Put together some notes on the Gemini terms-of-service: it looks like their paid API tier doesn't train on your inputs to the model, but the free API tier does:

simonwillison.net/2024/Oct/17/

Data used to create fine-tuned models, even on the free tier (they have a free tier!) won't be used for training. I hadn't realized you could fine-tune a model there for free, that's pretty wild: ai.google.dev/gemini-api/docs/