Have you ever stopped to consider what someone would find if they could see your internet search history? You might be embarrassed by what would be revealed. Most people don’t think twice about asking ridiculous, embarrassing, or even deeply personal questions:

“Why is my goldfish not swimming?”

“What’s the best place for sushi in Mexico City?”

“What rhymes with fart?”

But very little of our life on the internet is private. And nowhere is that more misunderstood than with AI tools like ChatGPT or Claude.

The core problem is the gap between how people experience AI and what’s actually happening. The conversational interface feels private – like talking to an advisor, a therapist, or no one.

Legally, it may be closer to speaking on a recorded line.

Recently, a case out of the Southern District of New York should raise every lawyer and client’s spidey senses about using AI. Conversations with AI may not be as insulated as they seem, and they may be discoverable, particularly depending on the subscription tier being used.

Roadmap
I. What Is Discovery / Attorney-Client Privilege / Work Product Doctrine?
II. History: Your Deleted ChatGPT Logs Are Discoverable - New York Times v. OpenAI/Microsoft
III. Fast Forward: AI Searches Can Destroy Privilege?! - United States v. Heppner
IV. Digging into the Terms of Service, Privacy, and Usage Policies of Claude and ChatGPT
V. Plot Twist: AI-Assisted Drafting Is Protected Work Product in Michigan — Warner v. Gilbarco, Inc.
VI. Reconciling Heppner and Warner
VII. Potential Ways to Mitigate Risk
VIII. Public Policy Thoughts
IX. Sources

I. What Is Discovery / Attorney-Client Privilege / Work Product Doctrine?

Let’s start with the basics.

A) Discovery

When a lawsuit is filed, the parties enter a phase called discovery—the pretrial process where each side must formally exchange relevant, non-privileged information and evidence to evaluate claims and defenses. My Cousin Vinny provides a helpful illustration of how discovery works:

Discovery is designed to ensure both sides have access to the facts before resolution or trial. Because discovery is broad, the law also recognizes important limits. Two of the most significant limitations are (i) the Attorney-Client Privilege and (ii) the Work Product Doctrine.

B) Attorney-Client Privilege (A-C Privilege)

A-C Privilege is a rule that protects confidential communications (oral, written, or electronic) between a client and their attorney made for the purpose of seeking or providing legal advice. The privilege protects both individuals and institutions. For example, communications between Vandeley Industries employees and Vandeley attorneys are protected if they are made in confidence and for the purpose of seeking legal advice about Vandeley’s legal matters. When the privilege applies, those communications cannot be disclosed to opposing parties or required to be produced in legal proceedings (with some exceptions).

C) Work Product Doctrine (The WPD)

The WPD protects materials prepared by or at the direction of an attorney in anticipation of litigation. Unlike A-C Privilege, it focuses on documents and strategy (notes, memos, research, etc.). In rare cases, it can be overcome if the opposing party shows a substantial need and inability to obtain the equivalent without undue hardship.

Why do we have privilege?

The A-C Privilege exists for a very practical reason: the legal system works better when clients can speak openly and honestly with their lawyers. If people fear that anything they tell their attorney could later be exposed in court, clients would be more likely to withhold facts, soften details, or avoid seeking legal advice altogether. That prevents attorneys from fully understanding the situation and providing effective representation. In this way, the privilege not only protects the client’s interests but also supports the fair administration of justice.

What are the limits to privilege?

Many clients assume that if a lawyer is involved in any conversation or copied on an email, the communication is automatically protected. However, that’s not always the case. The primary purpose of the exchange must be to seek or deliver legal guidance—not merely to discuss business strategy, operational decisions, or general information in the presence of counsel.

Additionally, you can lose the A-C Privilege when:

• The communication is shared with a third party (called waiver).

• There was no intent for the communication to be confidential.

• The advice was used to further a crime or fraud (crime-fraud exception).

• The client puts the advice “at issue” in litigation (e.g., “I relied on my lawyer’s advice”).

How far does privilege extend in terms of legal tools and service providers?

Not every disclosure to a third party automatically destroys A-C privilege.

Under United States v. Kovel, the protection may extend to third parties who are engaged to assist counsel in providing legal advice, so long as they are retained for that purpose, operate under the attorney’s supervision and direction, and are bound by confidentiality obligations. Courts have consistently applied this doctrine to professionals such as forensic consultants, translators, and e-discovery vendors when their involvement is necessary to facilitate legal representation.

The same principle applies to the WPD. It does not cover every document a lawyer touches; it protects materials prepared by or at the direction of counsel in anticipation of litigation. This means the material must be created because of a real and identifiable prospect of litigation—not merely as part of routine business operations or general risk management. Documents created in the ordinary course of business, even if reviewed by a lawyer, may not qualify. These protections are specific and conditional — not automatic or absolute.

II. History: Your Deleted ChatGPT Logs Are Discoverable - New York Times v. OpenAI/Microsoft

In N.Y. Times Co. v. Microsoft Corp. (2023), pending in the Southern District of New York, the Times alleged that OpenAI and Microsoft used millions of its articles without permission to train AI models like ChatGPT, which can at times reproduce or closely paraphrase its reporting in ways that threaten subscriptions and paywalls. OpenAI and Microsoft denied infringement, arguing that training on publicly available text was fair use because the models learn linguistic patterns rather than store or republish articles, and that verbatim outputs are rare and unintended. See a previous post for more information.

As the case progressed, discovery became a central battleground. The Times sought access to up to 120 million user chat logs; the court ultimately ordered production of a smaller, anonymized sample of roughly 20 million conversations under a protective order. OpenAI argued discovery should be limited to chats tied to specific Times articles or exemplar prompts, but the Times contended it needed broader data to assess how the models function in real-world use and whether they generate Times-like content at scale. The court rejected OpenAI’s narrower proposal, holding that privacy concerns do not automatically bar discovery of relevant electronic data where safeguards are in place.

The case also raised issues about preserving and deleting data. At one point, the court required OpenAI to preserve and separate certain ChatGPT log data that would normally have been deleted—including conversations users thought they had already deleted.

This ruling highlights a tension between what users expect and what the law requires during litigation. Even if a platform makes conversations temporary, the underlying data may have to be kept once there is a legal duty to preserve it as evidence.

More broadly, it underscores an important point for both AI companies and users: even deleted conversations may not be fully or immediately erased.

III. Fast Forward: AI Searches Can Destroy Privilege?! in NY - United States v. Heppner

I’m sure most lawyers at this point have had a client put whatever they said into an AI chat or use AI to try to teach their own lawyers. Well, this case is a bit of schadenfreude for legal eagles….

Bradley Heppner, a financial executive facing fraud charges, used the consumer version of Anthropic’s Claude after receiving a grand jury subpoena. He input information he had learned from his defense counsel, generated 31 AI-produced documents, and sent them to his lawyers. The FBI later seized those materials during a search.

Heppner argued they were privileged.

Judge Rakoff disagreed. The ruling turned on several key principles:

• Routing AI material through counsel does not create privilege. Simply sending AI-generated documents to a lawyer does not automatically shield the client from discovery.

• Independent research is not privileged. A client’s self-directed research—whether Googling case law, drafting personal notes, or posing legal questions to an AI platform—does not become protected simply because it relates to a legal issue. If the work was not prepared by or at the direction of counsel, and is not itself a communication seeking legal advice, it may be discoverable.

• AI tools are not lawyers. No attorney directed the AI use, and AI platforms cannot form A-C relationships.

• The purpose was not to obtain legal advice from counsel. The documents were generated through Claude, which expressly disclaims providing legal advice, undermining any claim that the communications were made for the purpose of seeking legal counsel.

• There was no reasonable expectation of confidentiality. Anthropic’s policies permitted disclosure to government authorities and use of data for model training, defeating any claim that the communications were confidential.

• Privilege cannot be created after the fact. Sending AI-generated documents to counsel did not retroactively cloak them in privilege, and the work product doctrine did not apply because the lawyers had not directed the AI research.

As the Casey Anthony trial demonstrated years ago, even internet search histories can become evidence. AI prompts are simply a more sophisticated version of the same concept.

IV. Digging into the Terms of Service, Privacy, and Usage Policies of Claude and ChatGPT

Let’s review the ToS for the consumer v enterprise versions of Claude and ChatGPT.

A) Claude

B) ChatGPT

1. Training & Privilege Risk (Consumer vs. Enterprise)
   Consumer plans may use inputs/outputs for training unless users opt out; business/enterprise plans are opted out by default. However, both permit internal access for safety, security, abuse prevention, and legal compliance. Courts have not yet addressed how these commercial AI arrangements affect A-C privilege. Because companies can access data internally in some situations, parties in litigation may argue that these communications are not completely private.

2. Disclosure of Confidential Information
   Both platforms allow disclosure when legally required. Claude states it will coordinate with Enterprise administrators to narrow compelled disclosures; OpenAI says it will make reasonable efforts to notify customers where permitted. Still, these policies do not remove the risk of subpoenas or discovery in lawsuits. Organizations should pay close attention to provider notices, especially if litigation is expected or ongoing.

3. Data Retention Controls
   Claude Enterprise offers clearer, configurable retention settings (e.g., minimum 30-day retention). OpenAI retains data “as long as needed,” with deletion subject to legal and security exceptions. While deleting data can lower risk, it does not completely eliminate it because AI companies may still have to keep information for legal or compliance purposes.

4. Webpage Structure, Contract Complexity, and Enforceability
   Claude’s policies are spread across several different documents, which can make them harder to follow. OpenAI’s policies are more centralized and organized by user type, though they are still divided across multiple documents. This structure can create confusion in litigation, especially when questions about privilege, confidentiality, and data access depend on how different provisions interact. It may also raise enforceability questions if certain terms appear only in a privacy policy—which is typically a notice rather than a contract—or on informational webpages outside the formal terms of service.

V. Plot Twist: AI-Assisted Drafting Is Protected Work Product in Michigan — Warner v. Gilbarco, Inc.

But that same week, a federal court in Michigan reached what appeared to be the opposite conclusion.

In Warner v. Gilbarco, Inc., Magistrate Judge Anthony Patti of the Eastern District of Michigan held that a pro se litigant’s use of ChatGPT to help draft filings did not waive protection under the work product doctrine.

Warner involved a pro se employment discrimination plaintiff who used ChatGPT to help prepare her filings. Defendants moved to compel production of all documents and information concerning her use of third-party AI tools, including prompts, outputs, and activity logs.

Judge Patti denied the motion. His reasoning turned on a few key points:

• AI as a “tool,” not a third party. The court treated ChatGPT more like a drafting tool than a human consultant. On that view, using AI is closer to using a word processor or research database than sharing strategy with an outside person.

• Work product waiver is not privilege waiver. Even if privilege can be fragile when information is shared outside the attorney-client relationship, work product is typically waived only by disclosure to an adversary (or conduct likely to reach one). Simply using AI to draft is not the same as handing strategy to the other side.

