Generative AI raises a host of legal questions and concerns. Some of these questions will challenge existing legal rules and require new laws and policy frameworks. Others have answers that are quite well settled, notwithstanding the new AI context bringing attention to them.
This explainer is an evolving project to provide everyone with the types of answers that legal experts might informally provide to each other. Each question includes a short response to help you understand the most likely answer in the most likely cases. That answer is then given a confidence score on a scale of 1-5, where 1 is “it really is anyone’s guess,” and 5 is “there are decades of case law to support this answer.” These are followed by a long answer that explains the answer’s legal context and how various factors could impact it.
Short answer: Probably yes. (Confidence: High.)
The non-consensual use of copyright works to train GenAI systems has been one of the most visible controversies surrounding GenAI lawsuits. If you’ve made something, the argument that you can prevent it from being used to train AI feels simple: (1) you own your copyrights; (2) copyright law grants you exclusive control over any attempt to make a copy of your work; and (3) when GenAI systems train on your works, the works are copied into its memory. Seems like a pretty slam dunk case, right?
It would be, except for the fair use doctrine. Section 107 of the US Copyright Act states that if the use of a copyright-protected work is a fair use, it is a legal use – even if it is done without any compensation, credit, or consent to the copyright owner.
Fair use has been around for centuries, and there have been hundreds of decisions finding fair use over copyright owner objections. The preamble of Section 107 includes examples of the types of non-consensual uses that are likely to be fair, such as criticism, comment, news reporting, teaching, scholarship, or research, but those are just a few of the many uses courts have found to be fair. Fair use has a four-factor test (which is too complex to explain here), but it's the underlying question behind it is whether the non-consensual use provides a significant enough scientific or social benefit to outweigh any harm it might cause to copyright owners in the process.
For computer programs, there have been dozens of fair use cases. Every time you access a digital image, video, document, song, or even a webpage, your computer has to copy that data into its memory in order to perform or display it for you. When these uses are done with permission, there is no copyright conflict. However, anytime someone loads content into a computer without permission, the courts usually look to fair use to decide if it should be legal or not to do so.
As to the question of whether training AI systems is fair use, US courts have addressed similar situations for over 30 years in dozens of cases, all of which overwhelmingly resulted in findings of fair use when the non-consensual copying was done for the purpose of improving computer systems, scientific knowledge, market competition, access to information, or some other socially beneficial purpose. So it is likely that they will find the same here.
That said, the fair use analysis of AI training AI datasets will likely turn on the answers to the two key questions:
Will courts identify a distinction between “copying” and “learning”? When a human reads a book or studies a painting, we usually think of them as learning from that work, not merely copying it to their brain. For example, going to the library to scour books about art or inspiration for your own novel doesn’t infringe copyright, even though you might remember dozens or even hundreds of examples from what you read. Similarly, machine learning (ML) models are often described as being “trained” on data, meaning they load copyrighted works into their memory as sources to analyze for patterns of production, not merely objects to be replicated. To the degree that courts recognize the “learning” approach to ML, they are more likely to find training to be fair use. However, if courts see training as simply reproducing copyrighted works for the purposes of replicating them, it is less likely that they will find the copying protected by fair use.
Are copies of the works in the dataset stored in the model? The legal community does not have a clear understanding of how individual works from a dataset are stored and represented in a trained model. If the individual works are represented in some specifically identifiable way, that might strengthen the case for copyright infringement. In contrast, if the works are significantly abstracted in the model, it might weaken that same case, as it is less likely that a court will find the abstracted version to be a “copy” in the same sense as when one downloads a copy of a file.
Short answer: Practically-speaking, probably not. (Confidence: High.)
Many tech companies creating GenAI products today do not disclose their sources of training data. But new tools are popping up that allow us to investigate whether our personal data has made its way into these training datasets. Examples include the “ Have I Been Trained ” tool, as well as the earlier Exposing.ai project. Through these tools, we have learned that many GenAI products are trained on massive datasets containing personal information scraped from popular websites, including social media platforms, video and photography sites, and online encyclopedias. According to one study extracting personal data from training datasets for early large language models like ChatGPT-2, GenAI products often produce outputs that contain peoples’ names and some form of contact information (like a phone number or social media handle).
Journalists have been reporting more recently on the ways that newer LLMs regurgitate personal information that is likely contained in their training datasets. OpenAI explains clearly how “[a] large amount of data on the internet relates to people, so our training information does incidentally include personal information .” One artist even found that her private medical records containing photos of her facial condition were in LAION, a foundational dataset for many GenAI products.
This is not unique to GenAI, however. Many companies are following the playbook of earlier AI companies like Clearview AI, a controversial facial recognition company that built its powerful surveillance tool from scraping billions of photos containing peoples’ faces from the web. When Clearview did it, several platforms whose users’ photos were scraped and converted into training data sent cease-and-desist letters demanding Clearview stop taking their users’ data in violation of their User Agreements. Now, companies like X (formerly Twitter) are taking the same action against GenAI companies for taking their users’ data to train GenAI products.
How can you remove personal data from training datasets when you do not know if your data is even there?
One way may be to ask nicely. For example, Mozilla has set up a petition asking Microsoft to divulge if the company will use personal data to train its AI models per its new Service Agreement (effective September 30, 2023).
Another way is to sue. In J.L. v. Alphabet , the plaintiffs are Github users who have long uploaded their computer code to the platform, only to allegedly have that code ingested as training data for Copilot. These plaintiffs allege that Github and OpenAI violated the CCPA by “incorporating” their personal information into Copilot with no way to alter or delete it, and with no way to share that data with them upon request. Compl. ¶ 227. Although the term “incorporating” is vague, this could be read to encompass data collected from Github that was used to train Copilot, specifically open source codes but not the personal information related to the codes (even if that information was incidentally part of the training dataset exfiltrated from Github to Copilot). But this is a strange claim under CCPA, which governs personal information (defined in § 1798.140(o)(1) of the CCPA)--a court will likely find it a stretch to argue that these users’ open source computer codes count as the kind of personal information protected under the CCPA.
