Natural Language Generation (NLG) is the branch of AI that turns structured data or a model's internal representation into fluent, human-readable text. It sits on the output side of language AI: where NLP reads and interprets, NLG writes — producing anything from a one-line summary to a full report.
Classic NLG follows a pipeline: decide what to say, choose how to structure it, then realise it as grammatical sentences. Today's large language models collapse much of that into a single generative step, but the goal is unchanged — take facts or signals and express them in language a person can read. A weather app writing "Rain clearing by mid-afternoon, high of 14°C" from raw forecast numbers is NLG at work.
An online retailer with 40,000 SKUs used NLG to draft a description for every product from its spec feed: material, dimensions, use case, care notes. What once took a copy team weeks became an overnight batch. The upside is obvious; the trap is publishing that output untouched. Google's scaled content abuse policy targets pages mass-produced with little added value, so the retailer routed every draft through an editor for accuracy and one genuine detail before it went live. NLG is a strong first draft, rarely a finished one.
NLG is the writing half of language AI; its reading half is natural language processing, and both build on generative AI.