Couldn't user data make it into the model, and be reverse-engineered from it, if you used it to train the model?

I mean, machine learning models aren't designed as one-way cryptographic hash functions, so you can't be sure that some of the training data can be inferred through the model's behavior.