lesson
15 min

How AI Thinks (It Doesn't)

A plain-language explanation of how AI models actually work — patterns, statistics, and a whole lot of data.

Let's clear something up right away: AI doesn't think. It doesn't understand. It doesn't have feelings or opinions or goals. When we say things like "AI thinks" or "AI knows," we're using shortcuts that can be misleading.

So what's actually happening under the hood? Let's break it down without the computer science jargon.

It's All About Patterns

Imagine you've eaten at a hundred different taco trucks in San Antonio. After a while, you start to notice patterns. The ones with long lines usually have better food. The ones near construction sites tend to open earlier. The ones with hand-painted signs tend to have more traditional recipes.

You didn't sit down and study taco truck theory. You just noticed patterns from experience.

AI works the same way — except instead of a hundred taco trucks, it's looking at millions or billions of examples. And instead of casual observation, it's using math to find patterns in data.

That's really it. AI is pattern recognition at a massive scale.

How a Model Gets Built

Here's the basic process:

Step 1: Gather a ton of data. If you want an AI that recognizes dogs in photos, you show it millions of photos labeled "dog" and "not dog."

Step 2: Let the AI find patterns. The AI looks at all those photos and figures out what the "dog" photos have in common. Maybe it notices that dogs tend to have four legs, fur, certain ear shapes. It's not thinking "that's a dog" — it's finding statistical patterns in the pixels.

Step 3: Test it. Show the AI new photos it's never seen and see if it correctly identifies dogs. If it gets 70% right, that's not good enough. Adjust and try again. If it gets 98% right, you might have something useful.

Step 4: Deploy it. Now you can use that model in an app, a website, or a product.

The AI that powers tools like ChatGPT and Claude went through this same basic process, but instead of dog photos, they were trained on text — books, websites, articles, code, conversations. They learned the patterns of human language so well that they can generate new text that sounds like a person wrote it.

The Key Insight: Prediction, Not Understanding

When you type a question into ChatGPT or Claude, here's what's really happening: the AI is predicting what words should come next.

It's like the world's most sophisticated autocomplete. Your phone suggests "you" after you type "thank" — because that's the most common next word. AI chatbots do the same thing, but across entire paragraphs and with much more sophistication.

When Claude writes a paragraph about the history of San Antonio, it's not remembering facts it learned. It's generating text that statistically resembles what accurate information about San Antonio looks like, based on all the text it was trained on.

This is why AI can sometimes be confidently wrong. It's not lying — it's predicting what a good answer looks like, and sometimes the prediction misses. We call these hallucinations, and they're important to watch for.

What "Training" Actually Means

When people say an AI model was "trained," they mean it processed a massive amount of data and adjusted its internal settings to get better at pattern recognition.

Think of it like this: imagine you're learning to shoot free throws. Every time you miss, you adjust a little — your angle, your force, your follow-through. After thousands of shots, you've developed muscle memory. You're not consciously calculating physics. Your body learned the pattern.

AI training works similarly. The model makes predictions, checks if they're right, and adjusts millions of tiny internal numbers (called parameters) to get better. GPT-4, the model behind ChatGPT, has over a trillion of these parameters. Claude has a similar scale. That's a lot of adjustments.

What AI Is Good At

Because AI is fundamentally about patterns, it's great at:

  • Finding patterns in huge amounts of data — way more than any human could process
  • Generating content that follows patterns — text, images, code, music
  • Classifying things — sorting emails, identifying objects in photos, detecting fraud
  • Predicting outcomes — weather forecasts, traffic, what you might want to watch next

What AI Is Bad At

Because AI doesn't actually understand anything, it struggles with:

  • Common sense — It doesn't know that water is wet or that you can't fit a car in a backpack
  • New situations — If something is very different from its training data, it guesses poorly
  • Moral judgment — It can't tell right from wrong. It can only mimic what it's seen
  • Knowing what it doesn't know — It'll give you an answer even when it should say "I don't know"

The Bottom Line

AI is powerful, but it is not alive and it does not think. It is math running on patterns from data. When you understand that, you're already ahead of most people — because you'll know when to trust it and when to double-check.


Exercises

Try This: Open Claude (claude.ai) or ChatGPT (chatgpt.com) and ask it: "What is Community Incubator in San Antonio?" Read the answer carefully. Does it seem accurate? Can you spot anything it might have gotten wrong or made up? This is good practice for catching hallucinations.

Try This: Ask an AI tool the same question three different ways. For example: "What's the weather like in San Antonio in July?" then "Describe San Antonio summers" then "How hot does it get in SA?" Compare the answers. Notice how the same AI can give different responses to similar questions — because it's predicting text, not retrieving facts.

Talk About It: If AI doesn't actually "understand" anything, should we be worried about it or not? Does it matter whether AI truly thinks, as long as it gives useful results? Where could this become a problem?

Quick Check

Quiz: Prediction, not understanding

  1. 01A chatbot gives you a confident answer about a San Antonio bus route. What is the best next step?

    Choose one.

  2. 02Why can the same question get different answers?

    Choose one.

Go deeper: Machine Learning, Minus the Hype, a Temerarii Media playbook by Dom Davis.