But likely the most promising method to learn whether your personal data is in GenAI training datasets and to seek its deletion is to send a request under the CCPA, which is only available for California residents. As part of an investigation into this right, Knowing Machines is currently testing out how various companies offering GenAI products respond to personal data requests under CCPA. So far, it is clear that companies respond in idiosyncratic ways and often ask for further information regarding the request and the requestor. This investigation could reveal the practical futility of data requests–including deletion requests–under CCPA as it currently stands, suggesting that the California Privacy Protection Agency should develop guidelines for compliances as well as enforcement mechanisms for companies that fail to comply.
For now, there is no foolproof way to force companies to disclose whether, how much, and what kinds of your personal data have been used as training data for their GenAI products. Even if there was a way, however, it is currently technically impossible to fully remove data from training datasets without influencing deep learning models in unpredictable ways. This issue is of regulatory concern for big players like Google, which announced a “ Machine Unlearning Challenge between mid-July and mid-September 2023. According to Google, the challenge involved “a realistic scenario in which an age predictor has been trained on face images, and, after training, a certain subset of the training images must be forgotten to protect the privacy or rights of the individuals concerned.” In a few months, we may see some technical progress on this issue; until then, companies may be unable to comply with data deletion provisions in laws like the CCPA.
Short answer: Probably yes. (Confidence: High.)
Defamation law allows you to sue someone who says or writes something about you that’s false and harmful to your reputation. You can use a lawsuit to stop them from repeating the falsehood or force them to compensate you for the harm they caused.
In the United States, defamation law differs a bit from state to state, but the basic elements of a defamation claim are substantially the same. You typically have to prove: 1) a false statement that purported to be fact about you; 2) publication of the statement to a third party (i.e., anyone but you); 3) some level of fault on the part of the individual or entity that published the statement; and 4) reputational harm.
But suing a GenAI company for defamation would not be a standard case. The most likely scenario would involve a suit against the company for a “hallucination”—something made up—that the GenAI returned to a user or users. Its success would hinge on a number of currently unresolved legal questions. Fortunately, there are at least two live court cases that might start to provide some answers.
If you tried to sue a GenAI company, the first hurdle that you’d have to overcome is convincing a court that the harmful hallucination constituted a statement of “fact.” Given that GenAI tools are simply “advanced pattern recognizers” trained to “predict the next word in a sentence,” it might be hard to see how their outputs could ever be thought of as supplying “facts” or “truth” in any meaningful sense.
But how a GenAI tool is trained and actually operates under the hood will matter less for this step of the defamation analysis than how the GenAI output at issue is presented in the context of a specific use case. Would an average user read the output and reasonably assume it’s making a statement of fact? In Walters v. OpenAI , a Georgia state court case arising out of an inaccurate ChatGPT output, a judge recently held that “multiple indicia” suggested that a reasonable reader would not have understood the output at issue to contain factual assertions, including the many general and specific warnings about accuracy and reliability that OpenAI and ChatGPT provide to users.
A closely related issue is that of “publication.” Here, you’d have to show that the harmful hallucination was somehow shared with or “published” to someone else. This is a pretty intuitive requirement—after all, if your worst enemy typed a fib-filled diatribe about you into a Word document that no one ever read, where’s the harm to you or your reputation?
OpenAI defended itself against the defamation case filed against it by basically arguing that ChatGPT is the Word document. The company referred to its Terms of Use —which assign users rights to their ChatGPT outputs and states that users are “responsible” for them—to describe ChatGPT as nothing more than “a private drafting tool” that “helps [a user] write or create content owned by the user.”
The court in OpenAI’s case did not ultimately address the question, but this is a novel argument that other courts will likely reject, particularly when it comes to publicly available GenAI tools like ChatGPT and Bing. Defamation law has traditionally recognized that publication “encompasses any communication of the idea to a third party other than the plaintiff [and defendant] themselves,” no matter who “owns” the communication. As long as a company operating a GenAI tool has shared the harmful hallucination with someone other than itself—even if it’s just the individual user who prompted the tool—you’re likely to clear this hurdle.
Then there’s the issue of “fault.” Generally speaking, you can’t hold a person or company liable for defamation unless you can show that at the very least they acted negligently in sharing the falsehood. (And if you’re a public official or public figure hoping to sue, you’d have to show that they acted with an even greater level of fault, what’s known as actual malice —that is, with knowledge that the statement was false or with reckless disregard as to whether it was false or not.)
What evidence would you need to prove that a GenAI company was negligent (or worse) in allowing its GenAI tool to output a harmful hallucination? Is it enough to point out that the GenAI company generally knows, or should know, that its tool can sometimes produce false information? Or must you somehow show that the GenAI company was specifically aware, or should have been specifically aware, that its tool produces the particular hallucination that harms you?
Significantly, the court in the OpenAI case held that, at the very least, OpenAI was not at fault merely because the company was aware that ChatGPT can provide inaccurate outputs. But because the plaintiff there failed to offer any more specific evidence to show why OpenAI was at fault, the court did not have occasion to analyze what steps OpenAI (and GenAI companies like it) might need to take to prevent defamatory hallucinations and thereby avoid defamation liability.
For now, it’s worth noting that you’re likely to clear this hurdle—that is, a GenAI company will likely be found at fault—if the GenAI company has been put on notice that its tool is producing a specific harmful hallucination about you, but then does nothing to try to rectify it. This may have major implications for the design of GenAI tools going forward. Will GenAI companies need to implement notice-and-takedown procedures for people to report harmful hallucinations? Will they need to finetune their filters or erect greater content-moderation guardrails against defamation?
Whether you can successfully sue a GenAI company for defamation will ultimately come down to the facts of your particular situation. Did the GenAI tool produce the harmful hallucination in a context, like a search engine, where users reasonably expect factual responses? After discovering the GenAI tool’s capacity to produce the harmful hallucination, did you or someone else notify the GenAI company? If the answers to these questions are both yes, then you just might have a defamation claim against the GenAI company.
