For any work with protected intellectual property, I use other providers, for the contractual guarantees, but I think it would be silly to think that OpenAI or Anthropic are not training on literally all data they get. How could you ever tell if they did? They can just claim the data was mislabelled, or ignore the accusations. If you have serious IP, use only local models.