Internet and Computer Law Year in Review 2025–2026: AI Copyright Fair Use, the CDA Circuit Split, and the Age Verification Flood Gate

Tuesday, March 24, 2026 

Executive Summary

Instructor(s)
Ian Ballon, Greenberg Traurig
Wayne Stacy, BCLT 

Keywords

AI copyright fair use training data — Bartz v. Anthropic — Kadrey v. Meta Platforms — N.D. Cal. 2025 • Cox Communications v. Sony Music Entertainment — DMCA repeat infringer policy — willfulness — Supreme Court 2026 • CDA Section 230 — algorithmic feed product liability — Anderson v. TikTok Third Circuit — M.P. v. Meta Platforms Fourth Circuit • Free Speech Coalition v. Paxton — age verification — intermediate scrutiny — material obscene as to minors • Thaler v. Perlmutter — DC Circuit AI copyright — Supreme Court cert denied March 2026 • Copyright Office AI Report Part 2 — human authorship requirement — insufficient human control • Yuga Labs v. Ripps — NFT Lanham Act — intangible goods under the Lanham Act • Warner Chappell Music v. Nealy — copyright statute of limitations — discovery rule — Petrella conflict • McLaughlin Chiropractic v. McKesson — TCPA — Hobbes Act — agency deference overruled — Loper Bright • State AI laws preemption — Trump executive order December 2025 — California Frontier AI Act • “can AI-generated content be copyrighted under US law in 2026” • “is using copyrighted content to train AI models a fair use”

 

Legal Analysis

AI Copyright Fair Use After Bartz v. Anthropic and Kadrey v. Meta: Intermediate Copying, Pirated Source Material, and the Limits of the Transformative Use Defense

The most practically significant AI copyright developments of 2025 came from two decisions issued within days of each other by different judges in the Northern District of California — both holding that using third-party copyrighted content to train large language models can constitute fair use, but under materially different analytical paths and with a critical distinction that will shape future cases. In Bartz v. Anthropic [N.D. Cal. 2025], Judge Alsup granted partial summary judgment for Anthropic, holding that the use of plaintiffs’ works for training and the print-to-digital format change constituted fair uses, but held that copying of pirated library copies was not a fair use. Ballon identified this piracy carveout as consistent with a pre-existing line of cases holding that “when what is copied is itself a pirated copy and not the genuine work, then the use can never be fair.” Following the ruling, a $1.5 billion class action settlement received preliminary approval from Judge Alsup in July 2025. In Kadrey v. Meta Platforms, Inc. [N.D. Cal. 2025], Judge Chhabria granted summary judgment for Meta, holding that copying plaintiff authors’ works to train AI models was a fair use, recognizing the highly transformative nature of the AI training use. Both decisions are district court holdings from the same district, Ballon cautioned, and do not carry the precedential weight of appellate authority; the important fair use questions in the AI training context remain open for circuit review.

The copyright framework governing AI training liability is structured, in Ballon’s analysis, around the concept of intermediate copying: when third-party material is ingested to train an algorithm, a temporary copy is typically made from which information is extracted, even if the model’s output does not resemble the input. “The output often doesn’t look like the input” in most cases, Ballon observed, meaning that direct substantial similarity cannot be demonstrated; what is at issue is the intermediate copying of the training data itself, and whether that intermediate copying qualifies as a fair use under the four-factor balancing test. A copyright owner faces a structural evidentiary difficulty in these cases: there is no legal requirement for large language models to keep detailed records of what they have ingested. But Ballon noted that access can be presumed from striking similarity even when not directly proven, and that where works are available for license and the licensee bypasses the license to ingest them for training, the fourth fair use factor — effect on the market for the copyrighted work — strongly weighs against the defendant. The statutory damages stakes are substantial: up to $150,000 per work for willful infringement for timely registered works, subject to a circuit split on what constitutes a “work” for purposes of aggregating damages — with most circuits applying an independent economic viability test that counts individual songs on a registered album separately, but the Fifth Circuit holding that the four corners of the registration certificate controls.

The question of whether AI-generated output itself can be copyrighted was resolved against the claim in Thaler v. Perlmutter, No. 22-1564 (D.C. Cir. 2025), where the DC Circuit held that the Copyright Office did not act arbitrarily or capriciously in denying registration of an autonomously AI-generated work, and the Supreme Court denied certiorari on the day of the webcast in March 2026. The Copyright Office’s 2025 Part Two Report on Copyright and Artificial Intelligence consolidated the governing legal principles: copyright protects original expression in works created by a human author even if AI-generated material is also present; copyright does not extend to purely AI-generated material or material where human control over expressive elements is insufficient; and, as a baseline rule, prompts alone do not provide sufficient human input to be deemed protectable. Ballon identified a potential exception that the Report’s reasoning does not fully foreclose: in software development contexts, the human operator’s decisions about what to include and exclude in a complex prompt-driven workflow may involve substantially more expressive decision-making than a single creative prompt, suggesting that AI-assisted software development may have “more of an opportunity to create AI-assisted works that are themselves entitled to copyright protection.” He also identified the transactional implication of the non-protectability rule: where a company can achieve the same output in either one minute through AI or eight hours through human effort, enforcement of copyright in the output may require the inefficient human performance.

Free Speech Coalition’s Age Verification Ruling, Cox Communications’ DMCA Repeat Infringer Question, and the Copyright Statute of Limitations After Warner Chappell

Free Speech Coalition, Inc. v. Paxton, 603 U.S. ___ (2025), which upheld a Texas law requiring age verification for content deemed obscene as to minors under intermediate rather than strict scrutiny, carries implications for internet intermediaries far beyond the adult content platform context in which it arose. The Court held that there is no First Amendment right to avoid age verification for this category of material, and that the Texas law advanced an important government interest while only incidentally burdening adult speech. Ballon identified the case as a doctrinal departure from ACLU v. Reno, 521 U.S. 844 (1997), in which the Supreme Court struck down the criminal provisions of the Communications Decency Act targeting material obscene as to minors, finding them to impinge on adult First Amendment rights. “In broad strokes, you see a very different Supreme Court in ACLU v. Reno in 1996 … very concerned about regulation and how it could impact free speech, and then, many decades later, a more conservative court” applying intermediate scrutiny to age verification requirements that only incidentally burden adult speech. Ballon predicted that the Paxton holding “does give an impetus to state legislators who seek to impose age verification obligations on companies” in a wide range of contexts well beyond adult content — including App Store access restrictions, social media age gates, and other state laws seeking to protect minors from digital harms.

Cox Communications, Inc. v. Sony Music Entertainment, pending (U.S. 2026), presents the Supreme Court with two questions that will reshape DMCA safe harbor practice across the country. The Fourth Circuit vacated the district court’s billion-dollar-plus jury verdict against Cox for contributory copyright infringement but found that Cox — whose graduated response program effectively operated on a thirteen-strike model rather than the three-strike framework most courts have associated with an adequate repeat-infringer policy — was not entitled to the DMCA safe harbor. The first question before the Court is whether a service provider can be secondarily liable merely because it knew users were infringing without proof it affirmatively fostered or intended to promote infringement; the second is whether knowledge of another’s direct infringement suffices to establish willfulness under Section 504 of the Copyright Act, 17 U.S.C. § 504. Ballon reported that the justices appeared “somewhat conflicted” — concerned about the massive verdict’s implications for universities that could lose institutional internet access over a single infringing student, but also skeptical of Cox’s position that no action on generalized infringement knowledge is ever required. He added the separate Second Circuit ruling in Capitol Records, LLC v. Vimeo [2d Cir. 2025], affirming the service provider’s position that loss of DMCA protection for red flag knowledge requires knowledge of specific files or specific infringing activity, not generalized awareness — the same specific-knowledge standard at issue in Cox.

The copyright statute of limitations landscape was complicated by the interaction of two Supreme Court decisions that Ballon characterised as “wildly inconsistent.” In Petrella v. Metro-Goldwyn-Mayer, Inc., 572 U.S. 663 (2014), the Court held that laches is not a defense to copyright infringement because the three-year statute of limitations adequately addresses delay, and that a plaintiff who sues in year fifteen of ongoing infringement may only recover damages going back three years. A decade later, in Warner Chappell Music, Inc. v. Nealy, 601 U.S. 366 (2024), the Court held that a plaintiff with a timely copyright claim under the discovery rule may recover damages regardless of when the infringement occurred — potentially reaching back well beyond the three-year Petrella window. Ballon observed that the Court in Nealy assumed without deciding whether the discovery rule even applies in copyright cases, leaving “for another day” a threshold question whose unresolved status encourages plaintiffs to bring time-barred damage claims in the hope that the discovery rule will be held applicable. The practical effect, Ballon predicted, is to “encourage copyright owners to bring cases where they have damage claims that otherwise would be time-barred under the Petrella rule,” making the statute of limitations a source of strategic uncertainty for internet platform defendants. McLaughlin Chiropractic Associates, Inc. v. McKesson Corp., 604 U.S. ___ (2025), which held that courts are not bound by FCC rulings in TCPA enforcement proceedings, was identified by Ballon as following the logic of Loper Bright Enterprises v. Raimondo, 603 U.S. 369 (2024), overruling Chevron deference, and predicted that it will substantially weaken TCPA troll litigation premised on broad FCC interpretations.