Short answer: Probably yes. (Confidence: High.)
It is very difficult to protect others from copying the style of human artists, as copyright law covers specific works of art, not general styles, methods, or genres. You can have trademark rights in style if it is part of your unique brand, but only if most consumers would associate it with you as a commercial artist by name and the GenAI outputs create confusion in the marketplace as to who made which works of art. And while you do have rights to control your image, name, and identity under your “right of publicity,” those are generally limited to exploitation of your personhood and not the works you create in the world. Deepfakes, on the other hand, might run afoul of right of publicity laws if they commercially exploit your image, voice, or other aspects of your identity.
Short answer: Probably Not. (Confidence: High.)
Right now, we’re in a moment of flux about who, if anyone, owns GenAI outputs. Traditionally, copyright law assigns ownership in a work to the author/artist. If the author/artist is an employee, the employer typically owns it, and when they are an independent contractor, who owns it typically gets spelled out in the contract.
With GenAI, however, the US Copyright Office has decided that most GenAI outputs are “authored” by the machines that create them, not the humans who prompt them. It has also decided that only human authors should own copyrights. This is primarily because of how copyright law is described in the US constitution and the Copyright Act, along with court cases specifying that the purpose of copyright law is to provide incentives for authors to create works. Since GenAI systems are machines that don’t create works based on financial or property-like incentives, there is no reason to grant them copyrights. This is similar to other cases involving non-human authors, both spiritual and simian .
As for the humans that prompt them, the Copyright Office has declined to consider them as authors unless they can show that they played a direct role in the creation of the outputs in an expressive or aesthetic sense, not just prompting the model with concepts or ideas or even specifications. For example, the Copyright Office recently denied one applicant any copyright in a GenAI image – despite prompting the model over 600 times – because he failed to show that any of the prompts contributed a specific expressive or aesthetic component to the output.
Short answer: Probably yes. (Confidence: High.)
When you use GenAI products, you generate data that companies can use to further enhance their products. This is a longstanding practice with most tech products – tech companies often design their products to maximize the amount and variety of your data that they can collect because the business insights these data can produce are incredibly valuable. The data broker industry that trades in user data is valued at roughly $260 billion .
This industry-wide thirst for your data has created the “Big Data” phenomenon of the past decade. In turn, this Big Data forms the core of GenAI: GenAI products are the latest application of advanced machine learning techniques for processing vast amounts of data to train algorithms to produce novel outputs. This means that data privacy issues with GenAI products aren’t new and evoke similar issues with earlier machine learning technologies . They compound past failures to pass meaningful data regulations that could have protected your personal data long before GenAI products existed.
But what does “your data” mean in this context? There are a lot of different types of information about you that GenAI products collect and process when you interact with them, and the specifics depend on which product you are using. Generally, there are two buckets of “user data”:
Inputs : Data that you input into the product, including the content of your written prompts, uploaded images, voice recordings, or other inputs. For example, if you input the prompt, “If my period is lasting longer than usual, and I am a 45 year old woman, am I perimenopausal? What are the signs of perimenopause in women with endometriosis?” into ChatGPT, your data includes personal information about your health, your gender identity, and your age. If you input your selfies into a text-to-image product like Stable Diffusion, your data includes any information that the product can ingest about your facial identity, as well as objects, people, or places discernable in the background. Inputs like selfies and facial images also include biometric data, or information about peoples’ unique biological features (here, facial features).
Metadata : Data that captures how you access and use the product, including things like the type of device you are using, your geolocation data, your name and/or phone number used to create an account, and any transaction history and/or account payment information if you purchase a subscription or upgrades to the product. In addition to the data that companies directly collect from their product, they can also collect your data when you interact with a separate website that has integrated their GenAI product plug-in . If you have a ChatGPT subscription and use Expedia or Instacart , for example, you can use each website’s individual chat feature to plan your next trip or generate new recipes. Websites with these plug-ins send data to OpenAI about how you interact with the product, which OpenAI can use to train ChatGPT further to improve the relevancy of its outputs.
Because we lack comprehensive federal data protection laws in the U.S. that constrain companies’ ability to collect your data, companies independently decide how much of your data they will collect and for which purposes. We rely on them to disclose their user data practices in privacy policies that can be ambiguous and even contradictory , but that also give companies legal cover when users “consent” to privacy policies in order to access and use GenAI products.
For example, to use ChatGPT, users agree to OpenAI’s privacy policy which lists a variety of personal information OpenAI collects from users, including their name, contact details, device identifiers, geolocation data, and more. The policy explains that OpenAI can disclose this information “to [their] affiliates, vendors and service providers, law enforcement, and parties involved in [business] Transactions,” without providing more details about which parties they do business with, who their affiliates are, and under what circumstances they interact with law enforcement.
Similarly, Lensa app developer Prisma Labs explains in its privacy policy how it collects personal data directly from users as inputs into Lensa, metadata about how users are interacting with the app (including your device information, app and country information, in-app purchases, and app usage), and from “third parties, for example, our service providers, partners and vendors.” When you use Lensa, Prisma collects the photos and/or videos of yourself that you upload to the app, information regarding your gender identity, any account information (for example, if you link your Apple ID or Google Account email address to Lensa), and any contact information you provide to the company (including your name, postal address, email addresses, and phone numbers). Even if you delete your account, Prisma may still share your personal data with third parties, including cloud providers Amazon Web Services and Google Cloud Platform and email delivery providers; the personal data Prisma can disclose to them may include your facial images and image data, defined as including “information about your facial position, orientation and topology”.
There are a handful of recent lawsuits challenging these companies’ ability to collect data while you use GenAI products, however. P.M. v. OpenAI and A.T. v. OpenAI , for example, both accuse OpenAI of violating wiretapping and anti-hacking laws. They argue that OpenAI effectively hacks users' platform access, exceeding users' authorized access for OpenAI when using ChatGPT, and intercepts their private information (i.e., user data) without their knowledge or consent.
