City communications wants a generative model that drafts plain-language summaries of council meeting minutes. How does this differ from pure classification inference?

Drafting plain-language summaries of council meeting minutes with a generative model is a generative content-creation workload: the system produces new text conditioned on the minutes rather than emitting only a class label from a closed set. That differs from pure classification inference, which assigns council meeting minutes to predefined summary template categories without composing a fresh summary. Treating the task as a multi-label classification workload that assigns council meeting minutes to predefined summary template categories collapses generation into discrimination and misses the token-generation serving profile. Calling it only a retrieval workload that selects a similar archived summary from a council vector database describes a lookup pattern that does not draft a new summary from the minutes. Treating it as a sequence-labeling workload that tags named entities in council minutes for indexing and search produces annotations rather than a plain-language summary and misses the generative content-creation task. Exam caveat: generative summarization may still use classification or retrieval as helper stages, but the primary workload distinction on the exam is generating text versus only labeling. Operational check: validate summary fidelity against approved minutes, set token and latency budgets appropriate for generation (not just classifier scores), and keep human review where municipal communications policy requires it.