Even if it's not used explicitly for "training", it's still valuable for model evaluation, A/B testing, etc. (e.g. "do users engage more with model A or B?")

Even if it's not used explicitly for "training", it's still valuable for model evaluation, A/B testing, etc. (e.g. "do users engage more with model A or B?") And to be clear, I'm not accusing Apple or OpenAI of being "evil" here wrt. handing user data, I just think it's a potential factor in their partnership.