Understanding Artificial Intelligence

Can machines really think? That question has haunted philosophers, scientists, and ordinary people alike since the dawn of the computer age. In Understanding Artificial Intelligence, Nicolas Sabouret argues that artificial intelligence (AI) is not magic, nor is it about building machines that think like us—it’s about creating tools that help us think and act more effectively. He explains that AI is fundamentally a human invention: it encodes our logic, creativity, and limitations into machines that can perform specific tasks efficiently. What distinguishes AI from earlier tools isn’t autonomy—it’s adaptability.

Sabouret invites readers on a journey through history and methodology to unpack how these machines are built, how they ‘learn,’ and what that word actually means. Rather than succumbing to the myths perpetuated by Hollywood or alarmist media, he grounds AI in computer science and real-world examples—GPS systems, chess programs, self-driving cars, and even online recommendation engines.

Demystifying Intelligence Itself

At its core, the book argues that what we call ‘intelligence’—human or artificial—is not a single thing. We mistake AI’s computational speed and consistency for intelligence, but Sabouret shows that intelligence involves memory, learning, reasoning, and perception working together in organic, not mechanical, ways. He references historical foundations, such as Alan Turing’s theoretical models and his now-famous ‘Turing Test,’ emphasizing that even Turing never claimed machines truly understood—they simply performed well enough to appear intelligent under narrow conditions.

This distinction between simulation and understanding is central. Machines operate on algorithms: structured sequences of instructions written by humans. Intelligence, by contrast, emerges in living beings through experiences and context. In other words, computers obey syntax but lack semantics—they can apply rules but not ‘mean’ anything by them (a point echoing John Searle’s ‘Chinese Room’ thought experiment).

From Calculators to Cognitive Tools

Sabouret traces AI’s history from mechanical calculators and Charles Babbage’s analytical engine to modern neural networks. Computing began as a way to mechanize mathematics, but as programs became capable of writing new programs (machine learning), we entered a new stage: computers started to generate patterns based on data instead of explicit instruction. This transition, he explains, mirrors human learning in structure but not in substance—it’s optimization, not imagination.

He also situates AI’s modern success stories, like Google’s AlphaGo or deep learning image recognition, within this lineage. These aren’t signs of machines achieving consciousness, but of algorithms refined by enormous increases in data and processing power (via GPUs). The lesson is that every AI innovation still rests atop centuries-old computational logic.

The Three Pillars: Data, Algorithms, and Human Intent

Throughout the book, Sabouret returns to one simple triad: AI is driven by data (examples), algorithms (rules), and design (human choices). These elements define all AI—from rule-based systems of the 1950s to modern machine learning. If the data are biased, or the algorithm incomplete, the resulting ‘intelligence’ mirrors those flaws. The supposed magic of learning machines is, ultimately, a reflection of the human programmers who created them.

This perspective also clarifies why AI’s limitations are not failures but natural boundaries of its design. A chess-playing program can’t drive a car. A medical classifier can’t write poetry. Each model excels narrowly because it focuses computational power where human intuition falls short. That, Sabouret insists, is what makes AI valuable: its ability to complement human intelligence, not replace it.

Why This Matters to You

AI is changing the world: from the way we navigate cities (A* algorithms in GPS) to how we shop, diagnose disease, and interact with media. Yet misunderstanding it breeds both unwarranted fear and misplaced awe. Sabouret’s conversational tone reminds readers that AI’s future depends less on superintelligence and more on our ability to use ethical design, interpret outputs responsibly, and prevent misuse. Like the printing press or industrial machinery before it, AI is transformative not because it’s alive, but because it amplifies our reach.

“Artificial intelligence is simply a tool—one of the best we’ve ever invented—but it has no will of its own.”

By the book’s end, you realize that understanding AI means understanding ourselves. Every algorithm whispers back our assumptions about intelligence, efficiency, and control. Sabouret’s real message is less about machines and more about human agency: if AI appears intelligent, it’s because we’ve encoded a fragment of our own thinking into it. What we do with that fragment determines whether AI becomes a creative partner—or just another mirror reflecting our limitations.

To understand AI, you must first understand what a computer does and doesn’t do. Sabouret begins here: a computer is a symbol processor, a machine that obeys instructions. The real magic lies not in hardware but in algorithms—the step-by-step recipes telling machines how to react to input data. Just as a cook follows a recipe, an algorithm follows logical rules. When these rules are encoded into programmable machines, they transform raw electricity into structured action.

From Babbage to Turing

Sabouret revisits Charles Babbage’s 19th-century idea of a ‘universal machine,’ which could perform any calculation following proper instructions. This idea was refined in 1936 by Alan Turing, whose conceptual ‘Turing Machine’ could read and write symbols on a tape according to simple rules. That abstract model remains the foundation of all computing devices. Whether your smartphone or a supercomputer, each performs millions of Turing-like operations per second—never thinking, always calculating.

