Older NLG systems used templates and rules. A bank’s system might take a spreadsheet of transactions and fill in “You spent $412 on groceries this month, 8% more than last month.” Sports results, weather reports and financial summaries were often written this way. Today most NLG is done by large language models, which generate text one token at a time.
NLG is one half of natural language processing. The other half, often called natural language understanding, is about reading and interpreting text.
Why it matters
Fluent generated text is easy to mistake for accurate text. Because modern NLG can produce confident, well-written sentences that are wrong, checking generated text against its sources has become a core part of building AI products.
No published Chain of Thought episode uses the term directly yet, so this entry relies on the sources below.
Related terms
AI hallucination describes NLG that sounds right but isn’t. Generative AI is the broader category that includes text, image and audio generation.
Sources
- Gatt and Krahmer: Survey of the State of the Art in Natural Language Generation — Survey defining NLG as generating text from non-linguistic or textual input, covering tasks, template- and rule-based systems and newer neural approaches; checked October 24, 2026.
- Jurafsky and Martin: Speech and Language Processing (3rd ed. draft) — Standard open textbook covering language modeling and text generation with neural networks and large language models; checked October 24, 2026.