Myths about AI

True or not? Answer before you read on. A wrong guess now helps you remember the answer later.

The whole story in three sentences. Many ideas about AI come from films, adverts and headlines. They sound logical and they stick. But once you see AI as a predictor, most of them fall apart.

You've probably heard a few. That a chatbot understood you. That AI is neutral. That it will soon do every job.

A myth here is an idea about AI that lots of people believe. But it doesn't match how AI works. Most myths start with something real. A chatbot does write fluent sentences. That's real. "So it understands" is the step too far.

These are the most common ones.

• "It thinks and feels." It predicts the next word from patterns. Fluent text is what a good predictor makes. Whether you can call that understanding, experts disagree (see below). Feelings: nothing in how it works needs them.
• "It knows everything." It only learned from its examples. Its predictions often contain mistakes.
• "It's neutral." A model shares the biases in its data. One study looked at face systems. They were wrong far more often for darker-skinned women than for lighter-skinned men.
• "If it sounds sure, it's right." Training and tests reward guessing. Saying "I don't know" scores worse. So it guesses, confidently.
• "It will take every job." The International Labour Organization did a large study. Most jobs are more likely to change than to disappear.

Myths stick because the predictor is very good at one thing: sounding like a person.

An example. You ask: "Are you sure?" The chatbot answers: "Yes, I'm sure." That sounds like confidence. But it's the most likely reply to "Are you sure?" in its examples. People usually say yes. So the predictor does too.

That's why your feeling isn't a good test. Text that sounds human, sure or caring was predicted. It wasn't felt.

Five reasons the myths stick:

• It sounds human. Fluent text feels like a mind behind it.
• Films told us so. Thinking robots are a great story. They're not how today's AI works.
• Adverts sell it. "It understands you" sells better than "it predicts your next word".
• It's often right. So you stop checking, until it isn't.
• It's hard to look inside. You can't see the patterns, only the result.

True or myth? Decide for each claim, then check.

Everything runs in your browser. Nothing you type is sent anywhere.

Ask four questions

Start today with the next AI headline you see. Ask these four questions.

Source

Who says it, and can you find where it came from? "Experts say", without a name, isn't a source.

Verb

Does it say the AI thinks, knows or wants? Translate it to predicts. Does the claim still make sense?

Money

Does someone earn from you believing it? An advert for a tool is not a test of that tool.

Test

Can you try it yourself? Ask the same question twice. Ask about something you know well.

Test the chatbot on what you know

Ask a chatbot about your street, your club or your hobby. Things you know well. See where it's right, and where it fills gaps with something that sounds right.

Not: "It answered confidently, so it's true." But rather: "It answered confidently. Is that true for something I can check?"

Three people meet a myth. A quick check shows what's really going on.

• Don't swing to "AI is useless"A myth-buster can overshoot too. Predictors are very useful for drafts, summaries and ideas.Use it, and check what matters.
• Don't fact-check a chatbot with a chatbotAsking "Is that right?" gets another prediction.Check with a source: a book, an official website, a person who knows.
• Don't trust a confident myth-buster either"AI can never…" is often a myth too.Ask the same four questions of every claim, including this page.

• explainerGoogle for Developers. LLMs: What's a large language model?.A language model predicts the next token, like a very large autocomplete. A token can be a word, part of a word or one character. Training needs a huge amount of text. Predictions often contain mistakes.
• explainerGoogle for Developers. LLMs: Fine-tuning, distillation, and prompt engineering.A base model is a platform, not a finished product; fine-tuning on examples for one task improves it. A prompt changes the output, not the model. Models share the biases in their data and cost a lot of computing power.
• researchKalai e.a. (2025). Why Language Models Hallucinate.Language models hallucinate because training and testing reward guessing over saying 'I don't know'. Like a student on a multiple-choice test, guessing scores better. Example: asked for one researcher's birthday, a model gave three different wrong dates. Three chatbots each made up a wrong title for his dissertation.
• researchBuolamwini en Gebru (2018). Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification.Commercial face systems got the gender of darker-skinned women wrong up to 34.7% of the time, and of lighter-skinned men at most 0.8%.
• researchMitchell en Krakauer (2023). The Debate Over Understanding in AI's Large Language Models.Researchers disagree strongly about whether language models understand language. The paper sets out the arguments on both sides.
• researchBubeck e.a. (2023). Sparks of Artificial General Intelligence: Early experiments with GPT-4.Researchers at Microsoft argue that an early version of GPT-4 could be seen as an early, incomplete form of general intelligence. They also list its limits.
• researchGmyrek e.a. (ILO) (2025). Generative AI and Jobs: A Refined Global Index of Occupational Exposure.One in four workers worldwide has a job with some exposure to generative AI. Office jobs are most exposed. Most likely, jobs change rather than disappear.

All sources of this course →

You get 7 questions about the whole page. Get 5 right, and you've mastered the topic.