AI, then, is not about imbuing machines with mystical intelligence, but designing algorithms that perform tasks humans find intellectually demanding: playing chess, understanding speech, classifying images. These algorithms execute faster than humans but depend entirely on the instructions they’re given.

Machine Learning as a Second-Order Algorithm

Writing an algorithm for every single task is tedious. So computer scientists developed programs that could write other programs—what we now call machine learning. Instead of specifying every rule, we supply examples and let the system infer statistical patterns. As Sabouret notes, this doesn’t mean machines learn independently; it means they automatically generate computational shortcuts based on data. Learning here means optimizing equations, not developing self-awareness.

A powerful example is Google’s AlphaGo. Behind its apparent intuition lies nested layers of algorithms: one to analyze board positions, another to evaluate moves, and another to refine its own predictions based on game results. Years of human programming and mountains of data made its victory possible—not spontaneous intelligence.

The Myth of Self-Modifying Machines

One of Sabouret’s central clarifications is that computers cannot truly ‘rewrite themselves’ in a creative sense. Every layer of learning still operates within human-defined constraints. A learning algorithm can adjust parameters, but it cannot decide new goals. Just as a child cannot reinvent the concept of arithmetic while learning to add, a machine cannot escape its architecture.

When people imagine machines designing other AIs without guidance, they misunderstand recursion. Sabouret humorously reminds readers that a race car can beat human sprinters but can’t bake bread. Each program is domain-specific—the illusion of general intelligence comes only from stringing together many narrow solutions written by different specialists.

“An AI program using machine learning can only produce a ‘good’ program if we give it good data and good instructions. None of this works all on its own.”

Why This Matters

Understanding algorithms isn’t just technical curiosity—it’s a form of literacy. When you grasp that all AI boils down to structured recipes, you can see through hype and evaluate real progress. It lets you appreciate that Siri’s voice recognition or Netflix’s recommendation system are built on the same logic. Sabouret’s takeaway is empowering: AI doesn’t mystify humanity’s future; it illuminates how far our formal reasoning has taken us—and where human creativity still leads the way.

What does it even mean for a machine to be ‘intelligent’? Sabouret calls this a deceptively simple question that reveals our biases. For centuries, humans have measured intelligence by proxies: memory, problem-solving, calculation, creativity. Yet computers already outperform us at some of these—does that make them intelligent? His answer: no, because these tasks miss the full scope of human cognition.

The Turing Test and Its Limits

Alan Turing’s 1950 thought experiment proposed a practical way to evaluate machine ‘thinking’: if you converse with a machine and can’t tell it’s not human, the machine passes. This idea inspired decades of chatbot competitions, from Joseph Weizenbaum’s 1966 program ELIZA to modern conversational systems. But as Sabouret points out, these programs excel by mimicking surface behavior, not by understanding meaning. The earliest chatbot ELIZA simply rephrased user statements—“Do you want to talk about your family?”—giving the illusion of empathy.

Even modern chatbots, despite massive databases, still operate by pattern-matching, not awareness. And the Turing Test itself, Sabouret notes, measures fooling ability, not comprehension.

The Chinese Room Argument

Philosopher John Searle challenged the Turing Test in 1980 through his ‘Chinese Room’ analogy. Imagine someone who doesn’t speak Chinese locked in a room with a manual matching symbols to responses. To an outsider, their replies appear fluent—but the person doesn’t understand a word. Likewise, AI manipulates symbols without meaning. The parallel is clear: syntax is not semantics. Computers can ‘say’ correct things without ever knowing why.

Why Comparison Fails

Comparing human and machine intelligence, Sabouret argues, is like comparing swimming to submarine propulsion: both move through water, but by entirely different mechanisms. As computer scientist Edsger Dijkstra quipped, asking if machines can think is “as relevant as asking if submarines can swim.” The machine’s power lies in brute calculation—testing millions of solutions instantly—while human intelligence thrives on abstraction, intuition, and social understanding.

Turing never intended his test to measure true intelligence, only to prompt reflection on whether mechanical processes could simulate thought. Modern researchers use tailored tasks—chess performance, language translation, image classification—rather than pretending computers ‘think’ in human terms. This distinction reappears throughout the book: the goal isn’t artificial humanity, but artificial capability.

Machines don’t think, they compute—and sometimes, computation gives us the illusion of thought.

Sabouret transforms these classic debates into clear insight for readers: stop asking whether AI is alive. Instead, ask whether it’s useful, ethical, and comprehensible. The real question is not ‘Can machines think?’ but ‘How can we think better with machines?’

Sabouret introduces a humbling truth: perfection is impossible in computation. Many problems, like building optimal school timetables or routing every delivery truck in a city, have astronomical possibilities. Even the fastest computers can’t explore all solutions. This is the heart of complexity theory—measuring not only whether a problem can be solved but how fast.

