Post AVlNhBGNUxgHVAlX84 by Arpie4Math@mathstodon.xyz
(DIR) More posts by Arpie4Math@mathstodon.xyz
(DIR) Post #AVlNhBGNUxgHVAlX84 by Arpie4Math@mathstodon.xyz
2023-04-07T01:05:33Z
1 likes, 1 repeats
Andersen v. Stability AI Ltd (3:23-cv-00201) District Court, N.D. California#DeviantArt, #StabilityAI, and #Midjourney are expected to reply (by which I expect motions to dismiss) on April 18 to this proposed #ClassActionLawsuit #lawsuit over #AI tools like #StableDiffusion violated #copyright in the scraping of images or the generation of new images.Stability AI's UK parent, Stability AI, Ltd. which is the lead defendant in the case caption, doesn't have filings indicating that they have been served or have waived service yet.To my mind, the act of web scraping itself seems harmless as it is literally necessary for the operation of libraries and search engines.Generating images in the style of artist X seems beyond the realm of copyright law. And the claim that generated images are collages created in violation of copyright seems to be deficient in the facts. Specifically, nothing that StabilityAI does seems to be akin to maintaining a library of art and creating collages from it.There seems to be a possible copyright issue with hypothetical image replication, but not only is that an unwanted behavior, but it has also been found to be exacerbated by unintentional overtraining on multiple copies of an image in the web-scraped inputs. "That's a [rarely seen] bug, not a feature." seems like a defense to this layman #IANAL. That it's rare and hard to reproduce for some of the generative AI products hurts the case for a class action where each plaintiff needs to be in basically the same boat.Also, will 47 U.S. Code § 230(c)(1) immunize online generative AI products for outputs they produce at the prompting of information from a third-party user?https://www.courtlistener.com/docket/66732129/andersen-v-stability-ai-ltd/
(DIR) Post #AVlNhExNlWrYxPuQ4m by Arpie4Math@mathstodon.xyz
2023-04-07T13:16:45Z
0 likes, 0 repeats
Andersen v. Stability AI Ltd (3:23-cv-00201) District Court, N.D. Californiatl;dr Artists sue #DeviantArt, #StabilityAI, and #Midjourney over purported #copyright violations inflicted by #StableDiffusion #AI and related tools.https://mathstodon.xyz/@hewittlaw@mastodon.social/109789753621184612The original complaint (¶¶ 80-81) has a nice cartoon of stable diffusion as a kind of interpolation, but since the cartoon uses a 2-D "space of all images" with but two images, basing your lawsuit on such a cartoon seems a bit naive and misleading. The cartoon came from the left part of Figure 8 in https://arxiv.org/abs/2006.11239 . It is not very large. Blink and you would miss it. But it lacks the context of what the computer knows, which is the discovered relations between images.¶ 76 links to that paper but puts the date as Dec 2020 when there are actually two similar preprints of the paper by Jonathan Ho, Ajay Jain, and Pieter Abbeel [v2] dated 2020-12-16 and [v1] dated 2020-06-19. The published version is in Neural Information Processing Systems, 2020. https://papers.nips.cc/paper/2020/hash/4c5bcfec8584af0d967f1ab10179ca4b-Abstract.htmlThe plaintiffs' lawyers have arranged for a late January media blitz of misinformation about the technology with a companion website that boosts the plaintiffs' story: https://stablediffusionlitigation.com/But there is what reads to me as a more mathematically-informed view suggesting that plaintiffs are using only reasoning by analogy with little foundation in what the computer is doing: http://www.stablediffusionfrivolous.com/There is no database of images. The rare toy example of overfitting ( https://arxiv.org/abs/2301.13188 ) does not mean an example of near-duplication proves that StableDiffusion is some form of lossy image compression.