• Prompts can reflect mental impressions. The court also recognized that forcing a litigant to disclose AI prompts and outputs could reveal mental impressions and legal strategy, which are the core protections of Rule 26(b)(3) of the Federal Rules of Civil Procedure. A discovery request demanding “how you drafted this” can become an indirect attempt to access protected work product.

VI. Reconciling Heppner and Warner

Placed side-by-side, Heppner and Warner seem to point in opposite directions.

• In Heppner, Judge Rakoff treated the AI system as a third party whose privacy terms undermined any reasonable expectation of confidentiality.

• In Warner, Judge Patti treated the AI system as a technological drafting tool, not a third-party receiving legal strategy.

The distinction is significant because modern legal practice already depends on cloud-based infrastructure. Email providers, document management systems, and collaborative platforms routinely process sensitive legal communications under terms of service that allow internal access for security, compliance, or operational purposes. If the mere presence of such terms were enough to defeat confidentiality, a wide range of everyday legal technologies—from Gmail to Microsoft 365—could potentially threaten privilege. Courts have historically rejected such an expansive view, recognizing that necessary technological intermediaries do not automatically destroy confidentiality.

The facts of the two cases also differ in important ways.

• In Heppner, the defendant had already retained counsel but independently used a consumer AI platform to generate documents after receiving a grand jury subpoena.

• In Warner, the plaintiff was representing herself, meaning the AI-assisted drafting directly reflected her own litigation strategy as the acting attorney.

Those differences help explain the outcomes, but they do not fully resolve the underlying issue.

What remains unresolved—and likely more important for future litigation—is the scenario neither case squarely addressed: the use of enterprise AI software—with contractual confidentiality protections and restrictions on data training—by an attorney or by a represented litigant acting under counsel’s direction.

Also importantly, neither Heppner nor Warner is binding precedent beyond their respective cases, as both are trial-court rulings that may serve only as persuasive authority for other courts. We’re still early in this AI era, and this area is still developing.

VII. Potential Ways to Mitigate Risk

Commercial-grade tools offer more protection but still require caution. Consumer-grade AI platforms generally do not provide the same level of contractual confidentiality protections as enterprise agreements, and even where users opt out of model training, other data access, retention, or disclosure provisions may still apply.

Don’t enter anything you don’t want seen. Sensitive information entered into these systems may not receive the same protections as communications made directly within the A-C relationship. In an abundance of caution, users should assume that information submitted to consumer platforms could be retained, accessed, or disclosed beyond their immediate control.

A) Considerations for Non-Lawyers (Clients)

Leave legal questions for the attorneys. Legal questions should be directed through attorneys, rather than having clients—or, in a business setting, managers or employees—independently rely on AI tools to conduct their own legal assessments outside of attorney direction. If research is conducted, documentation should clearly reflect that it was performed at the direction of counsel, demonstrating that the attorney was involved in and supervising the process.

Privilege isn’t magic. You have to work to keep it….Sending unprivileged AI-generated material to your lawyer does not retroactively make it privileged.

Be cautions with sensitive matters. Exercise extreme caution when using AI in connection with sensitive matters, especially for HR decisions, workplace complaints, investigations, or regulatory issues. Having a good AI policy in place at your company is key.

B) Considerations for Lawyers

Educate clients that AI usage may waive privilege or generate discoverable material. Clients should be cautioned against using consumer-grade generative AI platforms to analyze privileged or sensitive legal matters, as such platforms often retain, process, or train on user inputs. If you learn that a client is independently using AI to assess their legal situation, address the issue immediately and provide clear guidance.

Attorneys must also exercise caution with the information they input into large language models. Identifiable or sensitive client information should not be entered into AI systems unless adequate safeguards are in place. When AI tools are necessary, enterprise-level applications should be used where possible—particularly those that limit data retention, prohibit model training on user inputs, and impose strict contractual limits on use and disclosure. Lawyers should also understand and manage platform retention policies, including adjusting privacy settings and deleting prompts or histories where permitted.

VIII. Public Policy Thoughts

From a public policy perspective, especially given longstanding access-to-justice challenges, courts should not penalize individuals for researching their own legal issues. Legal representation is often unaffordable, so many people turn to available tools not to replace lawyers, but to understand when they need one. Just as consulting WebMD reflects a desire to be informed, the law should not discourage individuals from educating themselves about their rights.

Informed clients make better decisions, participate more effectively in the legal process, and are less likely to act out of panic or misinformation. The real concern is not self-education, but ensuring that the tools people use do not unintentionally undermine legal protections like A-C privilege.

Privilege should not hinge on technological literacy. Expanding waiver doctrine, as in Heppner, would disproportionately harm less sophisticated users who lack knowledge about data storage or platform practices or perhaps the funds to purchase an enterprise-grade version. If routine internal platform access defeated confidentiality, the same logic could threaten email, cloud storage, and other essential communication tools—an outcome courts like Kovel have rejected by recognizing that necessary intermediaries do not destroy privilege.

The justice system functions best when people have tools that help them engage with it responsibly. Research from The Bail Project similarly shows that supportive measures—such as reminders and transportation assistance—improve court participation. AI can serve a comparable supportive function: it does not replace lawyers, but helps individuals better understand and navigate the system. People engage more responsibly when given tools that promote clarity, not when barriers are raised.

When structured to preserve privilege, AI enhances meaningful participation without displacing professional representation. Public policy should reflect that reality by integrating modern tools in ways that expand access to justice while safeguarding core legal protections.

Share

Subscribe now

XI. Sources

1. United States v. Kovel, 296 F.2d 918 (2d Cir. 1961). https://law.justia.com/cases/federal/appellate-courts/F2/296/918/131265/

2. N.Y. Times Co. v. Microsoft Corp., et al., No. 1:23-cv-11195 (S.D.N.Y. filed Dec. 27, 2023). https://nytco-assets.nytimes.com/2023/12/NYT_Complaint_Dec2023.pdf

3. Motion for a Ruling that Documents the Defendant Generated Through an Artificial Intelligence Tool Are Not Privileged, United States v. Heppner, No. 25 Cr. 503 (JSR), Doc. 22 (S.D.N.Y. Feb. 6, 2026). https://storage.courtlistener.com/recap/gov.uscourts.nysd.652138/gov.uscourts.nysd.652138.22.0.pdf

4. Warner v. Gilbarco, Inc., No. 2:24-cv-12333, ECF No. 94 at 10-12 (E.D. Mich. Feb. 10, 2026). https://law.justia.com/cases/federal/district-courts/michigan/miedce/2:2024cv12333/379552/94/

5. Pete Brush, AI Docs Sent By Exec to Attys Not Privileged, Judge Says, Law360 (Feb. 11, 2026). https://www.law360.com/pulse/legal-tech/articles/2440082

6. Wednesday Testimony Delved into Internet Searches on Casey Anthony’s Computer, Jacksonville.com / Florida Times-Union (June 9, 2011). https://www.jacksonville.com/story/news/crime/2011/06/09/wednesday-testimony-delved-internet-searches-casey-anthonys-computer/15900821007/

7. Andrew R. Lee, Jason M. Loring & Graham H. Ryan, Your AI Conversations Are Not Privileged: What a New SDNY Ruling Means for Every Lawyer and Client, Jones Walker LLP, AI Law and Policy Navigator Blog (Feb. 13, 2026). https://www.joneswalker.com/en/insights/blogs/ai-law-blog/your-ai-conversations-are-not-privileged-what-a-new-sdny-ruling-means-for-every.html?id=102mif8

8. Andrew R. Lee, OpenAI Loses Privacy Gambit: 20 Million ChatGPT Logs Likely Headed to Copyright Plaintiffs, Jones Walker LLP, AI Law and Policy Navigator Blog (Jan. 6, 2026). https://www.joneswalker.com/en/insights/blogs/ai-law-blog/openai-loses-privacy-gambit-20-million-chatgpt-logs-likely-headed-to-copyright-p.html?id=102lzo9.

9. Parker Hancock, Your AI Conversations May Not Be Privileged: What United States v. Heppner Means for Every Organization Using AI, Baker Botts LLP, Our Take Blog (Feb. 13, 2026). https://ourtake.bakerbotts.com/post/102mibz/your-ai-conversations-may-not-be-privileged-what-united-states-v-heppner-means

10. Daniel Schwartz, A Court Just Confirmed What Employers Need to Hear: Your AI Conversations Are Not Privileged, Shipman & Goodwin LLP, Connecticut Employment Law Blog (Feb. 16, 2026). https://www.ctemploymentlawblog.com/2026/02/articles/a-court-just-confirmed-what-employers-need-to-hear-your-ai-conversations-are-not-privileged/

11. Michael Peltz, LinkedIn post re: United States v. Heppner ruling (Feb. 2026). https://www.linkedin.com/posts/mpeltz_ai-docs-sent-by-exec-to-attys-not-privileged-activity-7427479054242721792-n2LC/

12. Jason M. Loring, Andrew R. Lee, & Graham H. Ryan, Same Week, Different Frameworks: Why Heppner and Warner Both Got It Right on AI Privilege — and Why That’s the Problem, Nat’l L. Rev. (Mar. 5, 2026), https://natlawreview.com/article/same-week-different-frameworks-why-heppner-and-warner-both-got-it-right-ai

13. Joe Patrice, Does Asking ChatGPT a Legal Question Make It Discoverable? It Depends!, Above the Law (Feb. 24, 2026, 2:32 PM), https://abovethelaw.com/2026/02/does-asking-chatgpt-a-legal-question-make-it-discoverable-it-depends/

14. Attorney-Client Privilege, Legal Information Institute, Cornell Law School, Wex Legal Encyclopedia. https://www.law.cornell.edu/wex/attorney-client_privilege

15. Attorney Work Product Privilege, Legal Information Institute, Cornell Law School, Wex Legal Encyclopedia. https://www.law.cornell.edu/wex/attorney_work_product_privilege

16. Updates to Our Consumer Terms, Anthropic, https://www.anthropic.com/news/updates-to-our-consumer-terms (last visited Feb. 28, 2026).

17. Commercial Terms of Service, Anthropic, https://www.anthropic.com/legal/commercial-terms (last visited Feb. 28, 2026).

18. Consumer Terms of Service, Anthropic, https://www.anthropic.com/legal/consumer-terms (last visited Feb. 28, 2026).

19. Usage Policy, Anthropic, https://www.anthropic.com/legal/aup (last visited Feb. 28, 2026).

20. Privacy Policy, Anthropic, https://www.anthropic.com/legal/privacy (last visited Feb. 28, 2026).

21. Privacy Policy, OpenAI, https://openai.com/policies/row-privacy-policy/ (last visited Feb. 28, 2026).

22. Terms of Use, OpenAI, https://openai.com/policies/terms-of-use/ (last visited Feb. 28, 2026).

23. Service Terms, OpenAI, https://openai.com/policies/service-terms/ (last visited Feb. 28, 2026).

24. Data Processing Addendum, OpenAI, https://openai.com/policies/data-processing-addendum/ (last visited Feb. 28, 2026).

