What AI Actually Is

Imagine someone asks you to explain the difference between a bicycle and a jet airplane. They are both vehicles, but you would never govern them the same way. One requires a helmet; the other requires air traffic control. Before we can build sensible rules for artificial intelligence, we need to answer a deceptively simple question: What is AI, actually?

Here is the key insight: AI is not a single technology but a broad category, and the boundaries are fuzzy. At its core, the most useful definition for a governance leader comes from the OECDOECD: an AI system is a machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations, or decisions. The critical word here is infers. Unlike traditional software that follows explicit instructions, AI identifies patterns from data and uses them to make judgments about new situations.

This distinction matters enormously for governance. Think about your company's payroll system. It follows precise rules: if hours worked equals 40, then pay equals hourly rate times 40. There is no guessing, no inference. This is traditional automation.

Now consider a system that reviews job applications to predict which candidates will succeed. No human wrote explicit rules for "success." Instead, the system analyzed thousands of past employees and inferred its own patterns. This is AI. And because those inferred patterns can include biases invisible to the developers, governing AI requires fundamentally different tools than governing a spreadsheet formula.

To understand this distinction more deeply, consider the calculator versus mathematiciancalculator versus mathematician analogy: a calculator executes predefined operations perfectly, while a mathematician reasons through novel problems. Traditional software is the calculator; AI aspires to be the mathematician. The line between simple automation and true AI can be explored further in our discussion of RPA vs AIRPA vs AI.

Here is something many technologists miss: an AI system is never just code. It is a socio-technical system—a combination of algorithms, data, hardware, and the humans who build, deploy, and interact with it. When an AI system fails, the failure is rarely just a software bug. It is usually a breakdown in this larger system.

For example, an AI that denies someone a loan is not making that decision in a vacuum. A data team chose what information to collect. Engineers decided how to measure "creditworthiness." Business leaders set the threshold for approval. Users interpret the output and act on it. Effective governance must address this entire chain, not just the model at its center.

This is the concept of a socio-technical failuresocio-technical failure. The technical component may work exactly as designed, yet the overall system produces harmful outcomes because of misalignments between the technology and the human context it operates in.

The term "artificial intelligence" was coined in 1956 at a now-famous workshop at Dartmouth College. The Dartmouth WorkshopDartmouth Workshop brought together researchers who believed they could, in a single summer, make significant progress on making machines "think." That optimism was premature, but it launched a field that has, after decades of boom and bust, finally begun to deliver on some of its early promises.

Understanding this history helps you calibrate expectations. AI has been "five years away" from human-level intelligence for over 60 years. While recent breakthroughs are remarkable, a healthy skepticism about overpromises will serve any governance leader well.

Returning to our bicycle versus jet analogy: the governance tools you need depend on what you are governing. If you label every piece of software "AI," you will waste resources auditing your email filters. If you fail to recognize AI in a hiring tool, you may expose your company to significant legal and ethical risk.

The working definition we use throughout this learning path is practical: AI is a system that learns patterns from data to make inferences about new situations. If a system is learning and inferring, it needs AI governance. If it is just following rules, standard IT controls may suffice. Now that we have established what AI is, we can explore the different branches of this family tree. In the next article, we examine The AI Family Tree, tracing how different approaches to building intelligence have evolved over time.