(DIR) Post #AVlNhGjHAwleSFKBMm by Arpie4Math@mathstodon.xyz
2023-04-19T04:09:59Z
0 likes, 0 repeats
Andersen v. Stability AI Ltd (3:23-cv-00201) District Court, N.D. Californiahttps://www.courtlistener.com/docket/66732129/andersen-v-stability-ai-ltd/Today we have the expected motions to dismiss.#StabilityAI Doc 51: https://storage.courtlistener.com/recap/gov.uscourts.cand.407208/gov.uscourts.cand.407208.51.0.pdf• Many required copyright registrations not in place• #StableDiffusion does not store images, make collages; plaintiffs concede it does not output close copies• There is no claim of even a single act of direct or vicarious infringement, nor did they allege that copyright management information was stripped out of even a single image• "publicity rights in images" = "copyright holder rights", you can't just relabel them and double-dip• StabilityAI is not trading on plaintiff's identities• Copyright law preempts claims that might otherwise exist as unfair competition• You have to win or lose on your claims as stated; there is no such thing as requesting declaratory judgement in a complaint that your other claims are to be found true#DeviantArt Doc 49: https://storage.courtlistener.com/recap/gov.uscourts.cand.407208/gov.uscourts.cand.407208.49.0.pdf• Complaint does not specifically allege that Deviant Art did the bad things they accuse "defendants" of• Derivative works must be substantially similar in order to be infringing and Plaintiffs allege the opposite• DeviantArt did not beach its own terms of service, so there can be no breach of contract claim against it#Midjourney Doc 52: https://storage.courtlistener.com/recap/gov.uscourts.cand.407208/gov.uscourts.cand.407208.52.0.pdf• "there is no conceivable class that might satisfy the requirements of Rule 23 in this case."• "Plaintiffs seek to do the impossible by applying a single class definition to four distinct violations of law, each with its own elements and individualized, fact-intensive inquiries that overwhelm any conceivable common issues."Also some CA motions to strike
(DIR) Post #AVlNhIuh3T5ZEEgKRM by Arpie4Math@mathstodon.xyz
2023-04-19T04:40:39Z
0 likes, 0 repeats
Doc 58 is an amended version of Doc 51 from #StabilityAI https://storage.courtlistener.com/recap/gov.uscourts.cand.407208/gov.uscourts.cand.407208.58.0.pdf• Footnote 1 changed from"Stability AI consents to personal jurisdiction in this District solely for purposes of this action. Stability AI’s participation in this litigation should not be construed as consent to personal jurisdiction in this District for any other purpose or as a waiver of its right to contest personal jurisdiction in any other action."to"Stability AI consents to personal jurisdiction in this District solely for purposes of this action and any others raising substantially similar claims arising under the Copyright Act, DMCA and/or Lanham Act. Stability AI’s participation in this litigation should not be construed as consent to personal jurisdiction in this District for any other purpose or as a waiver of its right to contest personal jurisdiction in any other action."I'm not a lawyer, but this feels like there may be some land mine in case law if they don't phrase their agreement to personal jurisdiction is a specific manner.Doc 50 is a motion to strike by #DeviantArt the California Right-Of-Publicity claims which they think are forbidden by California Anti-SLAPP law and the #1stAmendment to the US Constitution. It's such a good idea #Midjourney and #StabilityAI jump in (Docs 54 and 59, respectively.)https://storage.courtlistener.com/recap/gov.uscourts.cand.407208/gov.uscourts.cand.407208.50.0.pdf
(DIR) Post #AVlU8kq3w8uNNgzKbY by Meier@socialgoblins.club
2023-05-18T02:07:03.597205Z
0 likes, 0 repeats
I personally believe that Stability AI did nothing wrong. The source of the training data is legit. Also how stable diffusion is trained appears transformative imo. https://youtu.be/J87hffSMB60 I also don't trust a California court to get this correct. I wonder how this will end up.
(DIR) Post #AVlswuZoP4iYxgpwOW by Arpie4Math@mathstodon.xyz
2023-05-18T05:10:54Z
1 likes, 0 repeats
@Meier Law Professor @pamelasamuelson has weighed in on the two #StabilityAI lawsuits and has opined that the California one (Andersen v. Stability AI Ltd., 3:23-cv-00201, (N.D. Cal.), https://www.courtlistener.com/docket/66732129/andersen-v-stability-ai-ltd/) is the easier one to defend because the plaintiff shoots themselves in the foot with respect to copyright infringement (Complaint, Doc 1, p. 21, ⁋ 93).More at: https://mathstodon.xyz/@berkeleylaw@sfba.social/110379592434811534The Delaware care (Getty Images (US), Inc. v. Stability AI, Inc., 1:23-cv-00135, (D. Del.) https://www.courtlistener.com/docket/66788385/getty-images-us-inc-v-stability-ai-inc/) is harder because it is multi-pronged, better written, and is not in the 9th District where many nice precedents with respect to web crawling were written.Also, if it goes to a jury, won't the layman place weight on any allegations there were attempts to sand off Getty Images watermarks or demonstrations that garbled watermarks sometimes are generated?What's harder to imagine is a process to agree on what the fair price for training a model on a curated set of data like the Getty Images which are particularly valuable due to the extensive textual labels which are inputs to the training of generative AI which is responsive to natural language prompts.Just what would they be paying for? I'm not sure, but I think it's not within the realm of copyright. I have some recent related thoughts: https://mathstodon.xyz/@Arpie4Math/110373253644824916
(DIR) Post #AVlswv9yEddolpkokq by Meier@socialgoblins.club
2023-05-18T06:45:01.482101Z
0 likes, 0 repeats
Thanks for all the info. :)I'll have to keep an eye on this.