1:00
Does an AI text watermark actually tell you who wrote something?
OpenAI's new watermark for ChatGPT text was one of Tuesday's stories. By OpenAI's own tests, its detector caught about 92% of watermarked text, and only 66% once one word in ten was swapped for a synonym.
On the show my co-host Nico Lafakis went one step further: pass the output through another model or two, and it's easy to wash out. For images, the mark is worth caring about. For text, in Nico's words: "this is not relevant," and "way too little too late on text."
If you're counting on a watermark to know what's human-written, here's a reason to plan for something else.
Watch the episode: https://valuefirststudios.com/media/episode/vf-ai-daily-2026-10-06
The next step