The CDA Section 230 Circuit Split on Algorithmic Feeds, NFT Trademark Doctrine After Yuga Labs, and the State AI Law Preemption Question

The most consequential and contested trend in Section 230 CDA jurisprudence is the emerging circuit split on whether an interactive computer service’s algorithmic curation and delivery of third-party content constitutes the service’s own expressive activity — placing it outside the CDA’s immunity — or remains integral to the publishing function that Section 230 was designed to protect. In Anderson v. TikTok, Inc. [3d Cir. 2024], the Third Circuit held that TikTok’s algorithm, which determined whether third-party content appeared on personalized user feeds, constituted TikTok’s own expressive activity, and therefore permitted product liability, negligence, and wrongful death claims to proceed outside the CDA’s immunity. Ballon characterised this holding as “absolutely wrong,” observing that “just because something is AI generated or algorithmically generated, it’s still third party content,” and that publication decisions inherently are protected by the First Amendment, which is foundational to the CDA’s rationale. The Fourth Circuit took the directly opposite position in M.P. by and through Pinckney v. Meta Platforms, Inc. [4th Cir. 2025], holding that Meta’s use of an algorithm to deliver content was integral to its publishing function and did not take it outside the CDA, and that the CDA preempted claims for negligence, strict product liability, and negligent infliction of emotional distress. Ballon cautioned that while the Fourth Circuit’s decision is correct, the mathematical reality of Supreme Court composition — with Justices Thomas, Alito, and Gorsuch already expressing hostility to CDA immunity — makes seeking certiorari to resolve the circuit split a strategically perilous choice for technology companies.

In trademark law, the Ninth Circuit’s decision in Yuga Labs, Inc. v. Ripps [9th Cir. 2025] produced the first circuit court opinion specifically addressing NFTs under the Lanham Act, holding on two important grounds: first, that an NFT constitutes goods under the Lanham Act, resolving the threshold question of whether non-fungible tokens are subject to federal trademark protection; and second, that there is no First Amendment defense under Rogers v. Grimaldi where the defendant’s use of the plaintiff’s mark functions as a source identifier for the defendant’s own goods — consistent with the Supreme Court’s holding in Jack Daniel’s Properties, Inc. v. VIP Products LLC, 599 U.S. 140 (2023). On the Rogers v. Grimaldi First Amendment question, Ballon highlighted the contrast with Hara v. Netflix, Inc. [9th Cir. 2025], where Netflix’s fleeting animated use of a drag queen performer’s likeness in its Q-Force streaming show was held to have artistic relevance and therefore was not actionable — illustrating the two ends of the post-VIP spectrum between source-identifying uses (to which Rogers does not apply) and genuinely expressive uses (to which it may). The Dewberry Group, Inc. v. Dewberry Engineers, Inc. [U.S. 2025] decision, holding that Lanham Act profits may only be recovered from a named defendant and not from separately incorporated affiliated entities, reinforces the practical importance of adding all entities from which monetary recovery is sought as defendants at the outset of trademark litigation.

The federal-state tension over AI law creates one of the most significant legal uncertainties in internet practice entering 2026. California, New York, Utah, Texas, and other states have enacted AI-specific laws addressing risk assessments, transparency disclosures, whistleblower protections, safety protocols, and high-risk AI interaction disclosures. In December 2025, President Trump signed an executive order purporting to enjoin enforcement of these state AI laws. Ballon assessed the order’s legal authority with careful precision: as a general proposition, a presidential executive order cannot preempt state law absent enabling legislation from Congress; however, an argument could be advanced under the field preemption doctrine that AI regulation is a matter uniquely susceptible to national regulation under the Commerce Clause, and “the existence of an executive order signed by the president could be cited in support of that.” The more pragmatically significant consequence, Ballon predicted, is that the Department of Justice — directed by the order to implement the enforcement ban — could enter litigation on the side of corporate defendants or independently sue to enjoin enforcement, making state AI law enforcement practically risky even where its legal validity is ultimately sustained. Ballon also identified the SEC’s enforcement in the AI washing area — the practice of exaggerating AI descriptions in securities filings — as a material compliance risk, and noted that more than 150 court cases in 2025 had called out, sanctioned, or chastised lawyers for submitting briefs containing AI hallucinations and citing cases that do not exist, a problem that “after that highly publicized case” in Mata v. Avianca [S.D.N.Y. 2023] remains pervasive despite widespread awareness.

Generated by AI based on the Interview/Transcript below.

Key Takeaways

• AI Training on Pirated Copies Is Not Fair Use: Ballon identified Bartz v. Anthropic’s most durable holding as the piracy carveout: copying that is itself a pirated copy cannot qualify as fair use, regardless of the transformative nature of the AI training use; where training data is sourced from legitimate copies, both Bartz and Kadrey v. Meta recognized the highly transformative character of AI model training.
• Cox v. Sony Will Define DMCA Safe Harbor’s Outer Limits: The Supreme Court’s pending decision in Cox Communications v. Sony Music Entertainment will resolve whether generalized knowledge of infringement — as distinct from knowledge of specific infringing files — can defeat the DMCA safe harbor and whether it constitutes willfulness under 17 U.S.C. § 504; Ballon described the justices as “somewhat conflicted” on both questions.
• Paxton Opens Age Verification Flood Gate: Free Speech Coalition v. Paxton’s application of intermediate scrutiny to age verification for content obscene as to minors gives state legislatures a new doctrinal foundation for a wide range of age-gate requirements beyond adult content; Ballon predicted that “we’re going to see more of” age verification laws in 2026 across multiple intermediary categories.
• Third/Fourth Circuit CDA Algorithmic Feed Split Is Live: Anderson v. TikTok (3d Cir. 2024) and M.P. v. Meta Platforms (4th Cir. 2025) are directly irreconcilable on whether algorithmic feed curation takes a platform outside Section 230 immunity; Ballon predicted companies will be reluctant to seek certiorari given the hostile posture of three Supreme Court justices toward the CDA.
• AI-Generated Content Remains Uncopyrightable: Thaler v. Perlmutter (D.C. Cir. 2025) was affirmed and cert denied by the Supreme Court in March 2026; the Copyright Office’s 2025 Part Two Report confirmed that purely AI-generated material or material where there is “insufficient human control over the expressive elements” is not protectable, with prompts alone insufficient to constitute authorship.
• Licensed Training Data Defeats Fair Use Fourth Factor: Ballon identified the market-substitution logic as the most powerful argument against AI training fair use: “if you use material that is available for license for training algorithms, and you do not pay the license, you are going to have a very difficult time establishing fair use” because the infringement adversely impacts the market for the copyrighted work.
• Nealy-Petrella Conflict Invites Time-Barred Damage Claims: The Warner Chappell Music v. Nealy discovery-rule holding, combined with the Court’s assumption-without-deciding that the discovery rule applies in copyright, creates an invitation for copyright owners to bring cases with pre-three-year damage claims that Ballon described as “wildly inconsistent” with the Petrella laches rationale.
• NFTs Are Goods Under the Lanham Act: Yuga Labs v. Ripps (9th Cir. 2025) held as a first-impression circuit ruling that an NFT constitutes goods under the Lanham Act, resolving a threshold question for trademark enforcement in the digital asset space, and confirmed that Rogers v. Grimaldi does not apply where the defendant’s use functions as a source identifier.
• State AI Law Enforcement Is Practically Uncertain Despite Executive Order’s Dubious Authority: Ballon assessed President Trump’s December 2025 executive order purporting to enjoin state AI law enforcement as legally questionable but pragmatically significant, because DOJ entry into litigation on the side of corporate defendants could deter state enforcement even if the executive order’s preemptive authority is ultimately rejected by courts.

B-CLE Recording (CLE: $115) | Youtube Recording | Resource(s)| Speaker Bio(s) & Contact Info

Download the interview/transcript and slides here!

Interview/Transcript

This transcript is one of a two-part program series, “Internet & Computer Law Year in Review,” held on March 24, 2o26. The internet continues to reshape how businesses operate, communicate, and innovate—bringing new legal challenges along with it. In this engaging discussion moderated by Wayne Stacy, BCLT, Ian Ballon, Greenberg Traurig, explores emerging developments in areas such as artificial intelligence, online platforms, cybersecurity, and digital content regulation. 