These claims are likely to fail. Although the plaintiffs claim that OpenAI is unlawfully intercepting these communications by collecting them without users’ consent, ChatGPT users do technically consent to OpenAI’s Privacy Policy and data collection practices when they use ChatGPT, and OpenAI’s Policy is explicit about the types of user data the company may collect when users interact with their product.
While plaintiffs claim not to have notice from third party websites that have integrated ChatGPT plugins, these websites generally have their own privacy policies that explicitly describe the companies’ ability to share user data with “business partners” and “service providers,” that would presumably include OpenAI.
For example, Kayak’s privacy policy states explicitly that the company shares user data with “other business partners,” in order to provide travel booking services; this is written broadly enough to encompass partners like OpenAI, whose ChatGPT plugin feature presumably helps Kayak provide travel booking services to users.
GenAI companies like OpenAI are a far cry from traditional wiretapping violators that intercepts private communications like phone calls without the consent of one or both speakers , nor are they akin to computer hackers (or free culture activists with technical chops ) who target rich databases prone to compromise . In order to use ChatGPT and other GenAI products, users constructively consent to the collection and sharing of their data by clicking “I Accept” to the terms of use and privacy policies unilaterally drafted by providers. Although these notice-and-consent regimes largely fail to protect user data , they have been accepted by courts as legally valid and enforceable.
Biometric data might be harder for companies to collect without complying with biometric privacy laws. In Flora v. Prisma Labs and P.M. v. OpenAI , the user-plaintiffs allege both companies violated several requirements under the Illinois Biometric Privacy Act (BIPA) that protect Illinois residents’ sensitive biometric data, specifically their facial image data in these cases.
Biometric data can include information about what you look like (i.e., what your face looks like, your irises, your thumbprints or palmprints, etc.), genetic data (which some laws include under the “biometric data” umbrella), as well as other types of personally identifiable information (PII) (under which some laws include biometric and genetic data). These are highly sensitive forms of data that many federal and state laws regulate in a variety of ways, because once your unique biometric data is compromised , it is nearly impossible to fix that breach (at least without undergoing extreme physical or genetic adjustments).
BIPA is a prime example of such regulations. It requires companies that collect biometric data to comply with several procedures before, during, and after collecting users’ biometric data. Companies must: (1) have a written policy establishing a schedule for destroying biometric information within a set timeframe (§ 15(a)); (2) first inform users in writing of data collection, the specific purpose and length of the collection, and receive a written release from users before they “collect, capture, purchase, receive through trade, or otherwise obtain” biometric data (§ 15(b)); (3) refrain from selling, leasing, trading, or profiting from users’ biometric data (§ 15(c)); (4) refrain from disclosing or disseminating user biometric data without user consent, among other requirements (§ 15(d)); (5) and must store, transmit, and protect biometric data from disclosure using “the reasonable standard of care” within that company’s industry or in a way that is consistent with how the company treats other sensitive data it possesses (§ 15(e)). Plaintiffs need only to show that a company has failed to meet any of these requirements to make a strong BIPA claim.
The BIPA claims in these two cases are important because BIPA has proven to be an effective, roundabout way to chastise companies for their invasive biometric data collection practices. Several powerful tech companies have chosen to settle BIPA lawsuits alleging they similarly collected and relied on users’ biometric data without following BIPA’s procedures: ACLU settled with Clearview AI over its use of facial biometric data; Facebook settled over its scanning of facial geometry in users’ photos ; Google also settled a similar suit, as did Snap (the company that owns SnapChat), and TikTok most recently. These were all multimillion dollar settlements, except for the Clearview settlement which severely restricted Clearview’s U.S. client-base to law enforcement agents only.
In Flora , the plaintiffs allege Prisma failed to meet all of BIPA’s procedural requirements before collecting Lensa users’ selfies, pulling their facial geometry information from those images, and training the model on those geometries to output personalized avatars. Importantly, the Flora plaintiffs rely heavily on the version of Prisma’s Privacy Policy that was in place when they downloaded the app in early December 2022. According to them, the wording of that policy was vague with the terms it used to describe the facial information Lensa collects, never using the terms “facial geometry” or “biometric” despite using the term “face data” in a separate part of the Policy (after claiming the company does not collect such data).
In P.M. , a subset of Illinois-based plaintiffs also allege that OpenAI violated BIPA’s procedural requirements. They claim that OpenAI collected and relied on their facial images from scraped photos from the internet to train image diffusion products like DALL-E, which can generate realistic depictions of human (and human-like) faces. Like Prisma, the plaintiffs allege that OpenAI did not have a public written policy concerning their use of facial data, did not receive a written release to collect and use this data from their images, and did not comply with several other BIPA requirements in developing and publishing DALL-E. If the case moves forward into discovery, OpenAI may have to finally divulge the internet sources of its training data beyond what has been uncovered by others .
Andersen v. Stability AI, 3:23-cv-00201 (N.D. Cal.) COMPLAINT DATE : 1/13/2023
In Andersen v. Stability AI , three named visual artists filed a proposed class action suit against Midjourney, Stability AI and DeviantArt, whose generative AI models were allegedly trained on billions of images scraped from the Internet without the permission of the copyright holder. The complaint alleged that there was direct and vicarious copyright infringement through non-consensual use of their works through works generated “in the style” of the artists. The complaint further claimed violations of the DMCA, violations of the right of publicity law, Cal. Civ. Code § 3344, violations of the common law right of publicity, unfair competition, and breach of contract of DeviantArt’s privacy policies.
In an order issued in October 2023, U.S. District Judge William Orrick of Northern District of California largely granted Stability AI’s motion to dismiss and gave Andersen a leave to amend to clarify their theories of each defendant’s differentiated roles in copyright infringement, DMCA violations, and violations of rights of publicity. The plaintiffs will be required to state and provide plausible facts on how compressed copies of training images contained in Stable Diffusion constituted direct copyright infringement. Another direct infringement claim is the allegation that output images from DreamUp are infringing derivative works, a claim that Judge Orrick required a showing of substantial similarity in order to survive. Additionally, the plaintiffs will have to provide facts that support Midjourney’s liability in direct infringement through the training process as well as facts that clarify DeviantArt’s liability.
