How does AI work, really? And why does everyone think they “get it”?

The big and burning question is: “How does AI work?” Lots of people claim to know the answer—but they’re usually off-base. Let’s keep it straightforward and actually see what’s under the hood.

Understanding how AI works is like peeling an onion: each new layer grows more complicated, and occasionally draws a tear. These days, you’ll hear plenty of self-proclaimed gurus tossing around phrases like “large language models”, “deep learning”, and “machine learning”, but they usually stop short at the AI onion’s outer layer. 

Here, in answering the burning question—“how does AI work?”—we’ll peel a few layers deeper, enough to go beyond the buzzwords, but not so far that we end up lost in the technical weeds. But first, let's deal with the wave of self-anointed experts.

Why everyone’s an AI expert these days

A plane can fly because its engines provide thrust and its wings generate lift, allowing it to overcome gravity and stay in the air. No matter how many times you read that explanation, you’ll still marvel at how that massive, many-ton hunk of metal manages to stay aloft.

In philosophy of science, one classic term for this kind of “explanation” is dormitive explanation. The concept originates from Molière’s satirical play Le Malade imaginaire, in which doctors assert that opium causes sleep because of its “dormitive virtue”—essentially restating the phenomenon (opium causes sleep) in different words rather than explaining how or why it does so. Dormitive explanations thus label a phenomenon without offering deeper insight into its underlying mechanism.

On the surface, we might feel as though we fully understand something merely because we can label its parts—like “lift,” “thrust,” and “gravity”. However, when pressed for details, our shallow grasp of the actual mechanisms quickly becomes apparent, revealing that our initial sense of understanding was largely superficial. The same is true for AI.

All these experts — they’re stuck in the dormitive trap

That’s why many people—knowingly or otherwise—get stuck in a dormitive trap: they accept the label (“machine learning,” “deep learning”) as if it fully explains how AI works. It’s not always a matter of faking expertise; it’s simply that they assume saying “the engine provides thrust” solves the mystery of flight—when, in reality, there’s so much more going on.

This superficial understanding can come at a cost. According to Forbes, as many as 85% of AI projects might fail to deliver on their intended outcomes. One reason is that businesses latch onto buzzwords and fall for the illusion of explanatory depth, thinking they’ve mastered AI when really they haven’t. 

The first step: A human goal and an AI model that’s up for the job

We’ll begin at the source—the human who wants something done. All AI systems start with a goal, meaning a human decides what problem the AI should solve or what task it needs to perform. Based on the defined goal, developers choose—or design—an AI model. 

An AI model is, at its core, a mathematical approach or algorithm tailored to the problem at hand; much like choosing an SUV for off-road treks rather than a sports car for city streets. Think of the model as the engine, and the AI system as the entire vehicle, including software, hardware, data pipelines, user interfaces, and other processes that make the model actually run.

One of the earliest recognized AI systems, the “Logic Theorist,” emerged in the mid-1950s. The AI system’s goal—which inspired the development of its logical model—was to prove basic math theorems, hardly mind-blowing by today’s standards, but revolutionary for its time. 

Today, a goal might be to make cars drive themselves, instantly translate speech in real time, or manage drone deliveries across an entire city. And there are plenty of AI models to match those goals, including: neural networks, decision trees, random forests, transformers, reinforcement learning, and more.

Key takeaway: All AI systems begin with a well-defined goal, which then dictates the choice of an appropriate model from the growing array of mathematical approaches, ensuring the AI is tailored to solve exactly the problem it was designed for. An AI model is the engine, the AI system is the vehicle. 

The second step: We feed the AI with experiences

You and I, and every human who has ever lived, learn what we know largely from experience. You wouldn’t know how to describe a vibrant sunset if you had never experienced color. We can’t simply reason our way to understanding; experience is the raw material from which reason crafts its insights. We need input in order to generate meaningful output.

The same is true for AI systems. They require input—images, text, or other forms of information. They need to “experience” situations and learn from them. Now, of course, what an AI can learn through experience differs from human learning in both scale and nature. An AI might analyze millions of images or billions of words, sifting through data far beyond the capacity of any single human mind. 

The input—the experiences we encourage the AI to seek—depends on the system’s goals: a computer vision system learns from images; a language model relies on text; and a personal AI assistant may integrate a variety of sources.

