Scott Aaronson spoke about this in a colloquium where he said that this was mooted at OpenAI before the decision was made by Altman to not implement it for the reasons you describe.
replyOpenAI will soon be adding watermarking to text as well, as it signed the EU Code of Practice on Transparency of AI-Generated Content: https://openai.com/index/advancing-responsible-ai-across-eur...
replyI’m more worried that this will degrade performance. I want the best results from a model, not the results that fit a constraint that’s not defined by me. Any increased cost or latency is also unacceptable.
replyThere are already small models trained specifically to prevent statistical detection, e.g., https://huggingface.co/kalpeshk2011/dipper-paraphraser-xxl
replyI guess whoever is the policy maker is assuming that some protection is better than none and that most people will not reach for such tools.
Either that or they want to comply with the EU AI Act when it affects them.
reply