Ofcourse the AI corps are incentivized to downplay the effects of watermarking where they can, as they stand to gain so much from rolling it out (prevent model collapse).

Also interested in how this watermarking push makes sense when considering RSI.