25. The Bail Project, “The False Promise of Bail: An Analysis of Release Mechanisms and Court Appearance in Tulsa County, Oklahoma,” June 10, 2025, bailproject.org/false-promise-of-bail.

    ---

    Disclaimer: This post is for general information purposes only. It does not constitute legal or tax advice. This post reflects the current opinions of the author(s) and are for educational purposes only. The opinions reflected herein are subject to change without notice.

    ---

    Images: All original images are AI-generated with using a combination of Dall-E, Firefly, DreamStudio and Pixlr.

Last week, a federal judge in the Northern District of California handed the AI industry a significant win regarding the use of copyrighted material to train large language models (LLMs). In Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson v. Anthropic PBC, the court held that training LLMs on copyrighted books constitutes fair use—so long as the books were lawfully acquired. In other words, the ruling doesn’t give carte blanche to go pirating. Just two days later, a different federal judge in the Northern District of California granted summary judgment to Meta (Facebook) in a parallel AI copyright infringement case, reaching a similar conclusion.

Taken together, these were significant rulings as they mark the first decisions to provide clarity on how AI companies can color within the lines as they train their models. But make no mistake—they don’t give free rein to swashbuckling AI firms eager to plunder the open seas of online content. Let’s dig into the courts’ reasoning—and what it means for those trying to build in this fast-moving space.

I. The Case: Bartz, Graeber, Johnson v. Anthropic PBC

A. Background

In 2024, a group of authors filed a class-action suit against Anthropic, alleging that the company had used pirated copies of their books to train its chatbot, Claude. They claimed their works had been downloaded from shadow libraries like Books3, LibGen, and PiLiMi without their permission. They were purchased in print, then scanned and digitized (destroying the originals).

Anthropic admitted to downloading and storing these materials, but argued that the use was protected under the “fair use doctrine” — a legal principle that permits limited use of copyrighted material without permission for purposes such as commentary, criticism, education, or transformative uses. For a refresher on the “fair use doctrine,” check out a prior post.

According to Anthropic, feeding books into an AI model was “quintessentially transformative.” The point wasn’t to replicate or compete with the original books—it was to teach the model how to write and reason like a human. The company also emphasized that its model did not reproduce specific works but rather generated original content. The intent was not to copy or mimic the original pieces and so was not a threat to said authors.

Anthropic’s library strategy was twofold:

1. To Train LLMs: It used selected books to train and fine-tune Claude, aiming to improve fluency, tone, and reasoning.

2. Amass a “Forever” Research Library: It maintained a massive, searchable internal repository of books—over 7 million, including pirated works—even if they weren’t immediately used for training.

It’s the 2nd part of the strategy that was problematic.

B. The Ruling

The judge had no problem with:

1. Using lawfully purchased books for training: Anthropic’s use of lawfully purchased books—even when scanned and digitized—qualified as fair use under U.S. copyright law. The training process was deemed “spectacularly transformative.” The judge likened it to a person reading thousands of books to learn how to write better. The model, Claude, didn’t reproduce or redistribute content—it synthesized new outputs based on patterns.

2. Digitizing printed books for internal use: The court compared this to converting a personal bookshelf into a searchable digital format. As long as the books were legally purchased, scanning them for internal AI training was not considered infringement.

However, Judge William Alsup ruled against Anthropic on the pirated material. Yarr.

The judge rejected the notion that downstream transformative use could cleanse the initial wrongdoing. Coupled with bad facts that Anthropic’s own communications revealed that it chose not to license books in order to avoid what its CEO described as a “legal/practice/business slog.” Such shortcutting undermined the fair use defense. We’ll have to await the court to rule on damages.

II. Kadrey v. Meta

A. Background/Ruling

The Meta case has a similar fact pattern to that of Anthropic: authors including Sarah Silverman, Ta‑Nehisi Coates, among others sued Meta for copying their books from “shadow libraries” to train its LLM model known as Llama.

The court granted Meta’s motion and dismissed the authors’ copyright claims tied to LLM training. Judge Chhabria applied a 4-factor fair use test of which 3 of the 4 weighed in Meta’s favor:

• Factor 1 (Purpose and Character): Point for Meta. Training Llama to generate new content, not to replace books, was a “highly transformative” use. Meta’s commercial motive and use of shadow libraries didn’t outweigh the shift in purpose.

• Factor 2 (Nature of the Work): Point for the authors. Their books were creative, but this factor carried little weight ultimately.

• Factor 3 (Amount of Work Used): Point for Meta. Copying full books was “reasonably necessary” to teach the model how to handle long-form, structured content.

• Factor 4 (Market Effect): Point for Meta—but with caution. Plaintiffs showed no real evidence of lost sales or regurgitated outputs.

  • Caution this with the fact that the court flagged that if LLMs were to flood the market with similar works and undercut authors indirectly, this factor could weigh in favor of future plaintiffs.

Ultimately, the judge implied to his ruling should be taken with a giant giant of salt: Meta won not because its use was clearly lawful, but because the plaintiffs “made the wrong arguments.”

III. Recent U.S. Copyright Office Report

These rulings land just weeks after the U.S. Copyright Office (USCO) released a report—cheered by many creators—that rejected the notion of inherent fair use in Generative AI (GenAI) training and tackled the key question: Is using copyrighted content without permission to train AI models considered fair use?

Its answer: …maybe. The USCO wrote that such a determination would be fact specific but courts should focus on the (i) transformative use and (ii) market effects to determine such use of copyrighted works.

For example, if an AI model is used for noncommercial purposes and doesn’t reproduce the copyrighted material in its outputs, that may weigh in favor of fair use. But if the model were trained on pirated data or if its outputs compete with the original work, expect heavy legal scrutiny. The USCO drew a clear line between benign, internal uses and high-risk practices that copy expressive works from pirate sources to generate unrestricted outputs — particularly when licensing options are available.

Crucially, the USCO rejected the argument that training on copyrighted content is inherently fair use, cautioning that language models don’t just learn meaning—they absorb the arrangement and structure of linguistic expression. If a model memorizes and regurgitates protected material, that could constitute infringement.

The USCO also avoided recommending compulsory licensing, favoring instead the emergence of voluntary and extended collective licensing systems to manage the complexity of content ownership and permissions. This framework could offer a scalable path forward, especially as courts begin grappling with the scope of market harm caused by GenAI applications.

And although the report isn’t binding, it’s already appearing in court filings. Its emphasis on provenance, market impact, and licensing alternatives reinforces the direction courts like that of Alsup and Chhabria appear to be heading.

IV. Takeaway

Bottom line: you gotta pay tribute for what you use. You can’t Captain Hook your way through training LLMs.To amass your treasure, you can’t loot the open seas of creative content without a proper Letter of Marque (i.e. you pay for using content).

These rulings are a partial victory for both creators and AI companies—and they chart out a map for future legal battles. Courts agree that LLM training is transformative, but how you get your training data still matters. Judge Alsup took a piece-by-piece approach, flagging piracy as a dealbreaker. Judge Chhabria looked at the end goal alone, but warned that economic harm could still torpedo a fair use defense.

Future plaintiffs should focus on provenance and market harm (bring better data like surveys, sales drops, or dilution studies) to tip the scale in this battle.

This is still the beginning of this battle between creators and AI companies. Expect appeals in both cases and a range of interpretations from other jurists until clearer appellate guidance emerges. Either way, these are the early days on the great battle on the high seas of GenAI and fair use.

The ships are built. The flags are raised. Grab the popcorn—this one’s going to be a blockbuster

.

Subscribe now

Share

Resources:

• The Ruling: https://fingfx.thomsonreuters.com/gfx/legaldocs/jnvwbgqlzpw/ANTHROPIC%20fair%20use.pdf

• Anthropic Scores a Landmark AI Copyright Win—but Will Face Trial Over Piracy Claims - Wired: https://www.wired.com/story/anthropic-ai-copyright-fair-use-piracy-ruling/

• Pirates, Privateers, and Civil War Maritime Laws - Library of Congress: https://blogs.loc.gov/law/2020/05/pirates-privateers-and-civil-war-maritime-laws/

• Copyright and Artificial Intelligence Part 3: Generative AI Training - US Copyright Office: https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Version.pdf

• US Copyright Office Issues Report Addressing Use of Copyrighted Material to Train Generative AI Systems - Schulte Roth: https://www.srz.com/en/news_and_insights/alerts/us-copyright-office-issues-report-addressing-use-of-copyrighted-material-to-train-generative-ai-systems

• Mixed Anthropic Ruling Builds Roadmap for Generative AI Fair Use - Bloomberg Law: https://news.bloomberglaw.com/ip-law/mixed-anthropic-ruling-builds-roadmap-for-generative-ai-fair-use

• Anthropic and Meta fair use rulings on AI model training - McDermott Will & Emery: https://www.mwe.com/insights/anthropic-and-meta-fair-use-rulings-on-ai-model-training/#:~:text=Comparative%20implications,tip%20the%20scales%20next%20time.

• Northern District of California Judge Rules That Meta’s Training of AI Models is Fair Use - Goodwin: http://goodwinlaw.com/en/insights/publications/2025/06/alerts-practices-aiml-northern-district-of-california-judge-rules

• Images: All original images are AI-generated with using a combination of Dall-E, Firefly, DreamStudio and Pixlr.

• Disclaimer: This post is for general information purposes only. It does not constitute legal advice. This post reflects the current opinions of the author(s). The opinions reflected herein are subject to change without being updated.

Here at The LawVerse, we don’t just take immense pride in what we write. We also meticulously craft the images you see in our articles. We’ll spend sometimes hours to get one image perfect through a mix of AI-generation and our own editing.

And, people have noticed. And by people, I mean my family member who has enlisted me to help them pull together images for a pregnancy-related app.

As we were building the app’s mise-en-scène, we debated whether to use stock photos (i.e. actual photographs) or AI-generated images (synthetic media). For a company on a budget, some of the pricier stock photo options were just out of reach. So, I turned to AI. The results were interesting.

I wanted to pass along what I learned through this process to help other startups who are trying to decide between stock photos and AI-generated images.

TLDR: you have to read the license to know what you are buying.

I. Why Choose AI and Not Real Photos?

AI-generated images can be a more attractive option than stock images for a few reasons.

• Customization: AI image generators let users custom-make images to fit a prompt within seconds. To get the perfect image, users can keep re-prompting the algorithm or edit the image later if they need to.

• Uniformity: AI image generators can keep the look consistent between the images, whereas stock photos may have different tones, lighting, or feel to them.

• Price: A subscription to an AI image generator can be much more cost-effective than buying licenses to many stock photos. Generating multiple images may be free up to a certain number of outputs or may require a subscription. Because I needed a lot of images, the cost for AI seemed more of a bargain.

• Ownership Rights: You can own the rights to use the output (i.e. images) without any restrictions....well, depends on the terms of use and the law in your country. Will discuss this more below.

II. AI-Generated Images: The Good, the Bad, the Too Perfect

Is this too good to be true? AI image generators have a few common pitfalls. Some are fixable. Others… not so much.

