In one sentence
Generation is producing new content, usually text, such as a draft reply, a summary or a description.
The simple model
Where classification chooses from fixed categories and extraction pulls out existing values, generation creates something new. A draft email, a summary of a document, a first version of a description.
This is the capability most people picture when they think of AI, and it is genuinely useful for the repetitive drafting that fills a lot of work.
Why you care
Generation removes repetitive preparation. Drafting a response, summarizing a long thread, writing a first version of a quote. A person then reviews and finishes it, keeping control over the result.
But generation is also the capability that needs the most review, for two reasons:
- The output is open-ended, so there is no fixed set of correct answers to check against.
- Generated text can be fluent and confident while being wrong or invented. This is a hallucination, and it is the main risk of generation.
Keep generation grounded and reviewed
Two habits make generation safe to use:
- Ground it in real sources. Have it generate from retrieved, current information rather than from memory, and cite what it used. See retrieve then answer.
- Keep a person in the loop for anything consequential. Generation is at its best producing drafts that a person approves, as in draft, approve, execute.
A small example
An AI drafted replies to common customer questions, grounded in the current help articles and citing them. Agents reviewed and sent. It saved real time on routine replies, and because every draft was checked, the occasional wrong or invented detail was caught before it reached a customer.
What this does not solve
Generation produces fluent content; fluency is not accuracy. It does not verify facts, follow policy, or know when it is wrong. Review and grounding reduce the risk but do not remove the need for a person on anything that matters.
