ML.GENERATE_TEXT vs AI.GENERATE: model objects no longer needed

The article frames the confusion as a generational split: older guides use CREATE MODEL to build a remote model object and then call ML.GENERATE_TEXT as a table function, while newer material calls AI functions such as AI.GENERATE directly as scalar functions. It notes that articles on the web mix both generations, so following a tutorial without noticing which generation it belongs to is a common cause of queries that fail to run.

The practical guidance for teams with existing assets is to start by finding where ML.GENERATE_TEXT is used, checking INFORMATION_SCHEMA.JOBS_BY_PROJECT and INFORMATION_SCHEMA.MODELS or searching the code repository with grep. It then recommends running the old and new styles side by side on a small sample, such as 50 rows, and comparing null outcomes and differences before switching. It also warns that downstream dashboards, filters, and assumptions about text length can break, especially for classification and judgment tasks where a small label change can make a WHERE condition miss.

What the outcome hinges on is whether the parallel comparison shows acceptable differences and whether downstream consumers survive the change, so the article's advice to keep the old model object for a while and not drop it immediately is a hedge against needing to switch back. For teams doing classification, the piece suggests this may be a natural moment to consider AI.CLASSIFY instead of free-text labeling.