A.T. et al. v. OpenAI LP et al., 3:23-cv-04557 (N.D. Cal.) COMPLAINT DATE : 9/5/2023
In A.T. v. OpenAI , users are suing OpenAI for allegedly violating federal ( ECPA ) and state ( CIPA ) wiretapping laws by collecting user data when they input content into ChatGPT or into third-party websites’ search features that have integrated ChatGPT APIs . Users are also alleging that OpenAI violates the Computer Fraud and Abuse Act of 1986 ( CFAA ), a federal anti hacking law, for the same action of collecting their user data (especially their metadata) when using ChatGPT and ChatGPT plugins on third party websites. They argue that OpenAI effectively hacks users’ platform access, exceeding user’s authorized access for OpenAI when using ChatGPT, and intercepts their private information (i.e., user data) without their knowledge or consent.
Authors Guild v. OpenAI, 1:23-cv-08292 (S.D.N.Y.) COMPLAINT DATE : 9/19/2023
Fiction writers and the Authors Guild that own registered copyrights brought a suit against OpenAI for copyright infringement for copying their works to train ChatGPT. The Guild’s complaint extensively discussed the threat and harm that generative AI like GPT-N and ChatGPT pose to authors and their livelihoods. The plaintiffs called attention to ChatGPT readily offering to produce summaries of copyrighted works, which they allege are derivative works. The plaintiffs laid out how ChatGPT was able to produce accurate summaries or hypothetical volumes of each of the plaintiffs’ existing work to allege that ChatGPT could not have generated the material if it had not ingested and trained on the specific works.
Battle v. Microsoft, 1:23-cv-01822 (D. Md.) COMPLAINT DATE : 7/7/2023
Battle v. Microsoft involves an Air Force veteran, Jeffery Battle, who happens to share the same name as Jeffrey Battle (first name spelled slightly differently), a member of the “Portland Seven,” a group of U.S. citizens who tried to join the Taliban in Afghanistan after 9/11. Battle sued Microsoft for defamation after he discovered that searching his name on Bing, the GenAI-assisted search engine, generated a short blurb that conflated the bios of the two Jeffery/Jeffrey Battles.
Battle is seeking monetary compensation from the company for the alleged reputational damage that its tool caused. He is also explicitly asking for a “remedy that will ensure that the damaging activities . . . cease and desist and that the false information will be . . . removed” from Bing search results “in perpetuity.”
On October 23, 2024, Judge Griggsby granted Microsoft's motion to compel arbitration and stayed the proceedings in court. This means the case is unlikely to generate a ruling that could be applied to other defamation cases.
Chabon v. Meta Platforms, Inc., 3:23-cv-04663 (N.D. Cal.) COMPLAINT DATE : 09/12/2023
Plaintiffs Michael Chabon, David Henry Hwang, and Matthew Klam are authors with registered copyrighted works. In their proposed class action complaint to Meta, the authors alleged copyright infringement on the basis that Meta used copyrighted materials to train the LLaMa model and the models themselves are unauthorized derivative works of the copyrighted training material. The complaint also outlined DMCA violation for removal of copyright management information, violation of the California unfair competition law, negligence, and unjust enrichment.
In November 2023, Northern District of California Judge Chhabria did not find plausible theories in Chabon’s claims and granted Meta’s joint motion to dismiss against Chabon for lack of plausible theory of infringement as in Kadrey v. Meta .
Chabon v. OpenAI, Inc., 3:23-cv-04625-PHK (N.D. Cal.) COMPLAINT DATE : 09/09/2023
Authors with registered copyrighted works brought a proposed class action complaint to OpenAI, alleging copyright infringement on the basis that OpenAI used copyrighted materials to train the GPT model rendering the models themselves unauthorized derivative works of the training material. The complaint also outlined DMCA violation for removal of copyright management information, violation of the California unfair competition law, negligence, and unjust enrichment.
In August 2023, plaintiffs in Tremblay v. OpenAI , who made indistinguishable claims against OpenAI as those made in Chabon v. OpenAI, moved to relate the two cases.
Concord Music Group, Inc. v. Anthropic PBC, 3:23-cv-01092 (M.D. Tenn.) COMPLAINT DATE : 10/18/2023
Major music publishers ABKCO, Concord, and Universal allege that Anthropic infringed publisher-owned copyrights in song lyrics. This infringement allegedly happened when lyrics were copied as part of the model training process, and when those lyrics were reproduced and distributed in response to prompts using Anthropic’s AI product, Claude.
The publishers argue that direct infringement took place at three stages of the model’s processing of the copyrighted material: copying taking place in initial data ingestion, additional reproduction in the fine-tuning of the data, and producing near-identical copies as outputs. Claude chatbot replicates significant portions of copyrighted lyrics when prompted to do so. The publishers highlighted that such service is provided by an already existing market of lyrics aggregators who have licensed those works. The publisher demonstrated that even when Claude is prompted to generate song lyrics without a request for a specified song’s lyrics, the chatbot replicated significant portions of existing copyrighted lyrics.
Furthermore, The publishers argued that Anthropic contributorily infringed and is vicariously liable for users’ direct infringement. Although the publishers point out Anthropic’s use of “constitutional” AI training model and its omission of protection of intellectual property rights as one of its principles, it does not appear that the infringement analysis would be affected by the values that govern Anthropic's AI model design.
Doe 1 v. GitHub, 4:22-cv-06823 (N.D. Cal.) COMPLAINT DATE : 11/3/2022
Software developers who own source code repositories made available on GitHub filed a proposed class action complaint against GitHub, Inc., Microsoft Corporation, and OpenAI. The plaintiffs alleged that the defendants trained Codex and Copilot on the code deriving from the plaintiffs’ published codes on GitHub, in violation of the Digital Millennium Copyright Act (DMCA), 17 U.S.C. §§ 1201–1205 and several other claims, most of which were dismissed in a mixed ruling in May 2023 by U.S. District Judge Jon Tigar in the Northern District of California.
