What Tristan just alluded to, is that you can always take a watermarked output from DALL-E 3 or whatever, and then use a small open source model to rephrase it. And once it’s rephrased, remixed, the watermark is gone and the original maker, OpenAI, doesn’t recognize the remixed picture anymore. But the remixed picture, it’s is as persuasive or even more persuasive than the one that’s generated by the larger model.