Let’s look at what I was trying to generate. Because this app is meant to appeal to a global audience, I tried to generate images showcasing diversity so that everyone could see themselves in the app. My prompting raised the following issues with using AI:

A. Homogeneity and Bias:

Just as humans have bias, so does the LLM creating the images. AI’s default person in an image will just look like the below: young, thin, pretty, and Caucasian. These default people are often too perfect–to the point of not looking human–and attractive by default.

Just as humans are biased, so too was the LLM creating the images. Now, I prompted it to reflect people with specific ethnicities to force it to be diverse, and the results were mixed.

Example Results - Good: I could get the result with some prompting.

Example Results - Bad:

Example Results - Bafflingly Bad:

After a few attempts, I finally realized that this is mixing an Orthodox man’s hat with the Orthodox woman’s outfit into something that looks like a cross between Guinan from Star Trek and Barbara Streisand from Yentil…..Oy, Captain.

I gave up trying to reprompt. Took the last image to Adobe Firefly to delete the hat thing and put on a winter beanie.

B. Weird Anomalies:

Besides bias, AI infamously generates weird things like extra limbs or random objects. Sometimes, everyone looks like they are wearing the same uniform, and the same person may appear in one image

III. What’s Causing These Odd AI Outputs?

Why do we get such strange results?

First, let’s look at what AI image generators do. Large language models (LLMs) like ChatGPT are complex algorithms designed to predict. In text, they act like fancy autocorrect machines: using the data they’ve been shown in the past (“trained” on) and what you give them, they play fill-in-the-blank to predict what comes next. AI image generators work the same way, using the data they’ve been given to make informed predictions about what you want to see.

While LLMs’ complexity makes it hard to pinpoint exactly why they produce these strange images, experts have a few theories:

• Training Data: LLMs only have the data we give them. If they’re trained on biased or outdated information, they’ll spit that bias back up.

• User Data: Besides training data, users’ prompts shape LLMs’ answers. A biased prompt creates a biased answer. To see this in action, look at Microsoft Bing chatbot, which produced less weird results when Microsoft limited how long users could chat with it.

• Dumb Computers: Unlike you and I, LLMs have no cultural or artistic knowledge. This means they can’t identify and correct their biased results unless their creators or users force them to. They also can’t use common sense or artistic principles to add texture to too-perfect skin, count how many fingers are on a hand, or stop hair from melting into the fabric of someone’s shirt.

IV. Do I Really Own the AI Output? Is There a Catch?

It feels too easy…..Let’s take a look at what ChatGBT’s terms of use say:

Ownership Means Commercial Usage Rights. The terms state that you own the images with certain limits and conditions. Ownership rights mean the right to use, modify, reprint, sell, merchandise, and distribute the images as you want – whether you generate the image for free or paid credit. [This wasn’t always the case. For example, with Dall-E 2, OpenAI owned the output.].

Note, this doesn’t necessarily mean you own the copyright.

Copyright (No Ownership in US). In the US, you can’t copyright AI-generated images because only works made by humans can be copyrighted (for more on this - please see this prior post). Having a copyright means you have the exclusive right to use or stop anyone from using your work (though there are some exceptions and the right doesn’t last forever).

Note, the American tradition isn’t followed in other countries:

• European Union (Ownership Is Possible): EU law doesn’t tell us much about the ownership of AI-generated works, so it’s really individual member states that must determine this question. Member states like Germany and France demand significant human input for copyright, excluding purely automated works.

• UK (Ownership Available): UK law requires that a work be the "author's own intellectual creation" showing a "personal touch." Under the Copyright Designs and Patents Act 1988 (CDPA), the UK extends copyright protection to "computer-generated works,” including those created by AI. The author is the person who made the necessary arrangements. The individual who made the “arrangements necessary for the creation of the work” would be considered the author of the work (though unclear who that person is making “necessary arrangements”).

• US (Not Really): What about if you really rework it…..are LawVerse images copyrightable?

  • In February 2023, the US Copyright Office awarded copyright protection to Kristina Kashtanova for the written content and the composite layout of text and images in her comic book "Zarya of the Dawn." However, the individual images produced by the AI tool Midjourney were not copyrighted because their unpredictable generation means the creator does not have enough control over their final appearance. Despite Kashtanova's significant involvement and creative input, the images alone were not eligible for copyright protection.

  • Most recently, in a January 2025 report from the U.S. Copyright Office, inputting prompts into generative AI, without more substantial involvement, does not grant the user copyright over the output, stressing that true authorship involves active control over the work’s expressive elements.

  • So for the LawVerse images, probably depends on how much I rework them….Ahem, I do!

Exclusivity. Because there is no copyright protection for an AI image, you don’t have any exclusivity rights to the output. That means that any AI-generated content is in the public domain (i.e., it’s not protected by IP laws….the public owns the works). So, while AI tools may promise you own the image, what they really mean is that you can use it (but so can everyone else). The fact that the LLM generated an image for you doesn’t mean you have exclusive rights. AI companies could also generate the same or similar output for someone else to use, and there’s nothing you can do to stop that.

Rights OpenAI Retains. Per the terms of use, OpenAI retains the rights to the model and use the generated content to improve its services. They may use your prompts and the resulting outputs for research, development, and improvement of their models and services. Additionally, OpenAI may employ the generated content for promotional purposes and marketing. They also reserve the right to monitor your usage to ensure compliance with their terms of service and content policies.

V. Can I Generate Anything I Want?

No, you have to comply with the license of the AI vendor.

For example, if you violate ChatGBT’s polices, OpenAI may suspend or terminate your access to its services, and you may be required to delete any content that violates their policies. Repeated or severe violations could include legal consequences, depending on the nature of the violation. Additionally, your account may be flagged or restricted from future use of OpenAI's tools.

OpenAI can also refuse to follow what it deems “harmful instructions…to produce harmful content.” Its universal policies require users to comply with laws, avoid causing harm, not misuse service outputs, and respect safety safeguards. This is known as content filtering, which effectively puts guardrails up before you can input anything into the model.

Specifically, the DALL-E 2 Content Policy prohibits creating, uploading, or sharing images that are not G-rated or could cause harm, including hate speech, harassment, violence, self-harm, sexual content, shocking subjects, illegal activities, deception, political content, health-related misinformation, and spam. Users must not mislead others about AI involvement and must respect the rights of others by not using images of people without consent, images without appropriate usage rights, or creating images of public figures.

VI. Do I Really Own the Output?

There are many cases pending before the US courts as to whether using copyrighted works without permission to train AI models that generate images are in violation of the original authors' works. If the algo is trained on images in violation of that creator’s copyright, AI companies promising that you own the image (or really the commercial use of that image) may be an empty promise. You could see creators trying to enforce their rights against AI users exploiting derivative works.

We just don’t know how the courts will come down on this. You will have to do your own risk/reward analysis here. Note, if you are trying to copy or imitate the style of another creator for your image, then that’s likely an increased risk.

VII. How Do My Rights under AI Compare to Those of Stock Photos?

Stock photos are professionally taken photographs available for licensing. The rights you get by using them are pretty similar to those of AI-generated images. Some of the most well-known photo companies include Getty Images, Shutterstock, and Adobe.

Ownership Means Commercial Usage Rights: When you purchase stock photos, you typically acquire a license that grants you the right to use the images under specific terms and conditions set by the stock photo provider. These rights generally include the ability to use, modify, reproduce, and distribute the images for commercial purposes. This does not mean you own the copyright to the image.

Copyright: Because photographs are taken by humans, they are eligible for copyright protection. Note that the copyright is usually held by the photographer or the stock photo agency, not by the customer of the stock photo company. Customers usually purchase a license to use the image, not the copyright. This means that while you can use the images according to the terms of the license, the copyright owner retains the exclusive right to enforce copyright protections against unauthorized use.

Exclusivity: Because photographs can be copyrighted, whoever owns the copyright can control its use and prevent others from using it. When a consumer licenses an image, you are likely obtaining the following types of licenses:

• Royalty-free licenses: You typically pay a one-time fee to use an image in various ways without paying any additional royalties per use. The license is non-exclusive, meaning the stock photo company can license the image out to someone else.

• Rights-managed licenses: Fees are based on a specific use, duration, or geographic region. Sometimes, this license may be exclusive for a fee.

• Exclusive licenses: The customer can obtain the right to be the exclusive user of the image (rarer in practice because it’s much more expensive).

• Duration: Stock photo licenses can vary significantly in duration depending on the type of license and the terms agreed upon. It could be into perpetuity, or it could be a few years.

Rights Retained by the Stock Photo Company: Stock photo providers retain certain rights over the images, including the ability to resell the same image to other clients (unless an exclusive license is purchased). They also set the terms and conditions of usage, which can include restrictions on how and where the images can be used, such as limitations on print runs, geographic regions, and types of media.

As an example, Shutterstock offers various types of licenses, including:

• standard licenses are royalty-free, allowing the same image to be used by multiple customers;

• enhanced licenses with broader usage rights but still non-exclusive;

• editorial licenses for non-commercial uses (i.e., news articles and educational materials); and

• premier licenses allow “sensitive use rights,” unlimited indemnification, third-party rights transferability, use in merchandise and templates, and high-resolution unwatermarked comps.

VIII. Bottom Line

IX. Extra Credit: Are NFTs like Stock Photos or AI-Generated Images?

An NFT gives the owner a unique digital certificate of authenticity, which bestows certain usage rights to the owner (sometimes to commercialize the underlying content). The copyright is retained by the original creator of the work unless the license itself transfers it. The rights should exist so long as you own the NFT.

When you buy a non-fungible token (NFT), it’s more akin to buying a license to a stock photo (assuming the NFT was not generated by AI). NFT projects are all over the place with how their licenses work (sometimes there is none).

So, NFTs are somewhere between an AI-generated image and stock photos.

Share

Subscribe now

[Special thanks to Jane Perov and Sam Silverberg.]

Resources:

• https://www.cooley.com/news/insight/2024/2024-01-29-copyright-ownership-of-generative-ai-outputs-varies-around-the-world

• https://www.reuters.com/legal/ai-generated-art-cannot-receive-copyrights-us-court-says-2023-08-21/

• https://www.novagraaf.com/en/insights/artificial-intelligence-and-copyright-who-owns-ai-generated-images

• https://www.forbes.com/sites/forbesbusinesscouncil/2024/04/16/who-really-owns-ai-art-what-businesses-need-to-know-about-generative-ai-and-intellectual-ownership/

• https://venturebeat.com/ai/who-owns-dall-e-images-legal-ai-experts-weigh-in/

• https://hbr.org/2023/04/generative-ai-has-an-intellectual-property-problem

• https://www.linkedin.com/pulse/who-owns-ai-generated-content-understanding-ownership-jha-pmp--bedoc/

• https://www.fairlicensing.com/en/blog/ai-generated-images-and-copyright#:~:text=Ultimately%2C%20the%20specific%20copyright%20laws,the%20necessary%20permissions%20and%20licenses.