Claims that were dismissed were for violation of Sections 1202(a) and 1202(b)(2) of the DMCA, tortious interference in a contractual relationship, fraud, false designation of origin, unjust enrichment, unfair competition, breach of the GitHub Privacy Policy and Terms of Service, violation of the California Consumer Privacy Act and negligence.
What survived the dismissal were plaintiffs’ claims that GitHub and OpenAI violated Sections 1202(b)(1) and 1202(b)(3) of the DMCA and breached the open source licenses under which the plaintiffs’ codes were made available through GitHub.
Flora v. Prisma Labs, 5:23-cv-00680 (N.D. Cal.) COMPLAINT DATE : 2/15/2023
In Flora v. Prisma Labs , the plaintiffs allege Prisma failed to meet all of the Illinois Biometric Privacy Act (BIPA)’s procedural requirements before collecting Lensa users’ selfies, pulling their facial geometry information from those images, and training the model on those geometries to output personalized avatars. First, they allege Prisma violated section 15(a), requiring Prisma to develop a written policy establishing a schedule to destroy biometric data after a certain time; they allege Prisma violated section 15(b) for collecting user’s biometric data without first notifying the user in writing of the purpose and length of collect and receiving a written release form; they also allege violations of sections 15(c) by profiting off of biometric data collected to improve Lensa, 15(d) by transferring users’ biometric data to third-party cloud servers without disclosing this and receiving user consent, and 15(e) for failing to store user’s biometric data in a way that meets the industry standard of care. Importantly, the Flora plaintiffs rely heavily on the version of Prisma’s Privacy Policy that was in place when they downloaded the app in early December 2022. According to them, the wording of that policy was vague with the terms it used to describe the facial information Lensa collects, never using the terms “facial geometry” or “biometric” despite using the term “face data” in a separate part of the Policy (after claiming the company does not collect such data).
The Flora plaintiffs also suggest that Prisma’s Privacy Policy states the company deletes users’ biometric data (gleaned from their uploaded selfies) within 24 hours of creating the avatars. But this claim by Prisma is spurious, as the ML community is still trying to understand the extent to which models memorize training data in ways that make it impossible to delete or remove that data from the ultimate models.
Getty Images Inc. v. Stability AI, 1:23-cv-00135 (D. Del.) COMPLAINT DATE : 2/3/2023
Getty Images brought claims against Stability AI for copyright infringement, DMCA violation, and trademark infringement and dilution. Getty Images alleged that Stability AI used more than 12 million Getty Images photos to train the model for Stability’s image generator, Stable Diffusion. In the complaint, Getty Images juxtaposed a copyrighted Getty Images photograph next to an image generated by Stable Diffusion to underscore the link between the two. The generated images included in the complaint have what Getty Images calls a modified version of the Getty Images watermark, which Getty Images claims to have harmed it by “creating confusion as to the source of the images and falsely implying an association with Getty Images."
Huckabee et al v. Meta Platforms, Inc. et al, 1:23-cv-09152 (S.D.N.Y.) COMPLAINT DATE : 10/17/2023
Mike Huckabee, former Arkansas governor, and a group of religious authors filed a class action suit against Meta, Bloomberg, Microsoft, and the EleutherAI Institute, arguing they trained artificial intelligence tools on the authors’ books without permission when they used Books3, a text dataset compiled by independent developers.
In their complaint, the plaintiffs emphasized how essential Books3, which contained their works, was in training the LLM products.They claim that the LLM products of Meta, Microsoft, and Bloomberg cannot function without the use of Books3, making the models infringing derivative works of the allegedly incorporated copyrighted materials within the dataset. Plaintiffs further argued that the defendants were vicariously liable for infringement because every output of the LLMs is also an infringing derivative work, but the plaintiffs did not identify who the direct infringer is in those instances.
J.L. v. Alphabet Inc., 3:23-cv-03440 (N.D. Cal.) COMPLAINT DATE : 7/11/2023
The plaintiffs, led by a New York Times best-selling author and journalist, J.L, who requested to keep their identity private, filed a putative class action complaint against Alphabet, Google Deepmind and Google LLC. The case alleges that Google misused content that the plaintiffs posted on social media and Google platforms, to train Bard. They alleged copyright infringement, DMCA violation, and other torts. In September, plaintiffs voluntarily dismissed Alphabet and Google Deepmind, leaving Google LLC as the sole defendant.
Kadrey v. Meta, 3:23-cv-03417 (N.D. Cal.) COMPLAINT DATE : 7/7/2023
The comedian and author Sarah Silverman and other published authors filed a class action lawsuits against Meta claiming it infringed upon their copyright by making copies of their books as part of training data for Meta’s LLaMA AI model, that the LLaMA model itself was a “derivative work” of their books, and that various LLaMA outputs were unlawful copies of their books. Plaintiffs also claimed Meta removed their Copyright Management Information (CMI) as well as engaging in unfair competition, unjust enrichment, and negligence.The complaint was filed along with the one filed against OpenAI by the same lead plaintiffs.
Meta moved to dismiss every claim except the copyright claim related to training LLaMa, reserving it to challenge later in the case. Judge Vince Chhabria of the Northern District of California granted Meta’s motion to dismiss across the board finding that there were no plausible theories for plaintiffs to win on. Specifically, Judge Chhabria rejected the claim that LLaMA models themselves are infringing derivative works, calling this a nonsensical argument, since the models are software code that create abstract mathematical relationships and could not be conceived of as a recasting or adaptation of any of the expressive text of plaintiffs’ books. The allegation that every output is an infringing derivative work and that Meta is vicariously liable for the users prompting such outputs was also rejected because of the plaintiffs’ failure to allege that any specific output was infringing or incorporated the copyrighted material. Without specific allegations that specific outputs incorporated their copyrighted works, there was no legal basis for the claim to move forward in the case. The court also dismissed the CMI claim because it also depended on specific allegations that LLaMA could produce plaintiffs’ books. The unfair competition claims, unjust enrichment claims, and negligence claims all failed as well because they duplicated the copyright infringement claims and thus were preempted by the Copyright Act. The court gave plaintiffs 21 days to amend their complaint to cure these problems.