The Imperfect Heroes of AI

Since perfect solutions are often unattainable, AI thrives on ‘good enough.’ These heuristics—educated guesses—mirror human shortcuts: you don’t measure butter to the milligram when baking; you eyeball it. The A* (A-star) algorithm is Sabouret’s favorite example. Used in GPS navigation, it prioritizes paths that look most promising based on proximity and known distance. It doesn’t guarantee the absolute best route, but usually finds one close to it fast enough to be useful.

A Win for Pragmatism

AI, in Sabouret’s telling, is the celebration of pragmatism in science. He reminds us that ‘good enough’ often outperforms perfect but impractical solutions. This philosophy explains why Deep Blue or AlphaGo rely on heuristics rather than exhaustive search—the key lies in clever approximations, not infinite exactness. By embedding human notions of sufficiency into machines, we bridge logic and intuition.

In an age obsessed with optimization, this insight is liberating: intelligence, even artificial, is about managing limits, not abolishing them. The strength of AI is precisely in knowing when to stop calculating.

When Sabouret turns to machine learning, he challenges another misconception: machines don’t learn as humans do. They don’t form concepts; they detect regularities. The system observes examples and adjusts its internal parameters until its outputs align with desired results. Yet, despite this mechanical logic, the results can be astonishing—speech recognition, image labeling, translation—all powered by this pattern extraction.

From Unsupervised to Supervised Learning

Sabouret distinguishes between two paradigms: unsupervised learning, where machines cluster data into groups without labels (as in genetic data analysis), and supervised learning, where examples come with answers (‘this is a cat, this is not’). The latter has driven most recent AI advances. In both cases, success depends on how humans define features: size, color, shape, frequency. These numbers form the ‘coordinates’ of the algorithm’s reasoning space.

The Role of the Human Teacher

A recurring theme emerges: humans remain the architects of meaning. They choose data, design labels, and interpret outcomes. Machine learning automates correlation but not understanding. As Sabouret jokes, if you feed a cat-recognition algorithm a map of Paris, it will still find ‘cats’—because numbers alone carry no context. Intelligence requires interpretation, and that’s still uniquely human territory.

“The computer surpasses humans in speed, not sense.”

Machine learning, Sabouret concludes, is a mirror to human logic. It shows how much of what we call ‘thinking’ is actually structured repetition—and how much of AI’s brilliance still relies on our curiosity to ask the right questions.

Among AI’s many metaphors, none are more seductive than neurons. Sabouret traces their history from Frank Rosenblatt’s 1957 perceptron to today’s deep neural networks. These systems, inspired by biological brains, process inputs through layers of interconnected nodes (‘neurons’) that activate based on weighted sums. Each layer captures patterns of increasing abstraction—edges, shapes, faces, sentiments.

From Perceptron to Deep Learning

Rosenblatt’s early system could recognize triangles in crude 20×20 pixel images—a marvel for its time. But enthusiasm waned after critics like Marvin Minsky highlighted its limits: simple networks couldn’t model nonlinear relationships (like XOR logic). Only with GPUs in the 2010s did neural networks regain life through deep learning—the use of many layers and massive parallel computation.

What Deep Learning Really Does

GPUs, developed for video games, revolutionized AI by speeding up the repeated calculations neural networks need. Deep networks excel at perception—recognizing an image, translating speech—but they remain opaque. Unlike Michalski’s logic-based systems that can explain their reasoning, neural nets are black boxes. They can say a banana is a toaster if you tweak a few pixels (‘adversarial patches’), and no one can trace why.

By replacing interpretable logic with statistical power, deep learning traded transparency for performance. Sabouret warns this trade-off must be managed carefully, especially in critical domains like medicine and law. Intelligence without explanation is not understanding—it’s automation without accountability.

As the book nears its conclusion, Sabouret turns philosophical. The dream of ‘strong AI’—a self-aware machine—is, for now, fiction. All current AI is ‘weak,’ meaning it excels only in defined tasks. Yet this weakness is also its safety net: a chess program won’t suddenly desire power. True risk, he argues, lies not in machines gaining intent but in humans misusing them.

When Tools Become Weapons

Algorithms currently filter search results, personalize news, and optimize logistics—but these same technologies can manipulate information or power autonomous weapons. Sabouret’s caution echoes Stuart Russell’s warnings: the existential risk of AI comes from unexamined deployment, not spontaneous rebellion. Like any powerful tool, AI magnifies intention.

The antidote? Explainability. AI must be able to justify its outputs to humans. New research in ‘explainable AI’ aims to identify which data influenced a decision—an ethical step forward. In a world shaped by algorithmic decisions, transparency safeguards democracy.

A Human Future

Sabouret ends optimistically: AI reflects human ingenuity, not obsolescence. When designed ethically, it can enhance creativity, reveal hidden patterns, and free us from routine work. But understanding AI’s boundaries—its inability to reason, feel, or imagine—is essential. The machines we fear or worship are, in truth, ourselves rendered in code. The question that remains isn’t whether they’ll surpass us—but whether we’ll use them wisely.