What generative AI is
- foundations
- task fit
Generative AI is software that produces new content — text, images, audio or code — in response to an instruction you write. It doesn’t look anything up: it constructs the most plausible response from patterns absorbed during training. That single fact explains both why it feels magical and why it can be confidently wrong.
What makes it “generative”
Plenty of software “knows” things without generating them. A spam filter sorts messages into two buckets; a recommendation engine ranks products that already exist. Those systems classify or rank.
Generative systems continue. Give one a partial pattern — your question — and it extends it into something new: a paragraph, a sketch, a function, a translation. The output did not exist a second earlier, and it was not retrieved from a database. It was constructed on the spot.
Why this matters
This changes how you choose tasks. Two everyday consequences:
- Transformations are its home ground. Rewriting, summarizing, structuring, translating, extracting — anything where you supply the raw material and the model reshapes it — works well and fast.
- Freshness, precision and provenance are not. Live prices, private databases, exact arithmetic, verified citations: a plain generative model has none of these. And because it always continues the pattern, it will rarely say “I don’t know.”
A bad example
Ask: “What’s the current EUR/USD exchange rate, to four decimals?”
You get a specific number — possibly yesterday’s, possibly invented. There is no live feed behind the answer; fluency is not accuracy.
A better example
Ask: “Draft a five-step checklist I can use to verify an exchange rate before quoting it to a client.”
Same underlying need, right division of labor: the model structures the process, and you keep the step that requires live data.
Why it works
A durable mental model: training stores patterns; generation samples from them. Once you hold that, three rules follow naturally:
- If everything the task needs can live in your prompt, generative AI is a strong fit.
- If it needs something the model cannot have — live data, your private files, exact computation, verified sources — pair it with a tool that provides that, or don’t delegate it.
- When the output will be acted on, verify it. Confident phrasing costs the model nothing.
Practice
Which task fits a plain generative model?
You are using a general chatbot with no web access and no calculator tool. It generates responses from patterns learned in training.
Which of these requests is the best fit for that setup?
Hint
Ask: does the task need something the model cannot have — live data, exact computation, or verified sources?
Why this is the answer
Generative models shine at transforming content you provide: structuring, rewriting, summarizing, extracting. They are weakest exactly where you need guaranteed freshness, precision or provenance. Matching the task to the tool is the first practical AI skill — it prevents most “AI got it wrong” stories before they happen.
Transfer
The test — “does it need something the model can’t have?” — travels everywhere:
- Drafting email or meeting notes: everything lives in your prompt. Good fit.
- Checking a price, a law or a statistic: needs a source. Use search or the primary source.
- Deciding something with consequences (taxes, health, contracts): wrong delegation entirely.
Next
Next: What an LLM actually does — the mechanism behind “plausible continuation.” Further foundations lessons arrive in waves, and the LLM glossary entry covers the basics.