Watermarks on Generative AI Art … and Copyright

I used AI to generate the image below. (I wanted new wallpaper for my laptop, but this was a bit more sci-fi than I had in mind.) Take a close look. You’ll see almost a dozen faint, diagonal lines of text. They look like watermarks.

generative AI artwork w/ faux watermark
I call it, “Garden Moonscape with Faux Watermarks.”

I’ve seen these faux watermarks on other images I created using AI. But I didn’t ask for watermarks in my prompts. How did they get there?

Training Data with Lots of Art Someone Tried to Protect

Artists add watermarks to online versions of their work so they can display them without giving them away. The visitor has to pay for a “real” copy: a version without watermarks.[mfn]See this example on dreamstime.[/mfn] It looks like the training data for my art-generation AI includes lots of this watermarked art – pulled from the web. It includes so much that the machine thinks watermarks are features of certain types of art (e.g., surreal landscapes). As a result, when the AI generates that type of art, it often includes images that look like watermarks.

If I’m right, the AI not only draws on copyrighted work, which we already knew. It also ignores artist attempts to protect that work through watermarks.[mfn]See also the vertical text in the image’s lower left corner. It looks like a faux credit or copyright notice – probably generated the same way.[/mfn]

Copyright, Fair Use, and Food for Thought

Of course, the conclusions above only apply to the generative AI product I’ve been using. But I’d be surprised if it’s the only one with watermarked art in its training data. And other forms of generative AI raise similar issues. I recently saw an AI-generated contract that looked a lot like one of my own form contracts posted here at TechContracts.com. I could be flattering myself, but that sort of heavy reliance on a particular set of texts wouldn’t be surprising.

None of this means AI vendors infringe copyright. Many legal scholars argue that generative AI’s use of training data is “transformative,” so that it qualifies as fair use under copyright law. Let’s hope the courts tell us soon whether that’s right.

But even if so, the faux watermarks make you think.

All of this reinforces one of the warnings in my recent issue-spotter on artificial intelligence contracts (bullets 20-22): generative AI outputs create significant IP infringement risks. Vendors should hesitate to grant IP warranties and indemnities related to outputs. And customers should use generative AI outputs with care.


© 2023 by Tech Contracts Academy, LLC. All rights reserved?

Share the Post:

Related Posts