First article in a three-part series on artificial intelligence: definitions, professional applications, and governance.
The first time a clustering algorithm I built flagged confirmed cases of supplier collusion, I didn’t feel like I was looking at an “intelligent” machine in the way we usually imagine one. No human reasoning, no intuition — just data points grouped according to statistical patterns invisible to the naked eye, yet pointing to something very real. That moment made me realise artificial intelligence doesn’t need to “think” like us to produce useful, sometimes more reliable, results than our own judgment.
That’s not entirely a coincidence: the term “artificial intelligence” itself, coined in 1956 at a research workshop at Dartmouth College, has always covered two rather different ambitions — imitating humans, or simply making good decisions. So what is AI, really?
A simple definition, a complex field
Artificial intelligence can be defined as an artificial entity capable of reproducing certain aspects of human intelligence: learning, perceiving, reasoning, and making decisions in new situations.
Here’s a good example of how misleading that definition can be: as early as 1966, a program called ELIZA, built at MIT, simulated a conversation with a psychotherapist simply by rephrasing the user’s statements as questions. Some users grew genuinely emotionally attached to it — even though the program was, in reality, doing nothing more than keyword matching, with no real “understanding” involved. That episode shows, 60 years before ChatGPT, our natural tendency to project intelligence onto a machine that merely gives the illusion of having it.
Two ways of thinking about artificial intelligence
Historically, two major visions have competed in AI research:
- Human imitation: designing machines that think and act like a human being — the ambition already behind ELIZA, and the one found today in conversational models like ChatGPT.
- Rationality: building systems that make the best possible decisions, whether or not they resemble human reasoning — the approach illustrated in 2016 by AlphaGo, which beat world Go champion Lee Sedol by playing moves no human would have considered, without ever trying to “think” like a human player. It’s the same logic behind my own collusion-detection model: no imitation of an auditor’s reasoning, just a rational search for patterns in the data.
These two visions branch into four concrete approaches:
| Approach | What it aims for | Concrete example |
|---|---|---|
| Acting humanly | Reproducing observable human behaviour — the machine is judged on its ability to appear human in the exchange, as in the Turing Test | ChatGPT (OpenAI) — a conversational model capable of simulating a credible human dialogue |
| Thinking humanly | Faithfully modelling the human cognitive process itself, not just its outcome | General Problem Solver (Newell & Simon, 1957) — a program designed to solve problems by reproducing, step by step, the process observed in human subjects asked to “think aloud” |
| Thinking rationally | Following logical rules to reach a correct conclusion according to a reasoning model, without trying to imitate human intuition | MYCIN — a 1970s expert system that recommended antibiotic treatments based on hand-coded medical rules |
| Acting rationally | Choosing the action that best achieves a goal, given available information and uncertainty | AlphaGo (DeepMind) — beat the world Go champion by playing moves no human player would have considered |
These distinctions aren’t just theoretical: they explain why two very different AI systems — a conversational chatbot and a fraud-detection algorithm, for instance — can both lay claim to the term “artificial intelligence” while pursuing radically different objectives.
The Turing Test: an outdated but foundational benchmark
No discussion of AI is complete without the Turing Test, proposed by Alan Turing in 1950 — six years before the term “artificial intelligence” even existed. The idea: a machine can be considered “thinking” if it manages to convince a human they’re conversing with another person.
ELIZA fooled some users in the 1960s with a handful of lines of code. Today, ChatGPT passes this test in the vast majority of short exchanges. But appearing intelligent is no longer a meaningful benchmark: a language model can hold a fluent conversation without actually reasoning, verifying facts, or making a reliable decision in a high-stakes context — a medical diagnosis, a credit decision, a financial audit.
When AI becomes a rational agent
It’s precisely this limitation that pushed research toward a more demanding notion: rationality. Rather than trying to imitate humans, much of modern AI research aims to build rational agents: systems capable of perceiving their environment, interpreting a request, and acting to achieve a defined goal — while accounting for constraints and uncertainty.
Today’s AI agents illustrate this well: given an instruction, such an agent interprets the context, calls on external tools if needed, chains together multiple reasoning steps, observes the result, and adjusts its action accordingly — never trying to “look human” in the process, only to reach the stated goal as effectively as possible.
This ability to “make the right choice” rather than “look human” fundamentally changes the kind of questions we should be asking about AI.
Thinking rationally is not the same as acting rationally
These last two notions are worth pausing on, because the distinction between them is one of the most useful for understanding modern AI.
Thinking rationally means correctly applying a logic or reasoning model to reach a conclusion — much like a rule-based system deducing “if A and B, then C.” Acting rationally is something else: it means choosing the best possible action given what is known, what isn’t, and the constraints of the situation — which is exactly what the kind of AI agent described above does.
This distinction has an important consequence: a rational agent is neither omniscient nor infallible. It can make a perfectly rational decision given the information available to it, and still get it wrong — because its data is incomplete, because its model of the world is imperfect, because its environment has changed since training, or because its objective was poorly specified in the first place.
This is a crucial point, and one that’s often misunderstood: an AI system that reasons “rationally” and still gets it wrong isn’t necessarily a broken system. It may simply have reasoned well from an incomplete picture of the world. This is precisely what makes the question of an AI’s decision reliability — and therefore its governance — so central.
Why this distinction matters
The question is no longer just “does the machine seem intelligent?” but “are its decisions reliable, justifiable, and aligned with our goals?”
This shift — from imitation to rationality, and from rationality to the question of reliability — explains why AI now plays an increasingly large role in high-stakes decisions: medical diagnoses, fraud detection, financial recommendations, autonomous driving. And it’s this very question that, as soon as an AI system starts making decisions in a professional context, leads directly into questions of risk and governance.
In the next article of this series, we move from theory to the field: how artificial intelligence is transforming functions like audit, compliance, and fraud detection — with a concrete example drawn from my own work as an internal auditor.
What’s your take — which vision of AI feels closer to the truth: imitation or rationality? Let me know in the comments.