Wayne Stacy  00:30

So, welcome everyone to the Berkeley Center for Law & Technology’s webcast. We have a special, special program that we try to do every year. I would say, if you’re in the world of computer law and internet law. If you watch one program, this should be it, because there’s nothing here that — in all the other programs we do, that’s all right here at once. So it is a program that, like I said, takes a little time to get through. It’s not your 10 minute quick hit, but it is what you need to know, what’s happened last year and what’s coming this year. So with that said, I’m going to introduce Ian. Doesn’t need much of introduction. If you’re here and you don’t know Ian, you probably tuned into the wrong program. Everybody in this industry knows Ian. Ian is the, I think, co-chair now of the Greenberg Traurig Global Intellectual Property Technology Group Practice. Ian has a treatise that gets bigger and bigger every year. But if you like, I said it’s the right title, E-commerce and Internet Law from West updated just recently. It is incredible treatise, and what you get here is the work product he’s been putting in to keep that treatise updated. So instead of having to read, somebody can help you walk through the highlights. And that’s the beautiful piece of it. Ian, as always, it’s great to be together again. I think — I was doing some calculations. We’ve been working together a little over 20 years on educational material, starting in Colorado.

 

Ian Ballon  02:10

Wow, time flies when you’re having fun, I guess.

 

Wayne Stacy  02:13

It is. I think from the very beginning, I’ve seen the growth of this program where it went from probably you could say everything you needed to know about internet and computer law in about 35 minutes, and now it’s all you can do in two hours to get the highlights out. And your treatise has gotten bigger and bigger because this area has gotten more and more complex. So it’s been been fun growth for a dumb patent litigator like me to see kind of what these other areas turn into. So with that, I appreciate it, and I look forward to learning myself. I’m going to turn it all over to you.

 

Speaker 1  02:52

Thanks, Wayne, thank you for that generous introduction. It’s always an honor and a pleasure to present to the Berkeley Center for Law and Technology. I’ve got two presentations. This one here, Internet and Computer Law Year in Review is going to be focused generally on Internet and AI, IP, Privacy, all of the the major issues. It’s a three day course that I’m compressing to about an hour and a half. And then there is another presentation focused more on data privacy, cyber security breach and AI class action litigation. And so we’re breaking it out into two separate presentations. This is the general one, and let’s jump right in. I do have some materials, and the slides are available. I have dense slides because it’s helpful for me in order to compress a three day presentation into a short period of time, to have everything right in front of me. And I always find it helpful to have more detail to refer back to even if you can’t go through all of that during the presentation. So let’s jump in. Wayne mentioned my treatise. I’ve included some materials from the treatise. It’s a couple of years old, because I’ve got a new version coming out later this year, which is more AI focused, internet law, e-commerce and litigation in the era of artificial intelligence, but the sections that I included are still highly relevant, and the slides have the most current case law, including cases from literally just the past month. So let’s jump in. Year in Review. One of the big developments last year we had the emergence of a popular artist, Solomon Ray. Here’s Solomon Ray’s YouTube page, 41,000 subscribers, 15 videos. Solomon Ray has an Instagram page, and yet Solomon Ray doesn’t exist. Solomon Ray is AI generated and this is where we are in 2026. We are in a world where what seems to be real and is real may be the result of algorithms, as opposed to human interaction. So with that, let’s jump in, and I’ll quickly give you an overview of what I’m going to cover in this presentation, as well as mention a few developments that don’t really fit in neatly anywhere else in the presentation, but I would be remiss to omit in a year in review program. So we will talk about AI law, litigation and trends, including some fair use decisions. Very interesting and important question is the extent to which state AI laws, which are now in effect in California, New York, Utah and elsewhere, have been preempted by an executive order that President Trump signed in December 2025. So we’ll drill down into that and talk about what law governs here. We’ll talk about important new case law in the IP space and First Amendment area, including the first NFT opinion, and that is an interesting new development from the Ninth Circuit under the Lanham Act. We’ll talk about CDA case law and trends, because I think there are some important things going on there that are worth talking about. I will talk a little bit about data privacy, cyber security breach and TCPA and AI class action litigation. But the main part of that is going to be in the companion presentation, which is also available this year from the Berkeley Center for Law and Technology. And so I’m going to take out of this presentation much of that, except high level that I’ll be talking about here, and then drill down into much greater detail in that separate program. In the cyber security breach area, we’ve seen in 2025 a continued growth in litigation. I don’t have the 2025 statistics, but in 2024 there were over 1500 data breach lawsuits. I have no reason to believe that that has slowed down. In 2025 about a quarter of the data breaches arose in the healthcare sector. We are seeing more ransomware incidents that actually surprised me a little bit. But ransomware is way up. Continued issues involving supply chain breaches and continuing costs. In the TCPA area, text and marketing, there’s a lower volume of cases, and I’ll talk about this more in the companion program, but there are some interesting developments, including a couple of US Supreme Court cases that are scaling back the potency of TCPA troll litigation because it is now harder for plaintiffs to rely on FCC rulings in that area. In the privacy and ad tech space, there are important issues involving CCPA/CPRA litigation. The Illinois biometric Information Privacy Act continues to be a big issue. I joke with startup clients that if they can only afford to comply with one privacy law that’s the one to comply with, because the cost of statutory damages is so high in that area. We also see a, really an explosion of wiretap troll claims. There’s also interesting developments in the air, interplay between class action litigation and mass arbitration. So as I said, I’ll talk, I’ll say a few things in this program, but really, to drill down, I’m going to do that in the separate year in review program that’s focused on data privacy, AI, and cyber security breach class action litigation. I’ll talk a little bit about online and mobile contract formation. That continues to be an area where, in the Ninth Circuit, it is difficult to get online and mobile contracts enforced the same way that you can in most other circuits, and that has significant implications for companies, both as plaintiffs and defendants and from a compliance perspective. A few specific things that I want to just mention, which I’m not going to cover in greater detail in the presentation, but they are important developments from the past year — California enacted the California Delete Act, this is applicable to data brokers and will allow consumers to provide a single opt out that takes effect later this year in August. Also enacted into law in 2025 the Take It Down Act, which is a federal law addressing non consensual online publication of real and computer generated sexually explicit images.

 

