IPSC Opening Plenary Session

 Matthew Sag, Copyright Safety for Generative AI

Not addressing whether training is always fair use in every circumstance; explain how generative AI fits w/in existing law (nonexpressive uses) and identify best practices to make generative AI fairer.

Non expressive uses: reverse engineering, anti-plagiarism, search plus text data mining, snippets. Although Authors Guild didn’t discuss it, that was very much about machine learning—the endgame was more sophisticated analysis using machine learning, not just frequency tables. His theory: rights of © are defined w/r/t communication of original expression to the public. Nonexpressive uses don’t involve human enjoyment of the original expression contained in a work and don’t lead to other humans potentially enjoying that original expression. Thus, nonexpressive uses don’t interfere w/the kind of interest © is designed to protect, which is why they are such strong candidates for fair use.

But is generative AI really the same? Are the outputs copies of the training data? Copying takes place prior to training; converted into tokens and training is a process of adjusting weights in the model, not copying tokens. Info acquired is much more abstract than the training data. The output is a novel synthesis, not just a cut and paste, usually combines abstract latent features learned from training data.

But CS literature on extraction attacks exists: in successful attacks, models may be trained on many duplicates of the same work; images associated with unique text descriptions; ratio of model size to training data is relatively large. Generally the extraction attacks fail; only succeed in edge cases. These are real risks, but rare. More common: The Snoopy problem: text to image models are set up to learn attributes of ©able characters: b/c same text description paired w/relatively simple images that vary only slightly. Not limited to characters; other images repeated w/ minor variations and consistently tagged w/same keywords: Banksy’s girl holding a red balloon.

Implications: too much memorization undermines arguments in favor of fair use. Hard to argue it’s highly transformative if training data come out as output. May also have implications under fourth factor. What to do? Ten suggestions in paper, but key: (1) reasonable measures to deduplicate; (2) reasonable measures in training and deployment to reduce probability of infringement—curate/pre-processing training data; using reinforcement learning through human feedback; installing restrictions on model outputs. (3) reasonable measures to safeguard privacy interests. (4) detailed records of works used and their provenance.

Sonia Katyal, A Trademark Theory of Rebranding

Lots of rebranding out there. Cultural and economic factors (change in what the business does over time, including startup drift and embrace then rejection of crypto). Practices are undertheorized and underregulated. Not transparent/information asymmetry for consumers. Internal contradiction in TM: rebranding could be characterized as contradiction b/t law and marketplace: marketplace gives incentives to rebrand, given lack of legal regulation thereof. A single user can obfuscate its past identity from its present. Source concealment/distortion.

Most common reason: change of ownership or structure, merger, acquisition, sponsorship change, shift from private to public, demerger

Change in corporate strategy—globalization or localization, diversify or divest. Change in competitive position or external environment (attracts the most commentary). TM doctrine assumes marks are chosen for potential for perpetuity. Allows tacking of priority w/continuing commercial impression. Material alteration requires new filing. This rule facilitates information asymmetry that can harm consumers by adding information costs. Distortion: expansion of boundaries by continuing commercial impression (actual TM boundaries). Also identity concealment—Worldcom to MCI; organizational misdescription—renaming, reorganization, shell companies; source obfuscation, where FB rebrands to Meta as distraction; affiliative disinformation—broader rebranding associated w social movements.

TM law should facilitate disclosure or it will fall behind market.

Sheff: What about licensing/merchandising in this system? That deprives consumers of info in the way that you’re talking about; the law used to treat it as a kind of deception but now we all love it.

Sag: Why isn’t this like bankruptcy? Fresh start.

A: good for companies, not so good for consumers.

Alex Roberts, Multi-Level Lies: Distinctive feature is that laypeople are making advertising claims to friends, social media followers, acquaintances, etc. Sellers make money by selling product but also by recruiting others. Exploitative: 99% make no profit/lose money; stuck w/unwanted inventory; exploitative based on false promises of wealth and miracle products—the seller gets free advertising. Not surprising that sellers make false claims—they’re desperate to sell. Lack knowledge/training; rules aren’t intuitive, including risks of sharing true but unrepresentative experience. Nondisparagement clauses, some of which purport to bind family members. Indemnification clauses (likely uncollectable, but still scary). Many contracts—77%--say sellers can only use marketing claims or materials provided by company; 71% say other claims must be pre-approved (that seems contradictory). 57% say sellers could be liable for false etc. claims, but only 24% reference or explain specific rules.

