Generative AI
Any model whose output is newly produced content — text, image, audio — rather than a classification, ranking, or extracted span.
Last updated
“Generative” describes a model whose output is made rather than found. Large language models, diffusion image models, and speech synthesis systems are all generative. The distinction is not architectural — the same transformer can be used generatively or not — it’s about what the model is asked to do at inference time.
Why it matters for your research. A generative answer is written, not looked up. Even when a generative system has “read” your sources, its paraphrase of them is a new object that may or may not faithfully track the original. The moment a claim is cited, it needs to point at a primary-source span, not at a paraphrase of one.
In Archēglyph. Deliberately not generative. The pipeline classifies, ranks, extracts, and clusters. Where a generative-flavoured tool is used (e.g. a VLM for page-layout assessment), the output is a structured decision with the model id recorded — not prose shown to the reader as content. See Why Archēglyph cannot hallucinate.
Not to be confused with. Extractive is the opposite stance — see Extractive QA. Generative ≠ AI: classifiers, embedders, and search rankers are AI too, and they are not generative.