Ian Ballon  10:31

The interesting development the FTC in 2025 — we saw enforcement of the consumer review fairness rule. That is a rule from the Biden era. And yet we saw in 2025 that the FTC under President Trump is sending out warning letters and enforcing that law, that law bans non disclosure agreement, retaliation clause and restrictive agreements in connection with consumer reviews. And so that’s an important law to be aware of, or at least an important rule to be aware of, based on FTC enforcement. Finally, there is a increasing trend of age verification laws. It was an important Supreme Court case, which I’ll talk about in a moment from 2025 which I think really opens the flood gates to state legislatures to require intermediaries to engage in age verification. And this is something that I think we’re going to see more of in 2026. So with that in mind, let’s jump in to Supreme Court cases. That always seems like a good place to start, from my perspective, because Supreme Court obviously sets the law in a way that the circuit courts and state courts simply do not. Some interesting developments and an interesting case to watch, as well as a case from 2024 that I mentioned last year. But I want to mention again, because in the internet space, it continues to be important. That’s the Warner Chappell Music, Inc. v. Nealy, which in 2024 the Supreme Court held that a plaintiff with a timely copyright infringement claim under the Discover rule, is entitled to damages regardless of when the infringement occurred. And the reason I mentioned this again in 2025 is this impacts copyright litigation. It impacts all internet companies. Back 10 years ago, in Petrella v. Metro-Goldwyn-Mayer, the US Supreme Court held that laches is not a defense to copyright infringement because of the three year statute of limitations, which takes care of delay. So if infringement started 15 years ago, but a lawsuit was only filed this year, a plaintiff generally could only recover damages going back three years. And so the Supreme Court 10 years ago said there was no need for laches. But then 10 years later, in the Neely case, the Supreme Court said that if the Discover rule applies, then you could reach back earlier. Those two cases, I think, are wildly inconsistent but the effect of Nealy is to encourage copyright owners to bring cases where they have damage claims that otherwise would be time barred under the Petrella rule. The other interesting thing about this case, which I think we see too frequently in Supreme Court cases these days, is that the Supreme Court assumed, without deciding, that the Discover rule applies. What that means is it leaves for another day the question of whether the Discover rule should should even apply in copyright cases. And you know, again, I can say facetiously, as a litigator, it’s always, always terrific when the Supreme Court leaves open cases to be decided at a later time, but as a practical matter, from a public policy standpoint, and for those of you who are transactional lawyers, avoiding the kind of certainty that you’d have from a definitive ruling is a little bit frustrating. Another case I want to mention, which hasn’t yet been decided, but is before the US Supreme Court, is Cox Communications v. Sony Music Entertainment. That is a case that you are probably aware of because it’s been in litigation for many years. It involved a very large billion dollar damage award, which was then vacated by the Fourth Circuit, and Cox Communications had been sued by Sony Music. And the district court, my former law partner, Liam O’Grady, a judge in the Eastern District of Virginia, found that that Cox was not entitled to the DMCA Safe Harbor defense because they did not have an adequate policy of terminating repeat infringers in appropriate circumstances. We know that most courts in the United States generally uphold a three strike rule in DMCA cases because Americans understand and like baseball and three strikes seems like a fair way to evaluate whether a company has a policy on terminating repeat infringers in appropriate circumstances. A repeat infringer, by definition, is an infringer who has infringed once before, but Americans like baseball and so no court to date has struck down a three strikes policy, but Cox had something more like a 13 strike policy. It can be justified, because for internet service providers, taking away home access to internet is much more significant than just terminating the right to access a website, and so many ISPs have what are called graduated response programs where there are increasing penalties and increasing warnings, and that’s essentially what Cox had. But you know, 13 strikes, that’s not baseball. That starts to look more like cricket. Americans don’t really understand cricket, and in any case, they were found to not have the MCA Safe Harbor. A jury entered a billion dollar plus award, and the case is now pending before the US Supreme Court on two questions: whether the US Court of Appeals for the Fourth Circuit erred in holding that a service provider can be liable for materially contributing to copyright infringement merely because it knew that people were using certain accounts to infringe and did not terminate access without proof that the service provider affirmatively fostered infringement or otherwise intended to promote it. And that gets back to a number of court cases that have held both with respect to secondary copyright infringement and the DMCA, that to either establish liability or take a company out of the Safe Harbor, what is required is knowledge of specific acts of infringement, specific files or specific actions, not more generalized knowledge that a site or service could be used for infringement. And so that’s a very important proposition, and one that will be addressed. Liability, of course, can be imposed for fostering infringement or for either contributory infringement or inducement of infringement. But the important question is, when there is generalized knowledge, is that sufficient? And then the second question whether the Fourth Circuit erred in holding that mere knowledge of another’s direct infringement suffices to find willfulness under Section 504, so that’s an important point too. The justices seemed somewhat conflicted, on the one hand, concerned about the massive size of the district court damage award, and particularly concerned about how that could play out in university settings. For example, could a university lose internet access and their science labs be deprived of the availability of the internet because, let’s say a college student was infringing music files, you know, but on the other hand, skeptical of arguments made on behalf of Cox that there really was no circumstance where a service provider really needed to act upon knowledge. So we will see how that case comes out, but it is one to watch closely. Let’s turn now to a few cases decided in 2025. One of them a trademark case, Dewberry Group, Inc. v. Dewberry Engineers, Inc., under the Lanham Act, holding that profits may only be recovered from a name defendant, not separately incorporated entities that are not parties to the suit, and that’s an important proposition, because for those of us who practice in the IP area, we are used to drafting injunctions or final judgments that apply to the defendant and all others acting in concert or on their behalf, affiliated entities, etc, all of the language that lawyers like to draft. And what the Supreme Court is saying is you need to sue any defendant that you want to get money from. You can’t take a judgment and then apply it to these related entities. And so I think what that means is that in the trademark area, as in the tort area, presently, when a violation is alleged, it’s incumbent on the plaintiff to add defendants and make sure they are joined to the lawsuit if the plaintiff seeks to recover damage from them. So that is an important case and important proposition. The next case I want to mention from 2025, Free Speech Coalition, Inc. v. Paxton, this is also an extremely important case, even though the case narrowly dealt with the question of whether a state could impose age verification requirements for adult material but it has much, much broader implications, and we’re already seeing the outcome of this case in the tech world. Free Speech Coalition, Inc. v. Paxton involved a Texas law requiring age verification for content deemed obscene as to minors. And that’s an interesting standard. We know generally about content that is entitled to First Amendment protection, and that potentially includes a lot of pornography, which is entitled to First Amendment protection. Then there are things outside the scope of First Amendment protection, such as child porn and material that is obscene. The US Supreme Court has held that that child porn and material, it’s obscene, is not entitled to First Amendment protection. This category of material, obscene, as to minors, goes back to a case from the 1960s where the US Supreme Court held that the state of New York could require store vendors to keep certain magazines behind the counter and to require that potential purchasers provide age verification, because those magazines involve pornography that was entitled to First Amendment protection when sought out by adults, but fell into this category of material that is obscene as to minors. And so in Free Speech Coalition, the Court held that that the Texas law requiring age verification for content obscene as to minors, in other words, meaning material, constitutionally protected by the First Amendment for adults, but obscene for minors, survived intermediate scrutiny because age verification advanced an important government interest, and it only incidentally burdened speech by adults because adults do not have a First Amendment right to avoid age verification. This is a very different approach than than the approach taken by the US Supreme Court in ACLU v. Reno, a case from 1996, where the Supreme Court struck down the criminal law provisions of the Communications Decency Act or the CDA, which we’ll be talking about in a few moments in connection with section 230, because those criminal provisions, in seeking to protect children from material obscene as to minors, impinged upon the First Amendment rights of adults. Now that case also involved a law that was very poorly drafted and so there were other aspects there, but I think in broad strokes, you see a very different Supreme Court in ACLU v. Reno in 1996 at the dawn of the Internet era, very concerned about regulation and how it could impact free speech, and then, many decades later, a more conservative court acting in light of the proliferation of sexual material on the internet and the problems of child abuse and and child sexual exploitation online, finding that age verification for this category of material obscene as to minors. Was only subject to intermediate scrutiny, not strict scrutiny, meaning that it was easier to pass constitutional muster, and finding that there is no First Amendment right to age verification. The reason I said this case has broader implications than just its narrow holding is because the issue of age verification comes up, not just in terms of protecting children from material obscene as to minors, but also in an array of areas where states seek to impose upon intermediaries obligations to protect children, such as, for example, state laws requiring, excuse me, App Store providers to limit access to children and various other state laws. So in broad strokes, the Paxton case does give an impetus to state legislators who seek to impose age verification obligations on companies. The final US Supreme Court case that I want to talk about from 2025, McLaughlin Chiropractic Associates, Inc. v. McKesson Corp. This is another case where the implications, I think, are much more significant than the narrow holding. The holding was that absent express statutory preclusion of judicial review, a district court in an enforcement proceeding is not bound by an agency’s pre enforcement statutory interpretation, but must determine the meaning of the statute itself, and that the Hobbes Act does not preclude a district court from disagreeing with the FCC statutory interpretation of TCPA. This is a good ruling for companies, because previously, a number of courts had held that the Hobbes Act provided that when an agency such as the FCC had ruled on an issue, District Court was bound by that ruling and generally could not relitigate it. And that, of course, was a problem for companies. A company sued in 2026 under a particular statute seeking to argue that the statute didn’t apply to its conduct might be hemmed in by an administrative ruling from 15 years earlier that it was not involved in and didn’t even realize that it should have been involved in that ruling. So in that sense, this is a very good ruling for companies. I think it follows on the US Supreme Court’s earlier decision in 2024 in Loper Bright Enterprises v. Raimondo, which overturned Chevron and generally eliminated the requirement that in many instances, courts must defer to agency interpretations, and I think this is generally consistent with the current Supreme Court’s view that the text of a statute generally should prevail over an agency determination. Doesn’t mean agency determinations are always irrelevant. But under McLaughlin v. McKesson, courts are not required to follow them, and in fact, they are required to make their own independent determination of what the meaning of the statute is for purposes of TCPA texting, which I’ll talk about more in the other program that I have for Berkeley this year, Year in Review for Data Privacy, TCPA, and cybersecurity breach class action litigation and trends, I’ll talk about it more there. But I think what this generally means is that so much of the troll TCPA litigation that was based on broad FCC determinations, rulings or even data really go out the window. And again, I think from a company perspective, this is a very good thing. So let’s jump into some of the specific topics, starting with AI Law Year in Review. We could spend several days talking about AI law from 2025 and so let me, let me provide an overview summary, and then we’ll drill down on some of the property law issues, particularly the Copyright Fair Use decisions from 2025. But let me start with the privacy area, because when we talk about AI law and litigation and AI policy, much of that involves data privacy law, and there are some very important developments there. First of all, as you know, many states have enacted privacy codes, and many of these privacy codes have provisions that implicate automated decision making, for example, the right to opt out, the right to access and notice rights and those are complicated. When you’re dealing with data use to train algorithms. They’re also, however, beyond data privacy, are AI specific state laws. For example, there is the California transparency in frontier AI act that requires risk assessments and disclosures, public, sorry, public training, data summaries and also includes whistleblower provisions, and that law is currently in effect. The New York responsible AI safety and Education Act, which establishes safety protocols, reporting obligations and oversight structures for AI systems. There is the Texas responsible AI Governance Act, which generally addresses government use of AI and prohibited practices, and then the Utah artificial Intelligence Policy Act, which addresses disclosure of high risk AI interactions and safe harbors. But interestingly enough, there is a question about whether and to what extent these state laws can be enforced because in December of 2025, anticipating this year in review program and trying to provide more interesting issues for the Berkeley Center for Law and Technology, President Trump signed an executive order purporting to enjoin enforcement of these state laws. And that raises some interesting questions about preemption. As a general matter, an executive order of the president generally cannot preempt state laws absent some kind of enabling legislation from Congress delegating that right to the President. On the other hand, there could be an argument that could be made by the federal government that state laws are preempted based on the doctrine of field preemption, which holds that when there is a matter uniquely susceptible to national regulation, then based on the Commerce Clause, state laws or enforcement of state laws would be preempted, and certainly the existence of an executive order signed by the president could be cited in support of that, I think more generally, just as a practical matter, whether whether these laws are preempted in law or in fact, the fact that the executive order directs the Department of Justice to implement this enforcement ban means, as a practical matter, that if California or Utah or other states seek to enforce their AI laws, it is possible that the federal Government could enter the litigation on the side of the company or defendant arguing against enforcement, or it’s possible that it might independently sue to enjoin enforcement. And so these are, these are very interesting issues that you know, that we need to follow up on, and we will see and hear more about in in 2025.