Many actual sellers (convenience sample) make effectiveness claims—80% almost always/half the time; 85% make testimonial claims; 73% make data-based claims (lose 1.8 pounds/week); 86% make positive subjective claims; 78% make business opportunity claims. 59% were aware that laws/rules constrain the claims they make, mostly made aware by training or other reps from company. Did they influence how you advertise? Many people said yes, but many people skipped the respondents—47% said they almost always follow the rules. Many expressed uncertainty that the rules applied to them. 46% said desire to sell outweighed desire to comply; didn’t expect enforcement. 37% said the company didn’t say anything about restrictions on claims or said that there were no restrictions.

Takeaway: companies, not lay sellers, must be held liable for false, misleading, or noncompliant claims disseminated by sellers. FTC and FDA think this already!

More enforcement; more training for MLMs with interactive onboarding/continuing education. Expand social media monitoring.

Contract issues: limiting arbitration and class action waivers would help; classify sellers as employees? Platforms should enforce own TOS against these claims.

Katyal: are the influencer rules different/should this be folded in?

A: biggest difference is that influencers get paid by the seller and MLM sellers don’t. Affiliate marketers get a cut of sales; that’s similar to how MLM works, so it’s pretty easy to see how they fit in, but you don’t really need that piece because FTC has said elsewhere that the companies are liable for claims made by sellers—the case law doesn’t require the endorsement piece to hold the company liable for the claims.

Q: endgame?

A: even with the ones that aren’t snake oil, and many are, there are risks of false income claims that should be guarded against. FTC could go harder against supplement market in general; FTC/FDA should team up more often.

Christopher S. Yoo, Common Carriage and Social Media

Hating on social media is bipartisan. Turning to common carriage analysis. Declare as matter of statutory announcement that they’re common carriers and hope that makes the regulation constitutional, as in Fl and Tex. Can that be right? No. Turner Broad. Sys. v. FCC says that the Court has to assess those judgments. Thomas once said that labeling a regulation a common carrier scheme “has no real First Amendment consequences.”

What about common law? Rationales are almost all debunked but one works.

“Affected with a public interest.” That’s not enough. It’s an empty category. Newspapers were critical for disseminating information, but that doesn’t deprive them of 1A protection for editorial judgments.

“Monopoly power.” That’s historically not been true. The root is not in 17th century English common law, but 20th century laws creating public utilities. Even if that were the case, should it make a difference to the 1A analysis? Not in Tornillo, even though most cities were one-newspaper towns at the time.

“Transportation/communication.” This is a useless category. Lumping things together doesn’t teach anything (Oliver Wendell Holmes). Too general to be helpful. There are tons of communications media that aren’t regarded as common carriers.

“Quid pro quo.” Section 230 as a quid pro quo. But not clear that common carriers are the beneficiaries of 230. Another statute enacted at the same time says “nothing in this section shall be construed to create interactive computer services as common carriers or telecom carriers.” There’s clearly some immunity for entities that aren’t common carriers.

“Holding oneself out as serving the entire public.” It’s an antifraud test, but easily evaded. DC Circuit says that FB, Google, etc. don’t hold themselves out as affording neutral, indiscriminate access and that net neutrality didn’t apply to any services involving editorial judgment.

The real motivation: syllogism, not 1A analysis.

Major premise: old regimes like common carriage must be constitutional.

Minor: new regulations of social media are like common carriage.

Conclusion: new regulations ok.

Major premise is problematic; SCt has never resolved. Precedents uphold access regulation only for media in which one actor exercises physical, not economic, control that social media lack. Invoking common carriage accomplishes nothing/serves as a distraction.

Lemley: What do you do with telephones? We clearly apply that to telephones.

A: dial-a-porn cases can help answer. Both sets of cases say that even a common carrier offering services, you can also offer video over which you exercise complete discretion. You can be a common carrier for only part of your services.

Lemley: but only b/c we put telephony and cable in different buckets. Why not exercise discretion over how and under what circumstances we deliver telephone calls?

A: is the logical implication that there are limits on that? Yes.

Interestingly, neither Fl nor TX bars discrimination in the ordinary sense—deplatforming and shadowbanning are not the same things. Not a true common carrier regulation.

Chris Newman: nondiscrimination in common carriage is about passage, not behavior on the vehicle/harassing other passengers.

A: agreed. The author of the 5th Circuit opinion said we reject constitutional challenges to nondiscrimination laws, but that’s about economic discrimination, not speech discrimination. That’s handwaving.

Nicholson Price, Empirical Studies of Medical AI Patents (w/ Mateo Aboy & Seth Raker)

Applications are up even after Bilski, etc. Grant rates dipped but then went back up. 3 of top 5 patent applicants are the 3 biggest MRI makers. Disclosure is only ok; universities do better. Predictions to contrary underestimated the cleverness of patent lawyers. We’ll get lots of patents anyway; people figure out ways over around and through doctrinal barriers.