In a nutshell People expect AI to be either an all-knowing mind or a dumb autocomplete — and both pictures get in the way. There's a third, accurate view: a generative model has four properties, and almost any strange behaviour comes down to one of them. Name the property that slipped, and you know the cause and the precise fix. This is the map of the section; each property gets its own article.
The four properties
- How generative AI works — what these models really are and where their "character" comes from (training). The foundation under the other three.
- Next-token prediction — where answers come from. The model builds a plausible continuation one piece at a time.
- Knowledge & hallucinations — what the model knows and why it can be confidently wrong.
- Working memory & context — what it looks at right now and what falls off the edge.
- Steerability — how far your instructions steer it.
And the finale — when properties collide: in a real task they work together, and errors stack.
Cheat sheet: which failure, which property
| What you see | Property | Where to look |
|---|---|---|
| Fluent but wrong; invented fact, number, link | Knowledge | knowledge & hallucinations |
| "Forgot" what was said early in a long chat | Working memory | working memory & context |
| Ignored the format or the rule | Steerability | steerability |
| Answers differently every time | Next-token prediction | next-token prediction |
| Refused to do it | Steerability (limits) | a built-in safety boundary |
The core skill isn't to "trust" or "distrust" AI — it's to see the kind of error and apply that error's fix.
Source
The section is built on the Anthropic Academy course AI Capabilities and Limitations (completed; the badge is in the Anthropic Academy block on the home page).