 

Ian Ballon  33:46

The other issues to be aware of — SEC enforcement. There is enforcement against AI washing, and that is the practice of exaggerating AI descriptions in securities filings. There are ethical issues in connection with training algorithms and most companies in this space have their own ethics committees, you know, to ensure that the algorithms are not undertaking bad practices, for example, that you know, there was this famous MIT study a couple of years ago showing that that machine learning using decades of data from the real estate industry, which included green lining and red lining, that the machine learning improved upon historic racial discrimination, and obviously that raises significant ethical issues. We also see AI issues in your legal practice. I think most people in this presentation today are familiar with clients who use AI for legal guidance. I have clients, particularly business clients, who will send me reports suggesting what the approach should be in litigation or with respect to a particular problem. And you know, AI is, you know, it’s a good tool. What AI is very good at doing is, is providing the most common response, which, for some legal matters, the most common response is a good response, and that’s not a bad place to start. Sometimes, though, when you’re dealing with complicated, cutting edge litigation matters, you don’t want the most common response. You need to be thinking about more strategic issues. Also, of course, there is bias. I had one client send a particular recommendation, and based on the recommendation, I had a pretty good idea what the prompt was to yield that response. But it doesn’t mean that it was necessarily the right response, because it suggests that what I think would have been a very solid strategy if litigation was inevitable. In other words, if we’re going to be sued anyway, this was a very good strategy to preempt that, but it was my assessment that we were probably not going to be sued, or at least my client wasn’t for various other reasons. And so if litigation was not likely, then a different approach to get a desired outcome one that I might suggest. So, again, something, something to be aware of. I had an interesting evidentiary issue in a trial in 2025. I was representing a client in a trade secret case, and the defendant had counterclaimed, alleging trade liable based on some information posted on my client’s blog describing the misappropriation and I had moved in limini to exclude evidence and an expert report claiming huge damages tied to these alleged, allegedly defamatory statements, because we were able to show that there were only less than 20 visits to the particular blog post that supposedly engineered millions and millions of dollars in damage. And the other side came up with a very interesting argument, which was the fact that virtually no one had looked at the particular blog posts at issue was not relevant because they argued that ChatGPT picked up the information, and therefore, regardless of the metrics of the number of websites visited, what you had was information that was widely disseminated via ChatGPT and was affecting their business. So that was an interesting question. I had a good associate of mine spend the weekend running the same search, both on ChatGPT and on other large language models, as well as what I thought were the more logical queries that third parties likely would have run in looking for information on this other company to come up Monday morning with a lot of conflicting evidence showing that actually, you really could not assume that these blog posts actually were being disseminated via large language models and that this was all speculative, so it was a very interesting issue. The case ended up settling because the judge allowed in an array of text messages from the defendants showing what we contended were knowledge of the willfulness of their trade secret misappropriation. So we never got a ruling on that particular AI issue. But I think it’s a good example of how in litigation, in your practice, in courts, AI is coming up in a lot of different ways, and is on the forefront and the front of mind for many people. The other thing, which I am surprised to still be mentioning three years later, is the problem of unreviewed legal research. That’s the Mata v. Avianca case from 2023 that involved a federal doctrine and a number of case sites beautifully well written, but the doctrine and the cases didn’t exist. And it goes to the problem of AI hallucinations, and after that highly publicized case, you would have thought this would no longer be a problem. And yet, in 2025, as of 2025, there were over 150 court cases calling out, sanctioning or chastising lawyers for submitting briefs created in part with AI and referring to cases that do not exist because of AI hallucinations. So as crazy as that sounds, it’s something that I need to mention. Also with respect to AI, there are interesting issues of privilege and confidentiality, and for many companies, I find it’s helpful to go back and revisit your non disclosure agreements to the extent you’re dealing with third party companies based on uses of AI. And then, of course, there are the property issues that I’ll be talking about in greater detail use of third party content or data or information to train algorithms, and a company is always free to use their own data or public domain data or open source data, as long as it complies with the license. But much of the disputes in the area involve the use of third party data, but AI is really, really an essential part of all our practices today. Someone said that 2025 was the year AI became real in the sense that companies started not just saving money, but making money from AI. And certainly it seems like 2025 involved some paradigm shifts in the AI area. And of course, AI has gotten much, much better. For example, here is a picture of me from the holiday season. If I did not tell you this was from AI generated, you would probably have thought this was an actual picture of me on a reindeer in a small town in in Colorado. On the other hand, some AI generated output still is easy to identify. For example. Here’s a picture where you’ve got a family, and that man seems to have a tail and a hoof, and I don’t even know what’s going on with all the — the dog there with all of their legs. Here is a painting which looks actually like an interesting piece of surrealist art with a man who has many different hands. Here are some yoga poses. Some of these are fabulous, and I’m sure even the most experienced practitioners of yoga watching this program cannot achieve some of these outcomes. And of course, here I am with three sets of teeth, and that’s that’s just wrong. But oftentimes the problem is not AI, but actually the prompts that are used to generate the output. For example, here’s a cocky expression. I mean that that seems like a cocky expression. And then here are some results for the request for salmon in the river. And we’ve got some salmon cutlets in the river. Again, a more precise prompt could have yielded perhaps a different outcome. So in terms of AI output, let’s start there and just briefly talk about the state of the law in 2026 as we look at protectability, we know that the patent office will not allow you to get a patent for AI generated inventions, and we know from the copyright world that that’s also true. In 2025 the DC Circuit in Thaler v. Perlmutter held that the copyright office did not act arbitrarily or capriciously in denying registration of an autonomously generated work. And just yesterday, as we, as I sit here in March of 2026 the US supreme court denied cert in that case. Now, even before Perlmutter, we knew from the monkey selfie case that it was going to be unlikely that algorithms could generate copyrighted works. The monkey selfie case involved a precocious monkey Naruto, shown here on the screen, and he was posed for a picture by a photographer because he was precocious, he reached out and clicked the button, and so it is a monkey selfie. The Copyright Office denied the copyright application of the photographer because the picture was taken by the monkey, and the Ninth — and because this is America, someone filed suit on behalf of the monkey — and the Ninth Circuit held that animals other than humans lack statutory standing to sue under the Copyright Act. In 2025 the copyright office came out with a very important report. They actually had several reports, one of which was somewhat controversial. The one that I’m going to talk about is copyright and artificial intelligence part two, and that was a report from 2025 really summarizing the law of copyrightability of output that is AI generated, or AI enhanced. And the Copyright Office held that copyright protects the original expression in a work created by a human author, even if the work, even if the work also includes AI generated material. And that’s again, very reasonable from a copyright perspective, because we’ve always known that a work can include both protectable and unprotectible elements, and what the copyright protects is the whole work. At the same time, the copyright office said copyright does not extend to purely AI generated material or material where there’s insufficient human control over the expressive elements, and that is an important concept. It also, frankly is important from a transactional perspective, to think about that for material that your company generates, because there are instances where you can have AI perform a task for you in perhaps less than a minute, or you could have an employee spend eight hours performing the same task. And if it’s important to enforce a copyright in the output, you may well want to pay the employee inefficiently to do the task that the algorithm could do more effectively, so something to be aware of. The report also then goes on to say whether human contributions to AI generated outputs are sufficient to constitute authorship must be analyzed in a case by case basis. That also makes sense. And then finally, this, the report says, based on the functioning of current, generally available technology, prompts do not alone, provide sufficient human input to be deemed protectable, and that’s an important proposition. But I will say this. I think when you’re dealing with AI generated software, there may be an argument that the human involvement with respect to prompts, what to include and what to exclude, involves much more, much more decision making from a human perspective. And therefore there is an argument that when you’re dealing with prompts for software development, there is more of an opportunity to create AI assisted works that are themselves entitled to copyright protection. If you look at the report, copyright and artificial intelligence part two, the report primarily cites the more creative works, books, motion pictures, television shows, things of that nature. And you know, if you’re dealing with a program that is a single prompt to create a photo, that’s a little bit different, I think, than software development. So something to think about. Turning to the output and to liability. This is an area where there are a wide array of potential causes of action that could apply. This is, I realize on the screen, there is a lot of content, and you should probably consider this your 2026 eye exam, because if you can see all this, you have 2020 vision. And if you can’t, you may need corrective lenses. But the slides are available, as I mentioned at the outset, in addition to other materials, slides are available, and if for any reason you’re having difficulty accessing them, feel free to just send me an email and I will send them to you. I also have included about 150 page excerpt from my treatise on AI and property rights and the extent to which liability can arise for using third party content, data or information without permission to train algorithms. So if you are having trouble sleeping in these difficult times, feel free to print that out and read it before bed, and you will undoubtedly sleep like a baby. In any event, there are a number of different theories. I talk about them at great length in the excerpt from my treatise, and I’m not going to go into significant detail in this year in review program, except to talk about copyright, where we’ve got some new case law that I think is worth talking about. The other kinds of claims, you know, like I tell clients, when you’re looking at this area, you have claims that may be easier for a rights owner to establish, such as breach of contract, if terms of use or EULA are circumvented in order to access content, data or information, but where damages may be harder to establish, and then, on the other hand, claims where the remedies are very strong but liability may be harder, such as under the Copyright Act, where there are statutory damages that are potentially available, potentially up to $150,000 per work infringe for willful infringement, but unclear whether liability may be established. The Copyright Act, which was enacted 50 years ago. We’re now celebrating the 50th anniversary of the 1976 Copyright Act. It provides a liability may be established where you can show access and substantial similarity. There’s no particular law requiring large language models to keep detailed records of what they have ingested, to train their algorithms. But if in litigation, if it can be established that there’s access, then liability is based on access and substantial similarity, unless there’s the defense, such as fair use that may apply. But even where access cannot be established, the case case law makes clear that if you can’t show access, access can be presumed, if instead of substantial similarity, you have something closer to a striking similarity. And so you know that’s something to be aware of. The reality when you’re using third party content, data information to train algorithms is that the output often doesn’t look like the input. You have what’s called intermediate copying, where third party material may be used without permission to train the algorithms, and unless that is happening in real time, typically, there’s a copy made somewhere. It may just be a temporary copy from which various information is extracted, and that temporary copy then raises the question of intermediate copying because there is no substantial similarity between the input and the output. I mean, in those rare instances where the input and the output look the same, that’s a different question. And maybe that’s that’s a different kind of case, but in most instances, the output doesn’t look like the input. And so the question, from a liability perspective is intermediate copying, and then from a defense perspective, whether that intermediate copying is a fair use. So from the past year, there are couple of cases I want to flag on liability, and then in a few moments, I’ll talk more about the fair use cases. But on liability, there was an interesting decision in Thomson Reuters Enterprise Centre GMBH v Ross Intelligence Inc. from the district of Delaware that is the first case to, excuse me, be decided on the merits. The case went to trial, but it involved very unique facts, and the court granted plaintiff partial summary judgment on keynote copyright claims, and so it dealt with West keyword, sorry, not keyword, keynote indexing system, and that’s a little bit different than the more typical case. In Concord Music v. Anthropic, very important decision from the north District of California in March of 2025 denying a preliminary injunction to establish an injunction a plaintiff needs to show like the prevailing on the merits and irreparable injury, and when you’re dealing with the allegation material that was ingested was infringing, it is hard to make an argument that that material needs to be removed in order to maintain the status quo. When you are asking for affirmative relief in an injunction, that’s what’s called a mandatory injunction, and it’s always harder in a court of equity to get a mandatory injunction than a negative injunction. Negative injunction simply enjoining a defendant from doing something in order to maintain the status quo. And so that’s that’s essentially what you saw in the anthropic case denying a preliminary injunction. There is this question about algorithmic disgorgement. In the prior administration, the FTC had spoken about the possibly, the possible need for a remedy of algorithmic disgorgement, at least in the privacy area, where information was ingested to train large language models in violation of various privacy laws. That is, that tends to be viewed as a more extreme measure, and it is a more complicated measure to implement, but it is something to be aware of, and it is one of the reasons that when injunctive relief is sought in this in these cases, it may be harder for a plaintiff to get a preliminary injunction, or indeed possibly even an injunction. There are other kinds of claims, including anti circumvention under the digital one and Copyright Act. All of those are addressed at great length in the excerpt from my treatise. In this area, when you’re dealing with these theories of liability, both particularly in the IP area, there are issues both of direct infringement and secondary liability. Secondary liability may arise if you pay a third party to access the data or acquire the data, and this sometimes happens when you’re dealing with llms, with companies offshore, they may be representing that they’re fully complying with applicable law. And it also can apply to end users. So for example, if your company uses an off the shelf AI program, even though the AI provider may be the company directly liable, it could be you as the user that would have secondary liability. And that’s also why most of the off the shelf llms include contractual indemnity provisions, which you should look at closely to the extent you’re using these these programs. There isn’t a direct analog in the non IP area for secondary liability, but there are similar concepts like interference with contract and the like. And again, I go into this in more detail, both in the slides and the materials.