Assuming plaintiffs cannot allege any new facts, the only remaining claim left out of the motion to dismiss is the claim on use of copyrighted works for training, a use that Meta maintains is quintessential fair use.
P.M. v. OpenAI, 3:23-cv-03199 (N.D. Cal.) COMPLAINT DATE : 6/28/2023
In P.M. v. OpenAI , users are suing OpenAI for allegedly violating federal ( ECPA ) and state ( CIPA ) wiretapping laws by collecting user data when they input content into ChatGPT or into third-party websites’ search features that have integrated ChatGPT APIs . Users are also alleging that OpenAI violates the Computer Fraud and Abuse Act of 1986 ( CFAA ), a federal anti hacking law, for the same action of collecting their user data (especially their metadata) when using ChatGPT and ChatGPT plugins on third party websites. They argue that OpenAI effectively hacks users’ platform access, exceeding user’s authorized access for OpenAI when using ChatGPT, and intercepts their private information (i.e., user data) without their knowledge or consent.
A subset of Illinois-based plaintiffs further allege OpenAI violated the Illinois Biometric Privacy Act (BIPA) when it collected their biometric identifiers without having a written policy, informing users in writing, receiving a written release, and meeting BIPA’s other procedural requirements (described above). Specifically, they claim that OpenAI collected and relied on their facial images from scraped photos from the internet to train image diffusion products like DALL-E. They allege OpenAI collected facial geometry data of Illinois residents, without which DALL-E could not exist. If the case moves forward into discovery, OpenAI may have to finally divulge the sources of its training data beyond what has been uncovered by others .
Registration of “SURYAST”, SR # 1-11016599571 (U.S. Copyright Office) Registration denied : 1211/2023
The Review Board of the Copyright Office affirmed the Copyright Office’s rejection of the registration of an image generated using RAGHAV, a 2-D image application. The image was generated by the claimant Ankit Sahni who inputted his own photograph as a base image and transferred style from Vincent van Gogh’s artwork through the program. The Office refused to register the work because it lacked the human authorship necessary to support a copyright claim, extending its uniform rejection of AI-generated outputs. See Thaler v. Perlmutter . The Review Board of the USCO maintained that it would register a work that contains sufficient human authorship, but that this work lacked “the expressive elements of pictorial authorship” provided by the claimant. The Review Board refused to recognize that RAGHAV was used as “an assistive tool” akin to a camera or photo-editing software subject to creative decisions by the author and further noted that the inputted choices made by Sahni merely constituted unprotectable ideas for the work.
Registration of “Théâtre D’opéra Spatial”, SR # 1-11743923581 (U.S. Copyright Office) Registration denied : 9/5/2023
Mr. Jason Allen applied for copyright registration of a two-dimensional artwork claim in the work titled “Théâtre D’opéra Spatial,” which was generated using Midjourney after employing as many as 600 text prompts. The USCO requested more information about Allen’s use of Midjourney and subsequently denied the registration. When Allen submitted requests for reconsideration, USCO review board found that the work contained more than a de minimis amount of AI-generated content which must be disclaimed in an application for registration. The USCO refused to register the work, because Allen was unwilling to disclaim the AI-generated material.
Registration # VAu001480196 (U.S. Copyright Office) Registration canceled : 2/21/2023
Kristina Kashtanova applied for copyright registration of a graphic novel work, “Zarya of the Dawn,” which was created with output from Midjourney that Kashtanova described as “the core creative input” for each image. Registration certificate was granted but the USCO canceled the registration due to a failure to exclude non-human authorship contained in the work. The USCO then replaced it with a new registration that covers the original authorship of Kashtanova, narrowing its coverage to the “text” and the “selection, coordination, and arrangement of text created by the author and artwork generated by artificial intelligence.”
Sancton v. OpenAI Inc. et al., No. 1:23-CV-10211 (S.D.N.Y.) COMPLAINT DATE : 11/21/2023
Writer Julian Sancton brought claims of copyright infringement and contributory copyright infringement against OpenAI and Microsoft on behalf of a putative class of owners of nonfiction works. The plaintiffs alleged that reproducing copyrighted works in datasets used to train their AI models constituted direct copyright infringement. Plaintiffs moved to relate the case to Authors Guild v. OpenAI , another class action brought by authors.
Silverman v. OpenAI, 3:23-cv-03416 (N.D. Cal.) COMPLAINT DATE : 7/7/2023
Comedian and author Sarah Silverman as well as other published authors allege that OpenAI infringed upon copyright by making copies of their books as well as works of those who are similarly situated. The complaint argued that the OpenAI’s large language models are themselves infringing derivative works because the models cannot function without the expressive information extracted from the infringed works.
OpenAI moved to dismiss the claims for vicarious copyright infringement, violation of DMCA § 1202, unjust enrichment, and negligence. Although OpenAI did not move to dismiss the direct copyright infringement claim, it laid grounds to discredit infringement claims by invoking the fair use doctrine, requirement for substantial similarity, and that copyright does not protect ideas, facts, or language. In a memo filed in September 2023, the plaintiffs strongly rebutted by arguing that copyright infringement does not require a showing of substantial similarity in direct copying cases in the Ninth Circuit and reiterated the strength of their direct copyright infringement claim.
Thaler v. Perlmutter, 1:22-cv-01564, (D.D.C.) COMPLAINT DATE : 6/2/2022
Dr. Stephen Thaler, who developed and owns an AI system called Creativity Machine, filed to register outputs generated from the AI system for copyright registration in November 2018. The U.S. Copyright Office (USCO) denied the registration application for the AI-generated work in a final agency action for the lack of “human authorship necessary to support a copyright claim.” Thaler filed against Perlmutter, the register at the USCO, for a judicial review of the denial under the Administrative Procedures Act, claiming that this rejection was contrary to the plain language of the Copyright Act and the statutory purpose of its mandate to promote the progress of science. In granting summary judgment in favor of the USCO, the District Court for the District of Columbia held in August 2023 that human authorship is an essential part of a valid copyright claim, emphasizing that authorship centers on human creativity. Thaler has since filed an appeal to the Court of Appeals for the D.C. Circuit.
