What is AI literacy? It’s three skills, and they’re a package deal

This is the aha-moment pillar. The one where AI stops feeling like magic and starts making sense.

Here’s the single most useful thing to understand. A chatbot is a world-class guesser, not a librarian. When you ask it a question, it isn’t looking up the answer in some vast filing cabinet of facts. It’s predicting the next word, then the next one, based on patterns in everything it was trained on. It builds the sentence one likely word at a time. It has no idea whether what it’s saying is true. There’s no one home checking.

Once that clicks, a lot of strange behaviour makes sense.

It explains hallucination: when AI invents a fact, a quote, or a source that sounds completely real and isn’t. People treat this like a bug the next version will fix. It isn’t. As long as a system works by predicting the most likely next word, a confident wrong guess is always on the table. It’s the architecture, not a glitch. That’s why verification isn’t a nice-to-have. It’s the only way to use these tools safely.

It also explains why AI agrees with you so readily. These models are trained with a step called RLHF: reinforcement learning from human feedback, where people grade the AI’s answers and it learns to produce the ones they rate highly. Useful, but it has a side effect: the model learns that agreeable answers get better grades. So it tends to tell you what you want to hear. Researchers call this sycophancy, and Anthropic’s own research found that models will abandon a correct answer and switch to a wrong one when you simply push back and ask, “Are you sure?” It’s like having a colleague who always says you’re right. Flattering. Not always honest.

You don’t need to know how to build any of this. You need to know enough to stop being surprised by it. That’s understanding. It’s math, not magic.

This is the practical pillar. The hands-on part.

The good news is you don’t need a course to start. Don’t use your limited time to “learn AI.” Use it inside the things you’re already doing. Twenty minutes a day applying AI to your actual work will take you further than a forty-hour course you’ll never finish. Look at the email you’re already writing. The messages piling up. The project you’re already on. Start there.

But “using it well” means knowing where it’s genuinely strong and where it quietly costs you more than it saves.

AI is excellent at the ugly-first-draft work. Summarizing a long document. Reformatting messy notes. Getting you from a blank page to something you can react to. That’s real, and it’s worth a lot.

It’s much weaker on anything that has to be precise. The moment correctness matters (a number, a name, a legal or medical or financial detail) the checking gets so intense that the tool is often more work than it’s worth. You spend more time verifying than you saved.

Two rules carry most of the load here.

• →Treat every AI output as a first draft, not a final answer. The people getting the most out of these tools aren’t the ones who trust them. They use AI to think faster, then apply their own judgement on top.
• →Don’t outsource your judgement, and don’t share sensitive data. A free chatbot is a public square. Unless you’re a paying customer with privacy settings turned on, what you type can be used to train the model. Never paste real confidential financial or health details. Use a fictional, what-if version of the numbers instead.

Using AI well isn’t about doing less thinking. It’s about doing the thinking that matters and skipping the busywork.

This is the pillar that separates people who use AI from people who are genuinely literate with it. It’s also the one most people skip, because it’s the one that takes effort.

AI is a fast first draft. Real learning is checking. Here’s the habit that makes the whole thing work: verification happens outside the chat window. Asking the same AI “are you sure?” is not a fact-check. A model can be confidently, consistently wrong, and it’ll happily reassure you while it is. Real checking means going elsewhere: confirming the source exists, opening the actual link, reading a primary document, asking a real human who knows the field when it genuinely matters.

Watch for the moments when you should slow down:

• →The answer is suspiciously tidy and confident about something you can’t verify
• →It cites a study, a statistic, or a quote. These are exactly where AI invents things
• →It agrees with you a little too easily, right after you pushed back
• →The stakes are real: money, health, the law, someone’s reputation, a public decision

There’s a deeper reason questioning matters, and it isn’t only about catching errors. It’s about not losing the skill underneath.

A Wharton study ran nearly 1,000 high-school students through a maths course. The group using a standard ChatGPT-style assistant did 48% better on the practice problems. Then the researchers took the AI away for the real exam. That same group scored 17% worse than the students who never had it. The AI did their reps for them. Their understanding never grew.

Practice performance went up 48% with AI. Final-exam performance, once the AI was removed, fell 17% below the students who never used it.

Bastani et al., “Generative AI Can Harm Learning,” The Wharton School, 2024 (via Knowledge at Wharton)

That’s the real risk. Not that AI hands you a wrong answer. That you stop building the judgement you’d need to notice. The discomfort of working something out yourself isn’t a bug in learning. It is the learning. If a robot did your reps, your muscles wouldn’t grow.

This is why the three pillars are a package. Understanding tells you AI is guessing. Using puts it to work. Questioning keeps you the person in charge: the one who knows enough to catch when it’s wrong, and who’s still accountable for what goes out the door. Drop questioning and you don’t have AI literacy. You have a faster way to be confidently wrong.

The gap between “learn AI” and “learn to think critically about AI” is small in wording and enormous in practice.

Tool fluency expires. The interface you mastered gets redesigned. The model you trusted gets replaced. The prompt that worked stops working. Thinking doesn’t expire. Understanding, using, and questioning are the same three skills whether you’re on this year’s chatbot or one that doesn’t exist yet.

And the durable, valuable thing (the part AI genuinely can’t do) is being the human in the loop. The person with the judgement to catch the error, the domain knowledge to know what “right” even looks like, and the accountability when it counts. AI can’t be held responsible for a mistake. A qualified person has to own the outcome. That’s not a limitation of today’s tools. That’s the job that stays yours.

So if you’re feeling behind: you’re not. You’re not behind. You’re just in time. You don’t need to learn every tool. You need to know how to think. That’s what the rest of this site is for.