 

Ian Ballon  58:42

Let’s quickly just stop on the issue of statutory damages, because damages can be quite significant in the AI area, if a work has been timely registered before it was infringed, a copyright owner generally may recover either actual damages, which are the defendants wrongful profits, and to the extent not duplicative, the plaintiffs lost profits or statutory damages, which generally range from 750 to $30,000 per work infringed, but can go up $250,000 for willfulness. So what is a work? There is a circuit split on what a work is, and that’s something to be aware of. We see this in the music area, where an album that is registered in most circuits, each song on the album would justify a separate award, even though there’s only one registration, because each song has independent economic viability. And most courts apply that test. However, the Fifth Circuit in an opinion which I think is perhaps more consistent with the perspective of the US Supreme Court, Fifth Circuit held that what is a work is determined by the four corners of the copyright registration certificate, not by any doctrine like, like the independent economic value test, that does not appear in the plain text of the Copyright Act. But this is something to be aware of, and quite important. In the Fair Use Area, this is where we’ve seen two important AI decisions in 2025, copyright fair use, as you know, is determined by a multi part balancing test. Courts must look at the purpose and character the use, including whether it’s for commercial nature or non profit educational purpose, and particularly whether the use is transformative, whether it’s using a copyrighted work for a completely different purpose, to further the constitutional objective of copyright law, to advance the useful arts and sciences, or whether it is simply in the same, you know, competitive to the plaintiffs use or otherwise for the same purpose and character. Courts also look at the nature of the work, the amount and substantially the portion used, and the effect of the use upon the potential market for about the copyrighted work. So if you use material that is available for license for training algorithms, and you do not pay the license, you are going to have a very difficult time establishing fair use, because your infringement has an adverse impact on the market for the copyrighted work. On the other hand, if the material is not available for license, there is an interesting question. We’ve seen two Supreme Court decisions in the Fair Use Area the last couple of years. I’m not going to go into detail except to flag them. One was Google v. Oracle holding that Google’s implementation of 37 of 166 Java SC application programming interfaces was a fair use. I represented Google in developing the copyright argument in the trial court. So I of course, am completely unbiased source and can’t really talk about the case beyond saying that the Supreme Court got it right here and saying this is a fair use. But then two years later, the Warhol decision came to a different conclusion involving these works, the Andy Warhol painting and the Lynn Goldsmith photo, and in that instance, the Supreme Court was only asked to address the question of transformativeness. I do believe that if it had been looking at all four factors, it could have reached the same conclusion by saying, in this unique, limited circumstance where both the painting and the photo were considered by Vanity Fair for a royalty to be used on the cover of Vanity Fair magazine. In this unique competitive situation, the use had an adverse impact on the market for the genuine product, but because the court was only tasked with considering transformativeness, what you had instead, as a majority, really changed the standard of transformativeness, saying that simply because something is transformative doesn’t mean it’s a fair use. The question is, how transformative is it? And this is particularly relevant when the works are used in competition with one another. So one could say, to some extent, this fourth Fair Use factor was imported into the first Fair Use factor. In the AI area, in 2025, we saw two important decisions. But it’s also important to underscore that these two decisions came from the same district court, Northern District of California, and so while we have two important fair use decisions, they’re both district court decisions from the same circuit, from the same district. In Bartz v. Anthropic, Judge Alsup granted partial summary judgment for Anthropic, holding that the use of plaintiff’s works for training and print to digital format change were fair uses, but that the use of pirated library copies was not a fair use. This goes back to a line of cases that hold that, something that when, when what is copied is itself a pirated copy and not the genuine work, then the use can never be fair, whether that’s rightly decided or wrongly decided. There is a line of cases that hold that, and so judge also simply applied that. Following that decision, there was a $1.5 billion class action settlement, and Judge Alsop gave preliminary approval to that in July. Within days of the Bartz v. Anthropic decision, Judge Chhabria, also from the Northern District of California, decided Kadry v. Meta Platforms. In that case, granting summary judgment for defendants, holding that copying of plaintiff authors works to train AI models was a fair use. And so we have two decisions, recognizing the highly transformative nature of using third party material without permission to train algorithms with the caveat that Bartz v. Anthropic, Judge Alsop found that that the copying of material that itself was counterfeited did not qualify as fair use.