• https://www.ropesgray.com/en/insights/alerts/2023/03/can-works-created-with-ai-be-copyrighted-copyright-office-issues-formal-guidance

• https://ipwatchdog.com/2023/12/12/copyright-office-affirms-fourth-refusal-register-generative-ai-work/id=170564/

• https://openai.com/policies/usage-policies/

• https://help.openai.com/en/articles/6338764-are-there-any-restrictions-to-how-i-can-use-dall-e-2-is-there-a-content-policy

• https://help.openai.com/en/articles/6425277-can-i-sell-images-i-create-with-dall-e

• Naruto v. Slater: https://law.justia.com/cases/federal/appellate-courts/ca9/16-15469/16-15469-2018-04-23.html

• Thaler v. Perlmutter: https://ecf.dcd.uscourts.gov/cgi-bin/show_public_doc?2022cv1564-24

• https://www.findlaw.com/legalblogs/legally-weird/who-owns-dall-e-images/

• https://www.nytimes.com/2023/02/26/technology/ai-chatbot-information-truth.html

• https://www.snexplores.org/article/ai-image-generators-bias#:~:text=Let's%20learn%20about%20bias&text=Bots%20can't%20know%20anything,images%20on%20which%20they%20trained.

• https://medium.com/@keithkisser/why-does-all-ai-art-look-like-that-f74e2a9e1c87

• https://www.thelawverse.com/p/so-what-if-my-ai-bot-wrote-this-paper-499

• https://support.shutterstock.com/s/article/what-is-shutterstocks-premier-license?language=en_US

• https://support.submit.shutterstock.com/s/article/What-is-a-Sensitive-Use-license?language=en_US

• https://fairuse.stanford.edu/overview/public-domain/welcome/#:~:text=The%20term%20%E2%80%9Cpublic%20domain%E2%80%9D%20refers,one%20can%20ever%20own%20it.

• https://openai.com/policies/row-terms-of-use/

• https://help.openai.com/en/collections/3643409-dall-e-content-policy

Images: All original images are AI-generated with using a combination of Dall-E, Firefly, DreamStudio and Pixlr.

Disclaimer: This post is for general information purposes only. It does not constitute legal advice. This post reflects the current opinions of the author(s) only. The opinions reflected herein are subject to change without being updated.

[AI-generated image:  Bob Ross was famous for saying “anyone could paint,” but he didn’t say anyone could own the copyright of their painting.]

As generative AI (GAI) continues to advance and blur the lines between human and machine creativity, we are forced to confront these fundamental questions:  is expression unique to humans, and are there limits to the protections of human expression?  As human expression becomes increasingly intertwined with technology, these questions will be stress-tested like never before, challenging our understanding of what it means to be creative, who deserves protection under the law, and whether the AI output of human expression will trip up copyright laws.  

In this article, I’ll dive into copyrights, the exemptions, and the lawsuits AI companies face for supposed copyright infringement.

I. How Does Copyright Work? 

Copyright is a legal right established by the US Constitution that protects original creative works once they are physically created or documented. This protection applies to both published and unpublished works. It covers a variety of creative expressions, including literature, drama, music, art, films, software, and architecture. However, copyright does not cover basic facts, ideas, or methods; it focuses on how these elements are expressed creatively.

A copyright is exclusively a human right.  The US Copyright Office (USCO) has issued guidance that anything generated by AI is not copyrightable except the parts not generated by AI -- to create is to be human.  Using copyrighted works without the copyright owner's permission violates her or his exclusive rights to reproduce and distribute these works.

[AI-generated image:  sadly, robots can’t hold copyrights.]

II. What Happens If Someone Violates Your Copyright?

You can defend your copyright by suing the violator in federal court; infringement could lead to penalties.  The penalty can be up to $30,000 for each copyrighted work misused. However, if it's proven that the infringement was done on purpose, this amount could go up to $150,000 for each work involved.  It could even lead to a criminal investigation by the US Attorney's office if someone intentionally uses your copyrighted work for their own profit.  

III. Are There Exceptions So I Don’t Need to Ask for Permission?

While the right to copyright is robust, it is not absolute.  There are exceptions to this right.  Educational and library uses, public performances, the first sale doctrine, accessibility provisions for disabilities, face-to-face teaching, and fair use are examples of copyright exceptions. 

Among these exceptions, fair use is worth exploring as applied to AI-generated content. An affirmative defense against copyright infringement, fair use supports freedom of expression by allowing certain uses of protected works without permission in specific contexts. Defined under Section 107 of the Copyright Act, fair use covers activities like criticism, teaching, comment, news reporting, scholarship, and research.  

Fair use is determined on a case-by-case basis and involves a court considering four specific factors: (1) the purpose and character of the use, such as whether it's for commercial or educational purposes; (2) the nature of the copyrighted work; (3) the amount and substantiality of the portion used in relation to the entire copyrighted work; and (4) the effect of the use on the potential market value of the copyrighted work.

As a generalization, if (i) you borrow a small portion of and even transform the original copyrighted work, (ii) that original work is more factual, less creative, (iii) you use it for noncommercial purposes, and (iv) using that work doesn't really hurt the copyright owner's original work or market, you likely can rely on Fair Use.  "Transformative uses" mean adding something fresh with a different purpose from the original content, bringing a new expression, meaning, or message to the work.

IV. Actual Examples of Fair Use Defense Failing or Winning

Fail:  Shepard Fairey’s Obama Hope Poster Was Not Transformative Enough

[Left: photograph taken by Mannie Garcia.  Right: Shepard Fairey “Hope” Poster”.  Images pulled from the New York Times article linked here]

The street artist Shepard Fairey created his artwork using a photograph of Barack Obama by Mannie García, which he did not initially credit. When the photograph's origin became known, Fairey claimed his work was protected under fair use. However, the court disagreed, ruling that Fairey's adaptation was not sufficiently transformative and did not qualify as fair use. The lack of original attribution to Garcia also weakened any argument that the poster was meant as commentary or critique of the original photo. 

Win:  Leslie Nielsen Spoofing a Vanity Fair Photo of Demi Moore 

[Left:  the original photograph by Annie Leibovitz for Vanity Fair.  Right:  a parody photograph of Leslie Nielsen for a movie ad. Images pulled from Mondaq article linked here]

In 1991, Annie Leibovitz photographed a pregnant Demi Moore for a Vanity Fair cover, creating a widely recognized image. Two years later, for the promotion of "Naked Gun 33 1/3: The Final Insult," Paramount Pictures parodied this photo by superimposing Leslie Nielsen's face onto the body of a pregnant woman, mimicking Moore's pose, lighting, and background, but with exaggerated features like garish lighting and a flashy ring, accompanied by a humorous caption. When Leibovitz sued for copyright infringement, the U.S. Court of Appeals for the Second Circuit ruled in favor of Paramount, declaring the ad fair use. The court recognized the parody, noting it did not harm Leibovitz's market, and dismissed concerns about deterring celebrity participation in such shoots as not a legitimate harm. This case has become a notable example in copyright discussions, illustrating the permissible boundaries of parody under the fair use doctrine.

[Note, I only provide these snippets of original images from news articles for purely factual purposes for my commentary, comment, and teaching about copyright lawsuits to illustrate how an image could “transform.”  I make no money off this.  I do this for the love of the game.]

V. AI Companies Don't Like to Ask for Permission

If Homer Simpson has a drinking problem, GAI companies have a data problem -- they need more and more data and can’t help themselves.  Why?  In AI, data equals power. Notably, datasets can provide more power to AI companies than the algorithm, for it is the data from which the algorithms learn to make decisions, function, interpret, and respond to new inputs.  The more diverse and extensive the data, the more intelligent and capable the AI becomes.  AI companies need vast amounts of data from varied sources—books, articles, online reviews—to teach their AI systems how to understand and interact with human language effectively.

[AI-generated image:  once you feed the monster you nurture, it won’t let you stop.]

It seems AI companies will go to extremes to get a hold of that data.  The New York Times reported that AI labs have exhausted nearly every source of reputable English-language text available online to train their AI systems. High-quality data, particularly from professionally written and edited sources like books and articles, is deemed most valuable for training AI to produce human-like outputs. While big tech companies like Google and Microsoft may have a user base generating vast amounts of user data, privacy laws, and corporate policies restrict their use for AI training. By 2026, the demand for high-quality data is projected to exceed its availability. The New York Times highlighted the lengths AI companies are going to for data: (i) Meta considered buying Simon & Schuster and using copyrighted internet data, (ii) Google amended its service terms to utilize what its users have written in Google Docs and Google Maps (reviews), and (iii) OpenAI developed a tool to transcribe YouTube videos despite potential policy violations.

VI. Data Addiction Leads to Lawsuits for Copyright Infringement

The lengths AI companies take to ingest data have prompted content creators to sue GAI companies for using copyrighted works without permission to train AI systems.  Many cases have cropped up against AI companies regarding their nonconsensual data scraping practices.  These lawsuits raise a bunch of questions as to whether: 

• (i) training a model with copyrighted data requires a license, 

• (ii) the AI output infringes on the copyright of the materials used to train the models, and 

• (iii) it is a direct copyright infringement or can it be considered fair use?   

Let’s look at Tremblay v. OpenAI, Inc./Microsoft as an exemplar of this trend.  A group of authors, including Sarah Silverman, sued OpenAI and Microsoft for using their books to train AI language models without permission.  The authors (plaintiffs) argue that AI companies should not be able to shield themselves by labeling AI output images as "new" when they are using the authors’ derivative works without permission.  Conversely, AI companies (defendants) will likely assert a fair use argument – that they are transforming the origin work into something distinct with a different purpose.  However, the “black box” nature of AI adds complexity to this argument.  The lack of transparency, where even the developers may not precisely know how their AI systems make decisions, especially in deep learning models, will make it difficult to prove fair use.  

The judge partially agreed with the defendants' request to dismiss the case, throwing out four of the six claims, but the authors can revise and resubmit. The following claims will move forward to trial: (i) direct copyright infringement claim (that OpenAI directly copied the authors' work and used it for commercial gain without permission), and (ii) unfair competition claim (that OpenAI unfairly competed with the authors by using their work without permission).  It was evident the defendants used Plaintiff’s copyrighted works without permission for profit so that could be an infringement and an unfair practice.  The court dismissed claims for (i) vicarious copyright infringement, (ii) negligence, (iii) unjust enrichment, and (iv) violations of the Digital Millennium Copyright Act.  

[AI-generated image:  If only Sarah Silverman and Sam Altman could resolve their differences via a thumb war.]

Vicarious copyright infringement is the idea that a company is responsible for the infringing activities carried out with its technology, even if it did not directly engage in the infringing activities itself.  Here, the judge reasoned that the plaintiffs did not allege the “ChatGPT outputs contain[ed] direct copies of the copyrighted books,” and therefore, the authors needed to demonstrate “substantial similarity” between the source materials and ChatGPT's outputs. AI-generated content is not automatically a copyright infringement.  

This decision follows a trend in similar cases, where courts are willing to consider claims of direct copyright infringement and unfair competition but are unsure about how to view AI outputs.  So far, courts haven't viewed using an author’s work without permission to train language models as clearly an automatic copyright violation.