Different ways of feeding the AI

Unlike humans, who can explore new places and experiences at will, an AI system cannot simply walk out into the world to gather information. It depends on us to feed it the data it needs, or, more precisely, the data we choose for it, based on the job it’s meant to do.

For example, if you’re training a computer vision system to recognize cats and dogs, you might feed it thousands—or even millions—of labeled images. The AI then goes through each image, notes key features, like fur patterns and ear shapes, and slowly figures out how to tell the animals apart.

Sometimes that information comes from structured data, which is often prepared and labeled by humans—think of tons of images tagged “cat” or “dog,” or lines of text with known categories. Other times, it may come from user interactions or prompts we type in when using a product. This is one reason many AI tools were initially offered at no cost: they could gather real-world input from curious users, further refining the system’s understanding. 

Key takeaway: Just as humans learn from their experiences, AI models require a continuous flow of data—whether meticulously labeled or gathered from real-world interactions—to develop their pattern recognition skills. The quality and variety of this input directly shape how effectively the AI can learn and perform.

The third step: From guesses to greatness

After devouring endless images, text, or any other data thrown its way, the AI model takes what it has “seen” and moves on to making guesses about new information. Think of it like a chef tasting a recipe as it simmers—constantly sampling, comparing, and adjusting to get the flavor just right.

For a model trained to recognize cats and dogs (somewhat trivial), this might mean encountering a fresh image and predicting “cat.” If the real answer is indeed “cat,” the AI is on the right track. But if it’s actually “dog,” that error prompts the real magic: the AI adjusts its internal settings—millions or even billions of mathematical “weights”—to guide its future guesses closer to the truth. A weight is a numerical value in a neural network that determines how much influence a given input has on the model’s output. These weights are adjusted during training to improve accuracy.

It’s essentially a giant game of “hot or cold,” played out over countless rounds. Each wrong guess nudges the AI in a slightly better direction, and each right guess cements the progress it’s made. Over time, these minute corrections add up: the model matures from a clueless amateur to a seasoned expert—whether it’s sorting images of cats and dogs, detecting tumors in medical scans, or generating text that sounds suspiciously human.

This iterative dance—guess, compare, adjust—lies at the heart of modern AI. By continually fine-tuning its internal parameters, the AI gradually refines its understanding, allowing it to tackle an ever-growing range of tasks with impressive accuracy and finesse.

Key takeaway: AI models improve by making guesses based on their experiences, comparing them to the truth, and adjusting their internal parameters accordingly. Each error becomes a small nudge that refines the model's accuracy, steadily transforming it from a novice into an expert over countless iterations.

Peeling back further: A glossary for the technically curious

For those tech-savvy minds eager to delve deeper into AI, here’s a handy glossary of common phrases:

1. Machine Learning. An umbrella term encompassing a range of algorithms and techniques that allow systems to learn from data.

2. Large Language Model (LLM). A type of AI model designed to process and generate human-like text by leveraging extensive datasets and deep learning techniques.

3. Deep Learning. A subset of machine learning that uses layered neural networks to learn representations of data with multiple levels of abstraction.

4. Neural Network. A system of interconnected nodes designed to mimic the human brain’s way of learning from data.

5. Gradient Descent. An optimization method that iteratively adjusts model weights to minimize the error (loss).

6. Backpropagation. The process of propagating errors backward through a network to fine-tune its weights.

7. Overfitting. When a model learns the training data too well—including noise—thus performing poorly on new data.

8. Underfitting. A scenario where a model is too simple to capture the underlying patterns in the data.

9. Regularization. Techniques used to prevent overfitting by penalizing overly complex models.

10. Learning Rate. A parameter that determines the size of the weight updates during training.

11. Activation Function. A mathematical function applied to a neuron’s output to introduce non-linearity into the model.

12. Loss Function. A measure of how far off a model’s predictions are from the actual target values.

13. Epoch. One complete pass through the entire training dataset during the learning process.

14. Batch Size. The number of training examples processed at one time during an iteration.

15. Convolutional Neural Networks (CNNs). Specialized networks that use convolutional layers for image and pattern recognition.

16. Recurrent Neural Networks (RNNs). Networks designed to handle sequential data by retaining information from previous inputs.

17. Transformer. A model architecture that leverages self-attention mechanisms to process sequences in parallel, revolutionizing language tasks.

18. Attention Mechanism. A method that enables models to focus on relevant parts of the input, improving prediction accuracy.