No. Most data generated this way is poor quality. It's not the user responses and or queries. If the user does not know better than the LLM, you can generate bad responses. The value is in taking a superior model, submitting a query, and getting a higher quality output than you yourself could have generated, and using that to boost your model.
You identify users doing real work and implementing a project over a long period of time and train on their traces.