AI Watermarks: Useful or useless?
A little while back, Claude announced that it will be watermarking everything its AI models create.
**collective gasps of “oh no” from thousands of firms ranging from law to professional services to creative to accounting to you name it **
What does this mean - fairly simple, you can take whatever text, images etc and use Anthropics own checker to determine if it has ‘touched’ any of its models. There is important nuance here, and Anthropic seems to prefer the word 'signals'.
But there is a GULF of difference between someone lazily asking Claude to generate a literal entire 100 page proposal and not even bothering to proofread it vs. someone who has been working side-by-side collaboratively with Claude for days or weeks using it to help refine & polish something that is inherently and originally theirs. Both would be watermarked.
The former you might reasonably be annoyed at if you’ve paid someone a lot of money to do it. The latter, I think it’s hard to be annoyed at someone who uses AI to genuinely improve things. Denying or outright banning the use of AI is like trying to fight the replacement of horses with cars in the early 1900s.
Will a Claude watermark be able to distinguish between the two scenarios? Unlikely.
So what’s the use / point of it then? Hmm… I can see why Anthropic have done it. It’s a useful way to say “we implemented traceability ourselves” and indeed they very clearly state this is to help ‘comply with their legal obligations’ - whether that traceability information is actually of any use isn’t really their problem - they’ve raised the ‘signal’, interpret it how you will.
Honestly, if Claude was the only frontier model out there, I reckon it would say 99% of content has at least been ‘touched’ by an AI. Such is how prevalent its use is - in personal situations, professional situations, you name it. If you’re NOT using AI, then....where have you been?! 😉
Anthropic have been careful to say all of these points already. It’s not foolproof, and as quickly as watermark technology originates, countermeasures and ways to combat it will develop in earnest.
But exactly what war is everyone fighting here and why? I understand some specific scenarios that would be beneficial - for example photography competitions that are strictly non-AI (this was literally a news story the other day). But my (blunt) view is in the majority of situations it’s like trying to watermark if someone’s used a calculator to do a math sum. Why can’t they? If it’s helped them get to the answer quicker and deliver value to the end user, then who exactly is against that?
Now exactly how that valuable 'AI-enabled' work is priced is another but important and linked matter. But I'll reserve that great debate for a blog coming soon!
Let me know your views!