on "expert systems" (software encoding the knowledge of human specialists as rules) failed to scale, and a second winter followed. Each cycle featured genuine advances oversold as imminent general intelligence. The current wave is built on real, measurable capability that the earlier waves lacked, but the historical pattern is a reason to separate what a system demonstrably does today from what it is projected to do, rather than a reason to dismiss the technology. Both overstatement and reflexive cynicism have been expensive postures in this field's history.

The current wave has a specific origin worth knowing. Around 2012, a type of machine learning called deep learning, loosely inspired by networks of neurons, suddenly outperformed everything else at recognizing images, after decades of being a research backwater. The same family of methods, scaled up dramatically and applied to text, produced the generative systems of the past few years. So when people say AI has "suddenly" gotten good, they are usually pointing at this one technical lineage maturing, not at the entire seventy-year field changing character. Most of the older layers are still running, unglamorously, all around you — the older a capability is, the less likely anyone still calls it AI at all.

That last point names a quirk worth recognizing, sometimes called the "AI effect": once a task is reliably solved, people tend to stop calling it intelligence and start calling it ordinary software. Optical character recognition, spam filtering, route planning, and chess engines were all once headline AI; now they are just features. The computer scientist Larry Tesler captured this as the only half-joking definition that "AI is whatever hasn't been done yet." The practical consequence for a leader is that the label "AI" drifts toward whatever is newest and least proven, which systematically inflates the sense that AI is more experimental and less dependable than it is. A great deal of genuinely reliable machine intelligence has quietly become invisible by working, while the word itself keeps migrating to the frontier. When you hear "AI," it is worth asking whether the speaker means a mature, boring, dependable capability or a bleeding-edge one, because the term gives you no help in telling them apart.

For an executive, the working takeaway is a vocabulary and a set of questions, not a verdict. "AI" is a category, like "transportation" or "chemicals" — useful as shorthand, useless for a specific decision. When the term appears in a strategy deck, a contract, or a news story, the productive move is to translate it into the specific layer in play: Is this rules, learned prediction, or generation? Is the claim about something it does now or something projected? Does the output need to be exactly right, or merely useful? Organizations that build this habit tend to avoid both the panic and the over-investment that the word "AI," used loosely, reliably produces. The rest of this curriculum unpacks each layer in turn, and almost every later concept assumes you are carrying this distinction with you.