VII. Bottom Line

[Civil litigation can take years from filing the initial complaint to trying the case.]

Given that the courts have been flexible in allowing plaintiffs to resubmit their complaints and point out the weaknesses in the cases, it’s too early to know how this will all ultimately play out.  For copyright holders to win a lawsuit against the AI companies, they need to show that the outputs from the models are substantially similar to the copyrighted content or contain parts of it.  Is the use of copyrighted materials to train a model and then generate new content transformative enough to fall under fair use?  

We shall see – but we're likely years away from a resolution.  Meanwhile, AI companies, realizing their vulnerabilities to lawsuits, are increasingly entering into licensing agreements with content holders. This development signals an acknowledgment that obtaining consent prior to using someone’s work is necessary.    In any case, litigation is expensive for companies so we may continue to see some CYA measures enacted by AI companies. 

[AI-generated image:  On the left we have the AI companies trying to break into the Elysian Fields of human creativity on the right.  OpenAI looks to be trying a click-wrap type strategy where you have to opt-out of data scraping.]

Lawsuit Tracker: 

• Paul Tremblay, et al., v. OpenAI, Inc. (NDCA) 

  • See above

• Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc. (DE)

  • Background:  Thomson Reuters filed a lawsuit against Ross Intelligence, alleging that the AI company illegally used Westlaw’s short summaries of points of law (headnotes) for training purposes, infringing on Thomson Reuters’ copyright.

  • Status:  In September 2023, the court ruled that the case could not be resolved on summary judgment due to disputed facts, and it is expected to go to trial in August 2024, with key issues including fair use and the transformative nature of Ross Intelligence’s AI training process.

• The New York Times Co. v. Microsoft Corp (SDNY)

  • Background:  The New York Times has initiated a copyright infringement lawsuit against OpenAI and Microsoft, arguing that Microsoft's chatbot, trained on its articles, hurts its business in two ways. Firstly, it diverts revenue by providing information that would have been accessed through Times' websites, like Wirecutter. Secondly, chatbots have emerged as a competing source of trustworthy information, threatening traditional news outlets like The Times and potentially damaging its reputation and revenue.

  • Status:  In  March 2024, OpenAI and Microsoft each filed motions to dismiss parts of the lawsuit, arguing that training their large language models (LLMs) didn’t disrupt the market for original news content.  

• Andersen et al. v. Stability AI Ltd. et al (NDCA)

  • Background:  The plaintiffs, a group of artists and content creators, allege that Stable Diffusion, Midjourney, and DreamUp were trained using copyrighted images without permission, resulting in copyright infringement. The claims include direct copyright infringement, creation of unauthorized derivative works, vicarious copyright infringement (based on user-generated content), violation of copyright management laws, unfair competition, violation of the right of publicity (for generating work "in the style of" specific artists) and breach of contract (against DeviantArt).

  • Status:  The defendants filed motions to dismiss, which the court granted in part in October 2023. Only one plaintiff's direct copyright infringement claim against Stability AI survived. The court allowed the plaintiffs to amend their claims, which they have done in an updated complaint. 

• Authors Guild v. OpenAI Inc. (SDNY)  

  • Background: The Authors Guild (including Jonathan Franzen, John Grisham, George R.R. Martin, and Jodi Picoult) is suing OpenAI for copyright infringement, claiming OpenAI used their books to train AI models without permission.

  • Status:  Still pending.

• J.L. et al. v. Google LLC (NDCA) 

  • Background: Filed by eight individuals who seek to represent millions of internet users and copyright holders, the lawsuit charges that Google's practice of scraping data from websites violates privacy and property rights. The plaintiffs contend that Google lacks the right to use their creative works, personal expressions, or any content shared online, simply because it is available on the internet.

  • Status:  In February 2024, Google filed a motion to dismiss a lawsuit, arguing that its generative AI models only use publicly available or lawfully obtained information for training. The company claims that the plaintiffs' privacy and security concerns are unfounded, as they have no property or privacy rights over information they shared publicly on the internet. Google asserts that, outside of copyright law and fair use provisions, there is no legal basis for controlling the use of publicly available information.

• Basbanes et al. v. Microsoft Corporation et al. (SDNY)

  • Background:  Journalists Nicholas Gage and Nicholas Basbanes sued OpenAI and Microsoft, accusing them of using copyrighted books and articles to train its AI model, ChatGPT. The complaint further points out that ChatGPT once had the capability to produce exact quotes from copyrighted texts but now offers summaries instead. According to the lawsuit, these summaries are considered derivative works that rely on the original materials that were copied.

  • Status:  Pending (filed January 2024).

• Kadrey v. Meta Platforms (now consolidated with Chabon v. Meta Platforms) (NDCA)

  • Background:  This case challenged Meta’s use of copyrighted books to train its LLaMA language models. 

  • Status:  The court dismissed most claims, including those asserting that all outputs of the model are infringing upon derivative works. The only remaining claim is for direct copyright infringement based on copies made during training.

• Makkai et al. v. Databricks, Inc. et al. (NDCA)

  • Background: Four prominent authors—Andre Dubus III, Susan Orlean, Rebecca Makkai, and Jason Reynolds—allege that Nvidia's NeMo Megatron models and Databricks' MosaicML models were trained on copyrighted works without permission. Nvidia has responded by asserting that their model creation complies with copyright laws, claiming fair use given the broad input sources and minimal influence of each work on the final model output.

  • Status:  Pending (just filed May 2024). 

Subscribe now

Share

Resources: 

• Image of Artist Robot – generated on Firefly and Dall-E

• Image of Rejected Robot – generated on Pixlr and DreamStudio

• Image of Sam Altman with Data Plant – generated on Firefly and Pixlr 

• Image of Thumb War – generated by DreamStudio and Firefly

• Image of Litigation Timeline - created with Visme

• Image of AI Companies Trying to Break Into the Elysian Fields of Human Expression – generated on Dall-E and Firefly

• What is machine learning (ML)?, IBM, https://www.ibm.com/topics/machine-learning (last visited Apr. 20, 2024).

• Randy Kennedy, “Artist Sues the A.P. Over Obama Image,” The New York Times, February 9, 2009, https://www.nytimes.com/2009/02/10/arts/design/10fair.html.

• Cade Metz, Cecilia Kang, Sheera Frenkel, Stuart A. Thompson and Nico Grant, "How Tech Giants Cut Corners to Harvest Data for A.I.," The New York Times, April 6, 2024, https://www.nytimes.com/2024/04/06/technology/tech-giants-harvest-data-artificial-intelligence.html.

• Cade Metz and Karen Weise, "Microsoft Seeks to Dismiss Parts of Suit Filed by The New York Times," The New York Times, March 4, 2024, https://www.nytimes.com/2024/03/04/technology/microsoft-ai-copyright-lawsuit.html.

• Michael M. Grynbaum and Ryan Mac, "The Times Sues Open AI and Microsoft Over A.I. Use of Copyrighted Work," The New York Times, December 27, 2023, https://www.nytimes.com/2023/12/27/business/media/new-york-times-open-ai-microsoft-lawsuit.html.

• Emilia David, "Sarah Silverman’s lawsuit against OpenAI partially dismissed" The Verge, February 13, 2024, https://www.theverge.com/2024/2/13/24072131/sarah-silverman-paul-tremblay-openai-chatgpt-copyright-lawsuit.

• Baker & Hostetler LLP, "Case Tracker: Artificial Intelligence Copyrights and Class Actions," BakerLaw, accessed May 10, 2024, https://www.bakerlaw.com/services/artificial-intelligence-ai/case-tracker-artificial-intelligence-copyrights-and-class-actions/.

• "A New Generation of Legal Issues Part 2:  First Lawsuits Arrive Addressing Generative AI," Perkins Coie LLP, April 20, 2023, https://www.perkinscoie.com/en/news-insights/first-lawsuits-arrive-addressing-generative-ai.html.

• “Recent Rulings in AI Copyright Lawsuits Shed Some Light, but Leave Many Questions," Perkins Coie LLP, December 14, 2023, https://www.perkinscoie.com/en/news-insights/recent-rulings-in-ai-copyright-lawsuits-shed-some-light-but-leave-many-questions.html

• Suzanne V. Wilson, "Artificial Intelligence and Copyright" United States Copyright Office, August 24, 2023, https://www.copyright.gov/ai/docs/Federal-Register-Document-Artificial-Intelligence-and-Copyright-NOI.pdf.

• Emilia David, "George R.R. Martin and other authors sue OpenAI for copyright infringement," The Verge, September 20, 2023, https://www.theverge.com/2023/9/20/23882140/george-r-r-martin-lawsuit-openai-copyright-infringement.

• Blake Brittain, "Google Hit With Class Action Lawsuit Over AI Data Scraping," Reuters, July 11, 2023, https://www.reuters.com/legal/litigation/google-hit-with-class-action-lawsuit-over-ai-data-scraping-2023-07-11/.

• Hailey Konnath, "Google Says AI Data Scraping Suit Still Doesn't Hold Up," Law360, February 12, 2024, https://www.law360.com/articles/1796997/google-says-ai-data-scraping-suit-still-doesn-t-hold-up.

• Stacey Chuvaieva, "Federal Judge Dismissive of AI Complaint:  Anderson v. Stability AI," Mitchell Silberberg & Knupp LLP, October 31, 2023, https://www.msk.com/newsroom-alerts-Federal-Judge-Dismissive-of-AI-Complaint.

• David Rabinwitz, "Using Copyrighted Works in AI Training Data May Infringe Even If the AI Output Doesn’t," Moses & Singer LLP, January 16, 2024, https://www.mosessinger.com/publications/using-copyrighted-works-in-ai-training-data-may-infringe-even-if-the-ai-output-doesnt.

• Isaiah Poritz, "OpenAI Hit With Another Copyright Suit From Pair of Journalists," Bloomberg Law, January 5, 2024, https://news.bloomberglaw.com/ip-law/openai-hit-with-another-copyright-suit-from-pair-of-journalists.

• Kyle Jahner, "Nvidia, Databricks Sued in Latest AI Copyright Class Actions," Bloomberg Law, May 3, 2024, https://news.bloomberglaw.com/ip-law/nvidia-databricks-sued-in-latest-ai-copyright-class-actions.

• Nancy Rubin and Steward McKelvey, "Canada:  The AGNS, Annie Leibovitz and the Naked Gun Lesson in Permissible Parody," Mondaq, September 4, 2013, https://www.mondaq.com/canada/advertising-marketing--branding/253026/the-agns-annie-leibovitz-and-the-naked-gun-lesson-in-permissible-parody.

• "Use of Excerpts from and Linking to Article Found to Be Fair Use," Winston & Strawn LLP,, October 26, 2010, https://www.winston.com/en/insights-news/use-of-excerpts-from-and-linking-to-article-found-to-be-fair-use.

