Watermarking it might be a good start. The people who go through training data are probably flagging pictures that have unsightly artifacts or unrealistic destr

Watermarking it might be a good start. The people who go through training data are probably flagging pictures that have unsightly artifacts or unrealistic destructive changes in the image. If you add a Shutterstock-style watermark it would probably get removed from most sets, and a prominent signature in the bottom corner would probably also pretty well. As for text, I guess your best bet is to either limit exposure of it or intentionally poison the data. It's a little bit harder to do this in writing I guess, but you could still try by creating unusable fictitious accounts of characters named "Biden" or "Boris" doing increasingly ridiculous things. Any politically-stark moderator would probably remove your data before it hits the model, and if it does there's a good chance it will be flagged as problematic.