> Training data and input data can be censored if the fed really wanted to, especially in the circumstances that they have the IP for Claude's foundation models.

You can't really, if it's widely covered. Even if you filter the articles a lot of information suggesting it will leak into the training data through contextual clues.