Spent a long time looking for TOS if the experimental models are training on prompt and response data. Not clear. I did notice lots of comments here that the EXP models have quotas that match the free-tier plans.
Other LLMs with web search capability all say Gemini Experimental train on our data. Is this true? If we use the experimental models, are we leaking data to the google data sets?
Welcome to the forum.
When using a paid plan, your data transmitted to the LLM remains your data. It will not be used to train future models.
For all experimental models and when using AI Studio, you do not have the option to use a paid plan. Follow the logical conclusion. I recommend using publicly available data (or made up data with no commercial value) when testing the experimental models; ditto when working on a prompt using AI Studio. That gets you almost all the way there, you can then do final prompt tweaks using the API and on a pay-as-you-go plan.
Hope that helps.
I only use it through our api keys on my internal dev app. I can query gemini models using the suggested method and get a response that includes all exp models and 1.5 pro-002, etc. But i am not sure i am actually being charged for the exp model on my api/gcp project. I was worried based on some comments here that google treats exp tos differently as it is “released to get feedback” and they allow access without charging the api key even if made using methods based on api keys. I dont have any programmatic way of knowing.
Do any exp models charge the pay as you go or do they fall under special tos and are essentially “free previews to get feedback.”. They say not for production use so perhaps that also means not available through api pay as you go even though i access it like i access any other pay as you go model.
On consumer side, Google one premium with gemini is a paid service, but they state after digging in tos that they train on consumer prompts even if paid! (Ridiculous and only feasible bc they own android). So that got me thinking i need to make sure i understand if there are hidden data mining associated with exp models
Accessing the experimental models doesn’t bump up any usage counters for your paid plan. You are not getting charged when using them. The experimental bunch are free-only, and that means data sent to them can be used to train future models. The experimental models also only get the free quotas, like 50 a day for the top-of-the-line model.