What AI can and can't do

We'll close the foundations with the most practical question of all: what is this stuff actually good at, where does it fall down, and — crucially — why? Getting this right is what separates people who use AI well from people who get burned by it.

What it's genuinely great at

Generative AI is remarkably good at working with language and patterns:

• Drafting — first versions of emails, summaries, outlines, responses.
• Summarizing — condensing long documents, call notes, or threads.
• Transforming — rewriting, reformatting, changing tone, translating.
• Explaining — breaking down a concept or a piece of jargon.
• Brainstorming — generating options and starting points.

Notice the theme: these are tasks where a strong, fast first draft is hugely valuable and a human is still in the loop to check it.

Where it falls down

• Facts and recency. It doesn't look things up by default — it predicts plausible text. It can be out of date or simply invent details.
• Math and precise logic. It can fumble exact calculations because it's predicting what an answer looks like, not computing it.
• Truly novel reasoning. It's pattern-matching on what it has seen, so genuinely new or highly specialized problems can trip it up.
• Knowing what it doesn't know. It rarely says "I'm not sure." It tends to answer confidently either way.

Why it's confidently wrong: hallucination

This is the single most important concept on the page. When AI states something false but sounds completely certain, that's called a hallucination — and it's not a glitch, it's a direct consequence of how the model works.

Remember from earlier: a model predicts the most likely next token. It is always generating plausible-sounding text — that's its only job. It has no separate step where it checks "is this actually true?" So when it doesn't have the right pattern, it doesn't stop — it produces the most plausible-sounding text anyway, delivered with exactly the same confidence as a correct answer.

There's a second, more surprising reason, and recent research points right at it: models are trained and graded in a way that rewards confident guessing over admitting uncertainty. Think of a multiple-choice test where a blank guarantees zero but a guess might score — you learn to always guess. Models are optimized to be good "test-takers" in exactly this sense, so when they're unsure they tend to produce a confident answer rather than say "I don't know." That's why getting an AI to admit the limits of its knowledge is still genuinely hard.

That's why a hallucination is so dangerous: it doesn't look like an error. It looks just as fluent and assured as the truth.

Where this leaves us

That's the foundation. You now know what AI physically runs on, what a model is, that it's all prediction, and where that prediction shines and where it breaks. From here, the rest of the curriculum is about putting it to work — starting with how to actually talk to these models to get reliable results.