Google Cloud Tech says LLM text output is poor for classification decisions

Google Cloud Tech post by Logan Hennessy argues that LLMs used as drop-in classifiers return text rather than probabilities, which blocks thresholding, abstention, and auditing.

Hennessy says discriminative models predict a fixed label set in a single forward pass and avoid the sequential cost of autoregressive generation. He cites Gemini's SafetyRating objects as an example of generation and classification returned together in one API response.