• "Exceptions for Libraries & Archives," Copyright Alliance. https://copyrightalliance.org/education/copyright-law-explained/limitations-on-a-copyright-owners-rights/copyright-exceptions-libraries-archives/

• "Exceptions for Performances and Displays," Copyright Alliance. https://copyrightalliance.org/education/copyright-law-explained/limitations-on-a-copyright-owners-rights/copyright-exceptions-performances-displays/

• "First Sale Exceptions Copyright," Copyright Alliance. https://copyrightalliance.org/education/copyright-law-explained/limitations-on-a-copyright-owners-rights/first-sale-exceptions-copyright/

• " Exceptions for the Blind," Copyright Alliance. https://copyrightalliance.org/education/copyright-law-explained/limitations-on-a-copyright-owners-rights/copyright-exceptions-for-the-blind/

• " Exceptions for Educational Institutions," Copyright Alliance. https://copyrightalliance.org/education/copyright-law-explained/limitations-on-a-copyright-owners-rights/copyright-exceptions-educational-institutions/

Disclaimer:  This post is for general information purposes only. It does not constitute legal advice.  This post reflects the current opinions of the author(s).  The opinions reflected herein are subject to change without being updated.

This post originally appeared as a guest post on May 16, 2024 on @aisupremacy:

AI Supremacy

So What If My AI Bot Wrote This Paper!? 🏛️

🎧 Audio Intro: 4:05 Hello Everyone, I’ve been very stuck trying to think about AI and copyright in this wild-wild west era of Generative AI. I’m not a lawyer, but the lack of rule of law in America regarding AI and copyright has been stunning in historical terms…

Read more

2 years ago · 40 likes · 3 comments · Michael Spencer and Jessica Furr

Thanks to Michael Spencer (@aisupremacy) for reaching out and for the opportunity. Check out his Substack for the latest and greatest on AI.

Achieving a consensus on regulating AI, given its rapid evolution, complexity, and borderless impact, presents a formidable challenge. The conflicting and passionate viewpoints of the various stakeholders — regulators, doomers, or accelerationists — make it seem that consensus could never be possible.  So, we put that to the test.

[AI-generated image:  The debate arena where wits do battle –  the accelerationist Nick Land v. Doomer Geoffrey Hinton as refereed by FTC Commissioner Lina Khan]

We surveyed 43 people working in the AI industry (a statistically large enough sample size for generalizability) to gather their thoughts anonymously on regulating AI. The folks we surveyed work at big tech companies, banks, nonprofits, and startups, and their job functions spanned engineering and tech, business, legal & compliance, and non-profit and other functions. 

In reviewing those responses, we realized there was a surprising level of commonality.  We hope that sharing these AI insider viewpoints may debunk the myth of tribalism in AI and also better inform lawmakers that they have allies within the industry.  Below, you will find our high-level takeaways, the survey results, and some closing thoughts.  Enjoy!

Key Takeaways from the Survey (n = 43) (detailed results below)

1. Most respondents think that AI should be regulated (72% Yes and 19% Maybe). 

2. That number goes up even more for “high risk” use cases, such as AI systems dealing with biometric data, education, employment, law enforcement, etc. (88% Yes and 7% Maybe).

3. An overwhelming number of respondents believed that if the AI industry were to self-regulate, the companies should pledge to have the following measures or procedures in place:

   1. Require continuous evaluation for bias, security, and vulnerabilities (91%)

   2. Require the maintenance of policies and procedures (76%)

   3. Set standards for content provenance (i.e. where does the content originate), including watermarking, etc. (76%)

   4. Election integrity (52%)

Therefore, these are the low-hanging fruits for either creating regulations or setting standards for AI.

4. Interestingly, most folks think that there should be some kind of accountability for companies that create AI software, but they are not sure about what to do or how to go about it. These two questions delineated the dilemma:

   1. Do you think it’s fair that an AI company should be expected to know how the output generated by its technology is being used by its consumers? (44% Yes and 19% Maybe).

   2. Should AI companies be held liable for what their users do with the generated content? (42% Maybe, 16% Yes and 33% No).

5. There seems to be consensus around requiring social media platforms such as Facebook to take down harmful content (84% Yes) and be responsible for ensuring any synthetic content uploaded by users is labeled as AI-generated (79% Yes).

Final Thoughts

While it’s clear there is an industry desire for regulation, it’s also clear that many are unsure what it should look like in an ideal world.  We want to leave you with some final thoughts as you form your own opinions about AI regulation:

1. Regulation of emerging technologies, especially one as transformative as AI, should avoid negative externalities for industries and consumers.  You need a balance between fostering innovation and ensuring ethical, social, and legal safeguards. Good regulatory frameworks encourage creativity and experimentation while safeguarding against potential harm. 

2. Remember that there are already laws on the books that regulate AI as they would any industry.  The approach towards regulating AI should be surgical, using a scalpel instead of a sledgehammer. You don’t want to crush a fledgling industry with many positive use cases.

3. There is an entire ecosystem of stakeholders that need to be considered in designing AI regulations, including the builders of AI, the users of AI, and the platforms where AI content is published.

4. Given technology moves so fast, regulations should be designed not to favor one technology over another and not affect the development of the technology itself.  

5. Consider whether it makes sense for a certain technology to be regulated based on its tech capabilities, user behavior, user reach, and use cases.  Regulations may need a hybrid approach of different philosophies.

6. Soft laws or self-regulations may be more effective for fast-moving industries and should be designed with public-private partnerships.

7. Regulations should be designed to not become stale too quickly.  They should stand the test of time. (ex. The EU’s MiCA regulation was designed with Facebook’s Diem project in mind, but Diem has since shut down.)

Detailed Survey Results

[We want to thank everyone who participated in the survey.  We hope you feel you got your voice heard.  More to come!]

Subscribe now

Share

Resources:

• Survey conducted using Google Forms

• Image of Boxing Match Generated on Pixlr, ChatGBT and BeFunky

Disclaimer:  This post is for general information purposes only. It does not constitute legal advice.  This post reflects the current opinions of the author(s).  The opinions reflected herein are subject to change without being updated.

Frameworks aren't just for McKinsey consultants.  We lawyers are all about frameworks; anyone who went to law school remembers applying IRAC (issue, rule, application, conclusion) to a fact pattern on an exam.  So, let's break down all the legal issues that may arise in artificial intelligence (AI) – or at least we’ll give it a try.  Feel free to add to this by dropping comments below. 

Why do this? As a lawyer in an emerging industry, you must anticipate every potential pitfall, think creatively, and craft new and innovative legal solutions to protect your client in unchartered territory where precedents may be scarce. It’s truly exciting and invigorating for those legal eagles with the mettle to tackle the unknown.

The legal brain may at times feel like Mel Brooks in History of the World Part I (“Panic!”), but if we can war game AI like his son Max Brooks does with zombies, AI won’t feel like a foreboding hoard of the undead about to destroy us all. We hope….

We've applied an industry spotting framework that can be useful in the practice of law or business to the AI universe.  When legal issue spotting a particular industry or business scenario, we think about:

1. Regulation

2. Product

3. Customer/End User

4. Company

5. Competition

You may have some issues that could fall into more than one bucket, but hopefully, using this framework will help you develop a 360-degree perspective on many of the legal issues within this nascent industry.

Overall Framework: 

If you want to dig deeper into these topics, below are full explanations. Have fun!

I. Regulation

Federal Consumer Protection Laws

• The Federal Trade Commission (FTC):  Lying can be deemed an unfair and deceptive trade practice (Section 5 of the FTC Act).  The FTC asserts that this prohibition on deceptive or unfair conduct also applies to generative AI tools, including any actions that mislead or harm consumers, such as phishing scams, identity theft, deepfakes, and other forms of deception - Link

• Federal Communications Commission (FCC):  The FCC has prohibited AI-generated voices for unsolicited robocalls. - Link

• Securities and Exchange Commission (SEC):  The SEC has issued some recent fines for “AI Washing” in public company filings (meaning companies made false and/or misleading statements about their use of AI). - Link

Terrorism/Anti-Money Laundering (AML) Laws

• Office of Foreign Assets Control (OFAC):  Terrorists could use AI platforms for nefarious purposes (automate fake social media, spread disinformation, optimize logistics and planning of attacks, etc.), and OFAC could impose sanctions on AI companies or its code if they are found to be engaging with entities or persons on the OFAC sanctions list (think Tornado Cash). - Link

• Financial Crimes Enforcement Network (FinCEN):  The agency could view AI software as a potential tool used for money laundering and terrorist financing, and it can bring enforcement actions against a company that fails to comply with AML regulations. - Link

US State Laws 

• Deepfakes in Elections and Non-Consensual Sexual Content:  Various US states have begun to pass laws around deepfakes in regard to non-consensual sexual content and elections (below are a few examples). - Link. 

  • California

    • California law allows individuals to sue for non-consensual image-based sexual abuse (IBSA) deepfakes, holds creators and distributors liable, but it exempts deepfakes of public concern or newsworthiness.

    • Additionally, political candidates have a cause of action against deceptive deepfakes during elections, unless they are clearly labeled as manipulated or a parody.

  • New York has updated its right-of-publicity law to protect against unauthorized commercial use of digitally manipulated images for 40 years post-mortem, and separately, it prohibits the distribution of unauthorized IBSA deepfakes.  Disclaimers saying the representation is fake is not a defense.

• Employment/Use of AI for Hiring Bias:  In employment, AI-driven hiring tools might perpetuate discrimination if they inadvertently learn and replicate biases present in historical hiring data. New York passed the first law in the country to require companies to reveal how algorithms influence hiring employees. - Link

• Automatic Decision Making (aka Profiling Technologies):  This refers to systems in which decisions are made by AI without human involvement.  Algorithm based decision-making could be biased, which can impact consumers depending on how they are profiled.  Some states (California, Colorado, Connecticut, and Virginia) have passed laws integrating AI into their privacy laws, requiring notice, opt-out options, and consumer access to AI-driven decision-making processes. - Link

• Insurance Industry/Discrimination:  AI models might unfairly calculate premiums or deny coverage based on biased data, potentially discriminating against certain groups. As an example, the Colorado Division of Insurance is first in the nation to publish a final rule obliging life insurance companies to report how their AI models work. The purpose is to put into place AI governance and risk-management tools to combat discrimination and bias. - Link

• State-by-state Tracker: For more information, here’s a state-by-state tracker on AI legislation.