 

Ian Ballon  1:06:24

Let’s turn to First Amendment and the Lanham Act. We have some earlier cases and some more recent cases I want to talk about. A couple of years ago, the Supreme Court decided this case, VIP Productions, and really changed the law involving Rogers v. Grimaldi. I had a case in this area which went into litigation for multiple years. Punch Bowl the AJ press, my client, AJ press, the publisher of Punch Bowl News. Punch Bowl News is called Punch Bowl News because the Secret Service, Secret Service code name for the US Capitol is the punch bowl, because if you look at the US Capitol, it looks like an inverted punch bowl. And so that’s why the company was called Punch Bowl News. You see there in non partisan violet, together with their slogan, power, people, politics, but punch bowl. The party planning service sued us for counterfeiting. The case was filed in front of my former law partner, Leo McGrady in the Eastern District of Virginia. We moved to dismiss, arguing that this case involved forum shopping. Punch Bowl was a Delaware corporation with the principal place of business in Massachusetts, and Punch Bowl News was a California — is a California company with a principal place of business in the District of Columbia. Case was dismissed, refilled in the Ninth Circuit and ultimately decided under Rogers v. Grimaldi. But then the mandate was stayed by the Ninth Circuit, pending the VIP case, and ultimately the case was vacated. We ended up winning on different grounds, but it is illustrative of the kind of case where you’ve got First Amendment issues. And so where are we, post VIP? We have a couple of interesting cases from 2025. The Supreme Court side stepped, what’s really a circuit split under Roger V Grimaldi, where you’ve got a broad ninth circuit test. This is the broad ninth circuit test. This is from my treatise. If the mark, if a mark is used as part of an expressive work protected by the First Amendment, then you ask two questions, does the secondary use have some artistic relevance? That’s a very low standard. Makes it fairly easy for defendants where there is an expressive use to sidestep the typical polaroid or sleep craft factors applicable to likelihood confusion and go into this Roger Grimaldi test, and if the secondary use has some artistic relevance, then the question is, is the use explicitly misleading? And this question of explicitly misleading, that’s a much higher standard, much harder for a plaintiff to show that a defendant’s use is explicitly misleading, than to show likelihood of confusion which is a lower standard. So in the Ninth Circuit, it’s a very free speech friendly, very defense oriented state test. In the Second Circuit, the courts look not only at — at these First Amendment questions, but also the Polaroid factors and likelihood of confusion. And so you come up with different results, potentially the same case could be decided differently in the Ninth Circuit than the Second Circuit, for example, but in VIP, the Supreme Court sidestep this circuit split again, leaving it for future case law to decide whether Rogers V Grimaldi even applies. It just held that Roger V Grimaldi was not appropriate where the accused infringer has used the trademark to designate the source of its own goods. And so that’s an interesting question, we’ve got a couple of interesting 2025 decisions from the Ninth Circuit. We have Yuga Labs, Inc. v. Ripps, which was an NFT case, holding that there’s no First Amendment defense where defendants use of plaintiff’s NFT mark was used as a source identifier, and that really is pretty faithful to the VIP case. On the other hand, in Hara v. Netflix, we have a ruling from the ninth circuit that Netflix ‘fleeting use’ of an animated version of a drag queen performer had artistic relevance to Netflix Q-force streaming show, and therefore was not actionable under Rogers v. Grimaldi. And I think that really shows you the two ends of the spectrum that you will see post VIP — between cases where the defendants are using the mark in some source identifying way, Roger Grimaldi and the First Amendment do not apply, or whether there is a First Amendment use, in which case the test may be applicable. And that leads us to NFTs. And as I mentioned, we have the First Circuit Court opinion on NFTs, the Ninth Circuit’s decision in Yuga Labs v. Ripps. And that was a significant decision, holding that an NFT, an intangible asset,  constitutes goods under the Lanham Act, and that is an important decision. I mentioned the different approach in different circuit under Rogers V Grimaldi. We see that in the Hermes v. Rothschild case, for example, it involved Hermes bags, which are shown here on the screen. And you know, I don’t pretend to be a connoisseur of ladies handbags, but by my estimation, these do not look like Hermes bags that you might buy, sorry, Birkin bags that you might buy in a store. These Birkin bags are NFT works of art by Rothschild, but under the second circuit test, entirely enjoined, enjoined, even to the extent that that they could not be displayed in an art exhibit in another country. So very different approach. Let’s move to Secondary IP liability and the CDA. And there’s a fair amount that one could talk about. I’m going to try to cover it as quickly as I can. Significantly and most importantly, in the copyright and DMCA area, we had yet another ruling in Capitol Records, LLC v. Vimeo, and in this case, the Second Circuit affirmed judgment for the service provider where there’s no red flag awareness. This gets back to the important question of whether a service provider can lose the MCA protection where it has more generalized knowledge. Excuse me, in the Second Circuit said no, it has to be knowledge of specific files or specific infringing activity. I mentioned already the Cox case before the US Supreme Court, not on DMCA, but on secondary copyright liability, and that’s going to likely lead to an important decision at some point between now and the end of June this year. In the trademark litigation area, of course, there’s no Digital Millennium trademark act, but there’s also a lot of litigation, including claims, typically by plaintiffs alleging counterfeiting and other claims. But let’s turn to the CDA, because I think there’s some interesting things to talk about in terms of trends. My friend Eric Goldman talks about how the CDA is under attack, and that’s certainly true. You have Supreme Court Justices, who are Justice Thomas Justice Alito and Justice Gorsuch, in an earlier ruling in the Paxton case, expressed some skepticism about the applicability of the CDA, or actually, I’m sorry, not the Paxton case, an earlier case that was not decided on the CDA grounds, but could have raised CDA protections. And you do have criticism from President Trump and his supporters of the CDA, even though President Trump personally, has prevailed in a number of cases by asserting the CDA, and then you also have, on the other hand, Democrat senators like Klobuchar and others expressing criticism of the CDA because of the way it limits the ability of plaintiffs in litigation to bring tort actions. And so what we see is really from both the left and the right. We see this kind of pressure, intention, hostility, and those kind of atmospherics can be very important in case law, in influencing judges, particularly judges who may not be steeped in CDA case law. The CDA, at least the civil law provision, section 230, provides that 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. And what that means is no liability for republication. You know, subject to certain exclusions, the exclusions include fosta sesta, dealing with child trafficking, including some provisions dealing with advertising. So most tech companies typically aren’t implicated by this, but to the extent that you take advertising that that could involve certain kinds of adult content, there could be issues. There also excludes federal criminal claims, claims under ECPA, any similar state law, and any law pertaining intellectual property. I’m not going to get into detail in that, because the decisions are not from 2025. We talked about this in past years, but it is addressed extensively in my materials, and of course, on the board here are the major cases. There is a circuit split on what constitutes any law pertaining to intellectual property. The Ninth Circuit construes that narrowly as federal IP, meaning that state IP claims, right of publicity, state trade secret misappropriation, state trademark claims, are going to be preempted by the CDA in the ninth circuit. The Third Circuit construes any claim pertaining intellectual property broadly as any state or federal claim, and therefore there is no CDA defense to a state right of publicity claim, state trade secret misappropriation or state trademark claim in the Third Circuit. So that that circuit split is important particularly, and in the trade secret area, it’s important to know also the Defend Trade Secret Act. The statute doesn’t make this clear, but the, sorry, not the not the codification, the codification of the US Code, doesn’t make this clear, but the statute actually does make clear that the Defend Trade Secrets law is not a law pertaining intellectual property. That sounds oxymoronic for anyone who has gone to law school, school and taken a class on IP, because, of course, a trade secret claim is quintessentially a law pertaining intellectual property. But in fact, what that means is that the CDA applies. So you have really the the odd situation in the Third Circuit where, if a service provider is sued under the federal Defend Trade Secrets Act, and under New Jersey, or Pennsylvania trade secret law that the CDA could provide insulation to the federal Defend Trade Secret Act, but not the state law claim. So something to be aware of, but I don’t want to drill down too deeply on that. I want to focus on where we are in 2025 and these are the major cases, and I think they’re illustrative of the particular trends. And so I want to flag these importantly, because we’re running low on time. And then then get into a couple of other quick points before I finish this, this presentation. So we saw in 2024 the decision from the Third Circuit Anderson v. TikTok. This is a challenging decision. I think it’s wrongly decided. Holding the TikToks algorithm, which decided whether third party speech on a platform would be included or excluded from a compilation and organized and presented the included items on users unique, personalized feeds of content was Tiktok’s own expressive activity, and thus outside the scope of the CDA, because it was not protected by the CDA, the Third Circuit allowed product liability, negligence, wrongful death claims. I think that is absolutely wrong when you’re dealing with — just because something is AI generated or algorithmically generated, it’s still third party content. And in 2025 the Fourth Circuit ruled that way in the case of M.P. by and through Pinckney v. Meta Platforms, that’s a Fourth Circuit 2025 decision holding that the CDA preempted claims for negligence, strict product liability and negligent infliction of emotional distress for improper content that metas use of an algorithm to deliver content was integral to Its publishing function, and did not take it outside the CDA, so we’ve now got a pretty prominent circuit split. I’m not sure too many companies are going to be anxious or excited to take it to the US Supreme Court, because even though it seems from this earlier case that Chief Justice Roberts is very concerned about the fact that you have three decades of case law construing the CDA by lower circuit courts relatively consistently. You do have three justices, Alito Gorsuch and Thomas, who are quite hostile to the CDA. And so that’s a that’s a hard, hard mathematical number to go up against when you’re already starting with three against you. In any case, it’s an interesting circuit split and one to be aware of. It certainly has implications for AI. I do think the Fourth Circuit decision is right when, when you’re dealing with third party content and the content is repackaged or republished, CDA should absolutely protect it, of course, if you’re dealing with hallucinations, that’s different and that may be outside the CDA. There are certainly arguments that can be made, but it may be outside the CDA. In any case, other 2025 decisions — Doe v. Grindr from the Ninth Circuit holding preempted the claims of underage users for defective design, defective manufacture, negligence, failure to warn And negligent misrepresentation. That’s an important decision, very consistent with existing case line most of the circuits, but I will say it also reflects the trend. Plaintiffs in this area increasingly are suing for claims based on product liability. They’re arguing that their cases are not about republication, but about product liability the way that a defendant or service providers service operates, and that really liability needs to be premised on product liability. So that’s certainly a trend to be aware of. Other cases, Doe v Twitter, another Ninth Circuit case from 2025, finding immunity on claims that Twitter knowingly benefited from sex trafficking and on their product defect claim based on failure to move posts, but not for claims of negligence per se and product liability based on a defective reporting infrastructure design rather than publication. Again, this is very consistent with that trend where I’m saying that that plaintiffs increasingly are trying to frame their republication claims in terms of product defect, that this is not about content, even though the case is about content, it is about the — it is, about, excuse me — the product itself. And a couple of other, well, one other case to mention, the US v. EZ Lynk, second circuit CDA inapplicable based on defendants alleged developments. Couple of other quick points about the CDA. There continues to be significant litigation about CDA marketplaces. I represent eBay in litigation. EBay has been successful in a number of cases either based on the CDA, such as in the Inman case that we won more than a decade ago. But also marketplaces, including Walmart, eBay, Amazon and others, even where the CDA is inapplicable, generally, have been successful in holding that state product liability laws do not apply to online marketplaces with a couple of exceptions, particularly in California, older cases, intermediate appellate decisions, Loomis and Bolger, but to a large extent across the country, holding that that these product liability cases do not apply to marketplaces. Again, consistent with this trend of plaintiffs trying to hold intermediaries liable for product liability. So some takeaways in the CDA area, content versus conduct. In my treatise, I argue, and for years, I basically said that conduct when you’re dealing with the internet or mobile phones takes the form of content, and therefore what you’re dealing with is publication. But as I’ve mentioned, there are now some courts that are distinguishing content cases, finding that actually it is the conduct of the defendant and plaintiff’s counsel try to frame claims in terms of development of content and conduct by the interactive computer service itself. We see some different approaches in different circuits. The Ninth Circuit has a number of very narrow, limited exceptions, but this tends to give life to plaintiffs in this area. Product liability law and the First Amendment may help where the CDA is unavailable, but I do think we’re seeing trends where plaintiffs are trying to hold algorithms — hold companies liable by treating algorithms as content rather than as publishing tools and publishing decisions. Publication decisions inherently are protected by the First Amendment, and that’s core to the CDA and CDA case law, going back to the Zaran case more than three decades ago. Now product liability minors. We do see a lot of litigation involving minors, sympathetic plaintiffs. There is some general hostility to the CDA, and as I mentioned at the outset of this section, that impacts the atmospherics, where some judges are more apt to find liability. And then there’s also the possibility of regulation in this area. Project 2025 takes the position that the CDA, because it is codified in the Communications Act Title 47, it is subject to regulation. And in the very last year of the First Trump administration, in 2020 there were proposed regulations so that could come back. It doesn’t seem to be a high priority at this moment, but that could happen. Let me stop here.

