On the latest episode of The Briefing by the IP Law Blog, Scott Hervey and Josh Escovedo discuss the lawsuit filed earlier this year by global visual content creator Getty Images against startup technology company Stability AI for allegedly scraping over 12 million photographs from Getty’s portfolio without consent or compensation.
Getty Images claims that Stability AI copied its photographs and associated text and metadata to train its Stable Diffusion model, which uses AI to generate computer-synthesized images in response to text prompts. The alleged act is said to constitute copyright infringement, provision of false copyright management information, removal or alteration of copyright management information, trademark infringement, unfair competition, trademark dilution, and deceptive trade practices under Delaware law.Getty Images earns revenue by licensing the right to use its assets to businesses and consumers. Stability AI, on the other hand, created an image-generating platform called Stable Diffusion, which relies on various images to generate a synthesized version of the user’s requested image. Getty Images claims that its assets are highly desirable for use in connection with AI and machine learning, and it has licensed millions of digital assets to technology innovators for various purposes.
However, Stability AI allegedly did not attempt to negotiate a license with Getty Images for the content but instead copied the images without Getty’s consent, which, according to Getty, constitutes clear copyright infringement.
Many AI technology companies believe that using copyrighted materials to train an artificial intelligence platform constitutes fair use under United States copyright law. The fair use doctrine takes several items into consideration when determining whether a use was fair. The purpose or nature of the use and whether the allegedly fair use has an effect on the market for the protected work are two of the most significant factors. It remains to be seen how the courts will apply the fair use doctrine in the context of AI technology, but it may be the case that using copyright-protected works to train a platform constitutes fair use, while using it to generate new content may not.
Clarity to Come, Perhaps
It is a complex issue dealing with sophisticated technology. However, the Supreme Court’s decision in the Andy Warhol Foundation for the Visual Arts v. Goldsmith case, currently before the Court, will likely affect some of the issues in this context. The Court will seek to provide clarity on the scope of the fair use doctrine and how different a work must be from the original protected work to be considered transformative and therefore not infringing. The decision will have far-reaching implications for creators, as well as some AI technology companies.
Getty Images’ lawsuit against Stability AI highlights the issue of copyright infringement and fair use in the context of AI and machine learning. While the fair use doctrine is a multi-factor analysis that takes several items into consideration when determining whether a use was fair, it remains to be seen how the courts will apply it in this context. The Supreme Court’s decision in the Andy Warhol case may provide clarity on the scope of fair use and how transformative a work must be to be considered non-infringing. The decision will have significant implications for creators and AI technology companies alike.
Join the Discussion
4 comments so far. Add my comment.
AnonApril 9, 2023 09:57 am
Nice to hear that you work in the encryption field (I have also done work in that field).
That being said, I would not base a determination of any inventiveness of AI based merely on one field. We both know that the modicum of inventiveness is a spectrum, rather than a sharp line.
As to “The query to ChatGPT has to be so specific that it indicates that the user already knew what he/she was looking for.” This is simply not the dispositive statement that you appear to trust it to be.
“Knowing what one is looking for” may well include “looking for an invention that the seeker has not made.”
Both in DABUS (taking the facts as provided) and in the larger environment (now with each aver growing capability of published AI), the easy and proper way to approach this is to simply apply the legal definition of inventor to the human in the loop.
If that human cannot meet that legal definition, it is simply improper to shift to that human the act of invention just because an AI was involved.
This is the crux of my “second person in another room opening a black box and reading an invention placed therein” example (now, getting to be about two years old).
As I have also provided, the pre-AI version of attempting to claim as one’s own the inventions of another was clearly prohibited in the (unfortunate) era of slavery in the US.
It really makes no difference as to this second person ‘opening the black box’ as to HOW that item ‘got’ into the black box.
Lab JedorApril 7, 2023 01:50 pm
Anon. I agree with your observation on volume. However, volume is a great motivator to infringe. Using a single or even a dozen Getty images is not that expensive. For $12.50 a piece that is less than $150. You can pay that or ignore it, in the conviction that Getty may not pursue infringement on that. But at 12 million instances, this becomes pricy. And, using Google’s escape from payment, a great incentive for Stability AI to try the same.
Strictly speaking, they may probably not literally use images or image parts from Getty, but create their own library using labels extracted from the Getty training. Technically there are enough arguments to confuse judges and justices and experts alike, why strictly and narrowly speaking there is no detectable use of Getty images.
Thanks for the link. I had not read this post yet. I found that related to inventions, ChatGPT by its nature and self acknowledgement is a hindsight oriented machine and is unable to come up with new ideas for instance gleaned from related areas. I tried it with my own inventions in cryptography and it came up completely empty, while it is very well versed in the subject area. It even generates complete C programs in encryption. But even when prodded with suggestions it fell back on well known approaches. It seems (for now) to be unable to evaluate let alone generate a new idea and even be close to being a co-inventor. The query to ChatGPT has to be so specific that it indicates that the user already knew what he/she was looking for.
AnonApril 7, 2023 07:30 am
I genuinely appreciate your posts as they typically show thoughtfulness and a clear distinction between feelings and an objective view of the law. You also alight upon the case most on point here (and that case does go against what I infer your feelings to be on the matter).
See also: https://patentlyo.com/patent/2023/04/should-patent-attorney.html#comment-849274
One probably should keep in mind that the Supreme Court is not constrained to actually make cohesive decisions, and the Warhol case may make yet another IP Gordian Knot.
(ps: sheer volume – be it millions or thousands – is simply not dispositive, even though there may be an emotional tug from the difference in scale)
Lab JedorApril 5, 2023 08:41 pm
“copyright-protected works to train a platform constitutes fair use, while using it to generate new content may not”
AI uses the training data to generate new content. This branch of AI doesn’t work without training. Fair use to me is using some Getty Images to check if the AI engine can recognize the content. Copying and using 12 million images from one owner Getty is technically an extraordinary interpretation of “fair use.”
However, Google’s copying and use of 11,500 Java APIs owned by someone else was declared “fair use” by SCOTUS and that pretty much sets the stage. So, why not 12 million photographs. Great work SCOTUS!
Like Google AI Stability concluded: you mean to say I have to take my own 12 million pictures (or Google, you want me to develop my own 11,500 APIs?). Do you know how much that will cost? While I can grab the …. (fill in: images, APIs, texts, code, data, videos, patents) right from the Web.