European Union (EU) Laws:  The EU has passed the first comprehensive AI regulatory framework in the world, which included requirements regarding transparency, disclosures, data quality, and human oversight.  It identifies high-risk and low-risk AI use cases. - Link

Climate:  The substantial computing power required for AI development may prompt the introduction of new climate regulations. - Link

II. Product

Intellectual Property (IP) Rights Issues

• Creation of Patents:  Per the US Patent and Trademark Office (USPTO), it is okay to use AI to create a patentable invention, but a "natural person" must provide a "significant contribution" to the invention to be eligible for a patent. The logic is that patents reward human ingenuity. - Link / Link 2

• Creation of Copyrights:  US Copyright Office (USCO) issued guidance that anything generated by AI is not copyrightable except the parts not generated by AI. - Link

  • AI-generated art cannot receive copyrights (Thaler v. Perlmutter of US Copyright Office (D.C. of District of Columbia). - Link

• Creation of Trademarks:  You base trademark rights on your use of the mark, not your creation of it. So, anything created by AI could be trademarked as long as it doesn't infringe on another's trademark. - Link

• Scraping Other Materials for Training:  There are open questions on scraping as to whether:

  • (i) training a model with copyrighted data requires a license,

  • (ii) the AI output infringes on the copyright of the training materials, and

  • (iii) it is a direct copyright infringement v. fair use? [The Fair Use Doctrine says there’s a limited exception for using a copyright without permission if it is “transformative.”]  If there's no direct copying, the owner of the data set being used by the AI company must show a substantial similarity in the output. 

    • Pending Litigation

      • There have been rulings by federal judges who have rejected the idea that the output of AI is an automatic copyright violation to those whose works were used to train the systems.  Direct copyright infringement causes of action have been permitted to go to trial. - Link

      • TLDR:   Presently, AI systems and rights holders are at odds over data collection, so you’re seeing AI companies enter into licensing agreements with content owners as a result.  Fair use may be a hard argument as a result.

• Ownership:

  • Custom AI Agents:  Companies are now letting you license their base models to build your own custom model.  What type of license are you getting?  Is it viral open-source software (OSS)?  If you're creating a custom model, who owns it?  How is liability allocated? Check the Terms of Use! - Link

  • AI-Generated Content:  Who owns the output that AI generates?  Check the Terms of Use! - Link1 / Link2

• Open v. Closed Source Code:  The code may be publicly available like any open-source software (OSS), meaning it can be scrutinized by anybody to see how it works.  With closed-source software (CSS), the code is not made public. - Link

  • Pros and Cons:

    • CSS allows the company to protect its IP, but it's also more of a black box regarding what it’s doing.  It leads to a concentration of power in the hands of centralized companies. On the flip side, CSS does mean that bad actors can't easily take its code and use it for nefarious purposes, so it may be easier to regulate AI at the company level and also impose liability. 

    • OSS allows for greater transparency, which can be good for strengthening the code for user development (less buggy, etc.) and regulatory scrutiny. 

    • The Biden administration has also weighed in on this debate to encourage CSS but VCs like Andreessen Horowitz disagree. - Link / Link2

• Data Sets: The algorithm (OSS vs CSS) actually may not be as important as transparency around data sets, which are paramount to train the models. The power of AI companies primarily stems from their control over extensive data sets through their various lines of business, which can give them more power than the code itself, as they determine the AI's capabilities and applications. - Link/Link2

  • “Synthetic Data:” Synthetic data refers to artificially created information by AI itself that’s used to train a model to avoid violating copyrights. Some refer to it as “laundering a copyright” because you wash the data to obfuscate the source. However, its repeated use can lead to degradation in quality, resulting in outputs that are merely less accurate copies of their originals (think AI inbreeding!). - Link

Product Harm

• Hacking:  Breaking into software can lead to theft of personal data.  If the code is impacted by bugs or viruses, it could lead to end-user lawsuits.  For example, if someone hacks into a Tesla’s computer and crashes the car, the manufacturer may be held responsible under a product liability claim.  - Link

• Image-Based Sexual Abuse (IBSA):  Without guardrails on technology, AI tools could unknowingly be used to create illegal or harmful content such as child sexual abuse material (CSAM).  Companies should institute policies to mitigate that risk and comply with local laws on CSAM reporting. 

  • Deepfake IBSA Material:  This is like the tort of nonconsensual IBSA material (colloquially referred to as “revenge pornography”) as it deals with a person's dignity and privacy of their likeness.  The difference is that with deepfake IBSA, there is a "distortion" that changes the way others may perceive you.  State laws differ on how they come down on deepfake IBSA.  States like CA and NY have instituted restrictions. - Link

  • Deepfake CSAM:  Under the Supreme Court decision Ashcroft v. Free Speech Coalition, fully fake CSAM can't be criminalized (i.e. no real children were involved in it).  However, if the CSAM pulled the faces of real children and morphed that into a new creation, that would probably be criminalized in each state as the government has a compelling interest to safeguard children.  - Link

• Deepfakes in Commerce:  The FTC is passing rules to make impersonation of government officials, businesses, and individuals a violation if deceptive to commerce.  Biden’s executive order also addresses deepfakes.  - Link / Link2

• Misinformation:  Internet companies are protected by Section 230 of the Communications Decency Act, which shields internet companies from liability for misinformation, which would include content generated by AI that users may post.  Question is whether Section 230 also applies to the AI output itself. AI could pull from data that's inaccurate or biased and that can lead to "hallucinations." - Link1 / Link2

  • FTC:  Lying can be deemed an unfair and deceptive trade practice under Section 5 of the FTC Act. - Link

  • Torts:  All of this could lead to torts lawsuits like for defamation, libel, right of privacy violation.

    • Example: There is a noteworthy Canadian case involving Air Canada being liable for negligent misrepresentation after its AI chatbot gave wrong information to a passenger. - Link1 / Link2

  • Election Laws:  States are beginning to pass laws to combat deepfakes in elections.

    • Example: California has a civil cause of action on the books for a political candidate deepfake video.  Check state laws! - Link

III. Customer / End User

Customer v. End User:  Who is the company's customer?  Sometimes, it's the end user and sometimes it’s a B-to-B situation.  Either way, the AI company should be concerned with how the product is used downstream and how liability is allocated amongst the various stakeholders.  - Link

Confidentiality / Privacy

• Confidential Information:  Many AI systems use the data a person inputs for training purposes, so users need to be careful to anonymize data, upgrade plans or change privacy settings. - Link

• Privacy Laws:  Given AI is trained on huge data sets, there may be an increase of personal data collected on people.  Companies may be collecting biometric or personal data on end users.  Each country has its own data privacy law (ex. GDPR in the EU) with which one would need to comply. - Link

  • HIPAA Violations:  Patient data may be collected. 

Discrimination and Bias

• Civil Rights Violations:  Title VII of the Civil Rights Act protects employees and job applicants from employment discrimination based on race, color, religion, sex, and national origin.  AI can show bias and may violate the Civil Rights Act, so under certain state or city laws, companies may need to test for bias. For example, NYC has passed such regulations (see Insurance and Automatic Decision Making in Section I) . - Link

IV. Company

Private Causes of Action:  AI companies and users should be prepared to be named in lawsuits concerning the unauthorized use of personal information or likeness, distribution of harmful or false content, and actions that invade privacy or cause emotional distress. Such “torts” include the following: defamation, libel/slander, false light claim, intentional infliction of emotional distress, right-of-publicity claim, non-consensual IBSA, CSAM, and expectation of privacy / invasion of privacy. - Link

Product Liability/Warranties/Liability/Indemnity

• Negligence v. Product Liability:  When there is no human involved in an incident (like a driverless car) or software gets a bug, it may turn a negligence case into a product liability case (manufacturing defect, a design defect, or a marketing defect).  - Link

• Allocation of Liability:  Liability may exist with the AI company (originator of algorithm) or maybe a seller of software incorporating that algorithm that then gets sold downstream to an end user.  Given this is a nascent industry, standards are still evolving. -  Link

• Software Developer Liability:  It has been hard to pierce the "tech veil" as courts have not recognized developers as fiduciaries or as having professional duties of care.  Courts have upheld posted disclaimers finding the downstream actors responsible.  This “tech veil” is always at risk of being eroded and some cases in the blockchain space are evidence of that. - Link

  • Deepfake Liability:  The FTC wants to hold software developers liable for deepfakes (more above).

Risk Mitigation/Self-Regulation

• Content and Safety Risk Assessments:  AI Companies should consider running red teaming and impact assessments.  The National Institute of Standards and Technology’s (NIST) AI Risk Management Framework provides a framework that can be useful.  - Link

• Content Moderation Policies:  Companies may consider having content regulation policies in place that put a guardrail on what can be produced to reduce consumer harms. - Link

• Audits:  Companies should consider allowing third-party auditors to evaluate their safety systems.  - Link1 / Link2

• Insurance:  Software providers may face liability for content produced by their software, so companies should be ready to be named as defendants in private lawsuits for defamation, right of publicity violations, and other legal issues. - Link

• Contracts/Terms of Service (ToS):

  • Implied Warranties:  AI companies want their ToS to disclaim the implied warranties of merchantability and fitness.  If left out, they are assumed as a matter of law.  - Link

  • Disclaimers, Waivers, Limitations of Liability and Consequential Damages:  Companies would want to make sure to include these in the ToS to mitigate their liability.

V. Competition

Antitrust:  General purpose AI requires substantial computing power that only the largest companies could afford.  Many of the largest AI companies are highly centralized and controlled by big tech.  It's very costly to create AI products, so it may discourage small companies from engaging in this new technology.  Big companies should play fairly and not stifle competition.  - Link1 / Link2

• Recently, the FTC launched inquiries into the big tech/AI deals. - Link

VI. Miscellaneous

Defenses for Regulating AI

• First Amendment Considerations:  Under Bernstein v. the Department of Justice (1999), the Ninth Circuit Court of Appeals ruled that software source code was protected speech under the First Amendment, and any government prohibition on its publication was unconstitutional.  - Link1 / Link2

• Section 230 of Communications Decency Act:  Section 230 of the Communications Decency Act provides limited legal immunity to service providers (and its users):  "No provider or user of an interactive computer service shall be treated as the publisher or speaker of any information provided by another information content provider."  Software developers may have limited legal immunity for unlawful acts done by their users.  - Link1 / Link2

• Supporting Growth of Technology:  Imposing liability on software developers may hinder AI development. Premature regulation risks creating harmful or outdated laws. Effective regulation balances public benefits with risks. Overly restrictive rules may increase costs, reduce product quality, or shut down projects, ultimately affecting users.  - Link1 / Link2

Practicing Law - Link

• Duty of Confidentiality:  Be careful what you input into an AI model.  Don’t use client names as it could violate attorney-client privilege.

• Local Rules:  Do the local bar associations and courts allow the use of AI in the practice of law?  Each state has different rules, and the Supreme Court hasn't opined yet. Check your local rules!

• Duty of Competence and Diligence:  Sometimes generative AI hallucinates, so you have to be careful you're not using incorrect information. Don’t trust. Verify and doublecheck…Triple check!

For those that made it this far, here’s a little humor as a reward:

Why did the hippy zombie eat the human brain instead of the AI bot’s neural network?

It never eats anything artificial.

Subscribe now

Share

Leave a comment

CREDITS:

Joke:  Human made

Images:

• MindMap created with EdrawMind

• Mel Brooks Chased by Zombies created by DreamStudio and speech bubbles

• Robot Comedian generated from ChatGPT

Disclaimer:  This post is for general information purposes only. It does not constitute legal advice.  This post reflects the current opinions of the author(s).  The opinions reflected herein are subject to change without being updated.