It is fascinating to me how many people feel they need to comment on Anthropic regarding watermarking in generated content. Two things really catch my attention here. First, why is this the problem instead of the fact that Anthropic was the last frontier lab not to do this anyway?
In case you're curious how they do it: SynthID is the technology that Google and OpenAI already integrate for that purpose (here are they describing their approach). This brings me, naturally, to a second fascinating aspect - why were people not expecting or demanding this from Anthropic in the first place?
Truthfully though, even if we solve the watermarking problem (and making it really easy to detect them), it won't solve for the fact that people blindly reshare anything that a) looks good and b) confirms their existing opinions.
I saw an example of this today. An impressive looking solar plant that allegedly floats in front of the canary islands went viral. I was curious, so I started digging for details. One might think it is pretty suspicious when the most respected medium carrying the story is a Yahoo page.
Humanity has been exposed to social media for over ten years now and we still haven't learned anything about how utterly untrustworthy the internet is, or how to validate any claim that comes your way. I'm pessimistic about watermarking truly solving for that problem but at least it is a step into the right direction.
Comments
Replies come from Mastodon and Bluesky. Reply on Mastodon or Bluesky to join in.