 

Ian Ballon  1:27:56

I mentioned that I do have another Year in Review program that you can find on the Berkeley Center of Law and Technology, Year in Review for Data Privacy, Cybersecurity Breach and Ad Tech Trends and Case Law. Let me just tell you very briefly what I cover in that area. I’ve included some slides here so you do have these materials, but I would encourage you to go to that other program, because I’ll be talking about the trends and the major legal developments, the claims that are brought, under the video Privacy Protection Act there is a circuit split, and the US Supreme Court in 2026 has granted cert to a case. So we will see another vppa decision from the Supreme Court, probably in late 2026 or early 2027. We’ll talk about the interplay between mass litigation and mass arbitration. This is something I deal with a lot in my practice, in dealing with multi case litigation, sometimes you want to go to the multi district litigation panel. Sometimes you’re dealing with mass arbitration. A positive development in 2025 is that the non refundable fees from triple A jams and other major providers have been greatly reduced and some mass arbitration does not present the kind of problems for companies in 2026 that it did in earlier years. You’re not likely to look at hundreds of millions of dollars in in non refundable fees. So that also has an has implications for strategy, and I’ll be talking about that in this other program. I’ll also be talking about CCPA and CPRA, and why the CCPA, even with statutory damages of up to $750 per claimant, is not a law that gives me a lot of heartache at night. I think it’s one that has a lot of difficulties for plaintiffs, and so we’ll be talking about that. I do think the Illinois Biometric Information Privacy Act is one that’s a concern, even though there are ways to defend those claims. We’ll talk about straight privacy policy litigation, and how I think there’s a nuance difference between the circuits on that, and I mentioned CCPA mitigation. And then finally, in this other program, I’m going to be talking in more detail, but online and mobile contract formation, I will say, for those of you who can’t join that program and just want the Reader’s Digest version. There continues to be a difference in approach between the Ninth Circuit and other circuits, where it is substantially harder to enforce an online and mobile contract or arbitration clause in the Ninth Circuit than in any other circuit the United States. And as I talk about in this other Berkeley program, I’ll give you some screenshots where you’ll see a case in the ninth circuit, where in other circuits there should have been a contract, and a screenshot of case from the Fourth Circuit from 2025 that I think would have been difficult for the company to establish contract formation in the ninth circuit. This obviously implicates transactional lawyers, and it implicates compliance, and it is ultimately a problem that that you have one circuit where it is so much harder to form a contract than in other circuits. So we’ll talk about that. I do have the slides here for those of you who can’t make that other presentation. We’ll also talk about the latest mass arbitration case law from 2025. But with that, I will leave you with the AI generated version of Ian Ballon. This is from Meta’s AI. I’m never going to look that good. So that is my Facebook profile picture, and it’s a reminder that AI allows us to be better than we really are. With that, I’ll turn it back to Wayne and thank you for joining me on this program. You can reach out to me by email or other modern means of communication. If you have questions, you can get these materials. And of course, I encourage you to also listen to the other program available through Berkeley Law —  Berkeley Center for Law and Technology on data privacy, AI cyber security breach, class action litigation and contract formation.

 

Wayne Stacy  1:32:43

Wayne in again. Thank you for this. And I want to remind everybody, a lot of programs out there, you don’t need to go look at the slides. You don’t need to look at the extra material, but there are two that I always say, go look. One is Mark Lemley’s Year in Review, and the other is Ian Ballon’s Year in Review, because the work is really there. You can have everything you need for yourself. For those that are working with you, pass it on. It’s very meaningful content. Ian, fantastic, and I look forward to working with you in the future.

 

Ian Ballon  1:33:15

Thanks, Wayne, I appreciate it, and thank you for your very kind compliment lumping me together with Mark Lemley, who is one of the great lawyers and great scholars in this area. So thank you for the compliment.