Thomson Reuters v. ROSS Intelligence, 1:20-cv-00613 (D. Del.) COMPLAINT DATE : 5/6/2020
Thomson Reuters, which owns Westlaw, a legal research platform, sought to recover from an AI startup, Ross, by alleging that Ross trained its AI on copies of Westlaw headnotes. Ross contracted with LegalEase, a legal-research company to conduct a Bulk Memo project that produced 25,000 question-and-answer sets with assistance of a text-scraping bot that accessed Westlaw. The LegalEase memos were then converted into training data by Ross. Thomson Reuters argued that the questions portions of Bulk Memo were essentially copies of headnotes.
On September 25, 2023, Judge Stephanos Bibas denied summary judgment and ruled the case to proceed to trial. Judge Bibas rejected Ross’s theory that its use of a portion of the large compilation was not enough to be infringing. Whether the headnotes themselves met the originality threshold to merit copyright protection is a factual issue left to the jury to determine. This issue on protected expressions will shape the theories of infringement liability.
In finding that the fair use determination should go to a jury, Judge Bibas declined to overread a recent fair use rulings in Andy Warhol Foundation v. Goldsmith , in which the Supreme Court heavily weighed commercial use to rule against fair use. Instead, he emphasized transformativeness and suggested that Ross may prevail in its fair use defense if factual determination supported that its use was transformative intermediate copying. Furthermore, he noted that the nature of Westlaw’s Key Number System and headnotes gravitate away from the core of copyright protection and may favor fair use, but the decisions on each factor will have to wait until trial.
Tremblay v. OpenAI, 3:23-cv-03223 (N.D. Cal.) COMPLAINT DATE : 4/3/2023
In Tremblay v. OpenAI , the named plaintiffs were authors who have registered copyrights in several books. The plaintiffs brought a proposed class action suit against OpenAI for direct and vicarious copyright infringement, DMCA violation, unfair competition and negligence. The plaintiffs alleged that OpenAI made copies of the books, made derivative works, publicly displayed, and/or distributed those copies or derivative works. The complaint identified that ChatGPT was able to generate fairly accurate summaries of books written by the plaintiffs, suggesting that the model had ingested and retained them. The complaint seeks relief on behalf of all authors whose books were copied as part of the training data set for ChatGPT.
In August 2023, OpenAI filed a motion to dismiss and the plaintiffs moved to relate this case to another case against OpenAI, Chabon v. OpenAI.
Walters v. OpenAI, No. 23-A-04860-2 (Ga. Super. Ct., Gwinnett Cnty.) COMPLAINT DATE : 6/5/2023
Walters v. OpenAI, No. 23-cv-03122 (N.D. Ga.) COMPLAINT DATE : 7/14/2023
In Walters v. OpenAI , a Georgia state trial court rejected a defamation lawsuit brought against OpenAI by radio talk show host Mark Walters. Walters had sought monetary compensation from the company for damage to his reputation allegedly caused when, in response to prompting from a journalist, OpenAI's ChatGPT tool produced an output that wrongly described Walters as a defendant in a lawsuit who had been accused of fraud.
The court initially denied OpenAI’s motion to dismiss in January 2024, after which the parties engaged in a few months of discovery. OpenAI then filed a motion for summary judgment. In order to aid the court’s consideration of this motion, NYU’s Technology Law & Policy Clinic filed a friend-of-the-court brief emphasizing that the court should consider the core purposes of defamation law and conduct a fact-specific inquiry to adapt the law to this new and increasingly widespread technology.
In its opinion, the court granted summary judgment in favor of OpenAI on three independent bases. The court first concluded that the ChatGPT output at issue could not be “reasonably understood as describing actual facts” about Walters, in part because ChatGPT specifically warned the journalist that it possessed insufficient information to answer his prompts and because OpenAI generally and repeatedly warns all of its users that “ChatGPT can and does sometimes provide factually inaccurate information.” The court additionally found significant that the journalist admitted that within a short time after prompting ChatGPT, he “had confirmed that the output did not contain actual facts.” As the court noted, “[i]f the individual who reads a challenged statement does not subjectively believe it to be factual, then the statement is not defamatory as a matter of law.”
The court also likened OpenAI to a “publisher” and explained that Walters failed to show how OpenAI was at fault, under either a negligence or actual malice standard, in allowing ChatGPT to produce the output. Walters had merely alleged that “because ChatGPT is capable of producing mistaken output, OpenAI was at fault simply by operating ChatGPT at all.” The court rejected this proposed rule, which “would impose a standard of strict liability” on OpenAI, holding the company liable for any injury caused by ChatGPT, no matter what steps OpenAI might take to prevent such injury.
Finally, the court held that, in any event, Walters had not suffered and was not entitled to any damages arising out of the inaccurate output.
Kyland Young v. NeoCortext, Inc., 2:23-cv-02496 (C.D. Cal.) COMPLAINT DATE : 4/3/2023
Kyland Young, a former contestant of a reality television show, brought right of publicity claims in a proposed class action against NeoCortext Inc., a Ukrainian company that made the AI-based face swapping app, Reface. The complaint alleged that the app exploited Young’s image and likeness without consent or compensation, violating the right of publicity under California Civil Code § 3344.
In September 2023, U.S. District Judge Wesley L. Hsu of Central District of California denied NeoCortext’s motion to dismiss and motion to strike, writing that the right of publicity claims were not preempted by the federal Copyright Act. Under California law, significant transformative use would warrant an affirmative defense to a right of publicity claim under the First Amendment, but Judge Hsu found that NeoCortext had failed to show that Reface app’s use of Young’s face was a transformative use as a matter of law.