Spring is here, and hope springs eternal in the human breast that competition enforcers will focus on welfare-enhancing initiatives, rather than on welfare-reducing interventionism that fails the consumer welfare standard.
Fortuitously, on March 27, the Federal Trade Commission (FTC) and U.S. Justice Department (DOJ) are hosting an international antitrust-enforcement summit, featuring senior state and foreign antitrust officials (see here). According to an FTC press release, “FTC Chair Lina M. Khan and DOJ Assistant Attorney General Jonathan Kanter, as well as senior staff from both agencies, will facilitate discussions on complex challenges in merger and unilateral conduct enforcement in digital and transitional markets.”
I suggest that the FTC and DOJ shelve that topic, which is the focus of endless white papers and regular enforcement-oriented conversations among competition-agency staffers from around the world. What is there for officials to learn? (Perhaps they could discuss the value of curbing “novel” digital-market interventions that undermine economic efficiency and innovation, but I doubt that this important topic would appear on the agenda.)
Rather than tread familiar enforcement ground (albeit armed with novel legal theories that are known to their peers), the FTC and DOJ instead should lead an international dialogue on applying agency resources to strengthen competition advocacy and to combat anticompetitive market distortions. Such initiatives, which involve challenging government-generated impediments to competition, would efficiently and effectively promote the Biden administration’s “whole of government” approach to competition policy.
[C]ompetition may be lessened significantly by various public policies and institutional arrangements as well [as by private restraints]. Indeed, private restrictive business practices are often facilitated by various government interventions in the marketplace. Thus, the mandate of the competition office extends beyond merely enforcing the competition law. It must also participate more broadly in the formulation of its country’s economic policies, which may adversely affect competitive market structure, business conduct, and economic performance. It must assume the role of competition advocate, acting proactively to bring about government policies that lower barriers to entry, promote deregulation and trade liberalization, and otherwise minimize unnecessary government intervention in the marketplace.
The FTC and DOJ have a proud history of competition-advocacy initiatives. In an article exploring the nature and history of FTC advocacy efforts, FTC scholars James Cooper, Paul Pautler, & Todd Zywicki explained:
Competition advocacy, broadly, is the use of FTC expertise in competition, economics, and consumer protection to persuade governmental actors at all levels of the political system and in all branches of government to design policies that further competition and consumer choice. Competition advocacy often takes the form of letters from the FTC staff or the full Commission to an interested regulator, but also consists of formal comments and amicus curiae briefs.
Cooper, Pautler, & Zywicki also provided guidance—derived from an evaluation of FTC public-interest interventions—on how advocacy initiatives can be designed to maximize their effectiveness.
During the Trump administration, the FTC’s Economic Liberty Task Force shone its advocacy spotlight on excessive state occupational-licensing restrictions that create unwarranted entry barriers and distort competition in many lines of work. (The Obama administration in 2016 issued a report on harms to workers that stem from excessive occupational licensing, but it did not accord substantial resources to advocacy efforts in this area.)
ACMDs refer to government-imposed restrictions on competition. These distortions may take the form of distortions of international competition (trade distortions), distortions of domestic competition, or distortions of property-rights protection (that with which firms compete). Distortions across any of these pillars could have a negative effect on economic growth. (See here.)
Because they enjoy state-backed power and the force of law, ACMDs cannot readily be dislodged by market forces over time, unlike purely private restrictions. What’s worse, given the role that governments play in facilitating them, ACMDs often fall outside the jurisdictional reach of both international trade laws and domestic competition laws.
The OECD’s Competition Assessment Toolkit sets forth four categories of regulatory restrictions that distort competition. Those are provisions that:
limit the number or range of providers;
limit the ability of suppliers to compete;
reduce the incentive of suppliers to compete; and that
limit the choices and information available to consumers.
When those categories explicitly or implicitly favor domestic enterprises over foreign enterprises, they may substantially distort international trade and investment decisions, to the detriment of economic efficiency and consumer welfare in multiple jurisdictions.
Given the non-negligible extraterritorial impact of many ACMDs, directing the attention of foreign competition agencies to the ACMD problem would be a particularly efficient use of time at gatherings of peer competition agencies from around the world. Peer competition agencies could discuss strategies to convince their governments to phase out or limit the scope of ACMDs.
The collective action problem that may prevent any one jurisdiction from acting unilaterally to begin dismantling its ACMDs might be addressed through international trade negotiations (perhaps, initially, plurilateral negotiations) aimed at creating ACMD remedies in trade treaties. (Shanker Singham has written about crafting trade remedies to deal with ACMDs—see here, for example.) Thus, strategies whereby national competition agencies could “pull in” their fellow national trade agencies to combat ACMDs merit exploration. Why not start the ball rolling at next week’s international antitrust-enforcement summit? (Hint, why not pull in a bunch of DOJ and FTC economists, who may feel underappreciated and underutilized at this time, to help out?)
Conclusion
If the Biden administration truly wants to strengthen the U.S. economy by bolstering competitive forces, the best way to do that would be to reallocate a substantial share of antitrust-enforcement resources to competition-advocacy efforts and the dismantling of ACMDs.
In order to have maximum impact, such efforts should be backed by a revised “whole of government” initiative – perhaps embodied in a new executive order. That new order should urge federal agencies (including the “independent” agencies that exercise executive functions) to cooperate with the DOJ and FTC in rooting out and repealing anticompetitive regulations (including ACMDs that undermine competition by distorting trade flows).
The DOJ and FTC should also be encouraged by the executive order to step up their advocacy efforts at the state level. The Office of Management and Budget (OMB) could be pulled in to help identify ACMDs, and the U.S. Trade Representative’s Office (USTR), with DOJ and FTC economic assistance, could start devising an anti-ACMD negotiating strategy.
In addition, the FTC and DOJ should directly urge foreign competition agencies to engage in relatively more competition advocacy. The U.S. agencies should simultaneously push to make competition-advocacy promotion a much higher International Competition Network priority (see here for the ICN Advocacy Working Group’s 2022-2025 Work Plan). The FTC and DOJ could simultaneously encourage their competition-agency peers to work with their fellow trade agencies (USTR’s peer bureaucracies) to devise anti-ACMD negotiating strategies.
These suggestions may not quite be ripe for meetings to be held in a few days. But if the administration truly believes in an all-of-government approach to competition, and is truly committed to multilateralism, these recommendations should be right up its alley. There will be plenty of bilateral and plurilateral trade and competition-agency meetings (not to mention the World Bank, OECD, and other multilateral gatherings) in the next year or so at which these sensible, welfare-enhancing suggestions could be advanced. After all, “hope springs eternal in the human breast.”
The 117th Congress closed out without a floor vote on either of the major pieces of antitrust legislation introduced in both chambers: the American Innovation and Choice Online Act (AICOA) and the Open Apps Market Act (OAMA). But it was evident at yesterday’s hearing of the Senate Judiciary Committee’s antitrust subcommittee that at least some advocates—both in academia and among the committee leadership—hope to raise those bills from the dead.
Of the committee’s five carefully chosen witnesses, only New York University School of Law’s Daniel Francis appeared to appreciate the competitive risks posed by AICOA and OAMA—noting, among other things, that the bills’ failure to distinguish between harm to competition and harm to certain competitors was a critical defect.
Yale School of Management’s Fiona Scott Morton acknowledged that ideal antitrust reforms were not on the table, and appeared open to amendments. But she also suggested that current antitrust standards were deficient and, without much explanation or attention to the bills’ particulars, that AICOA and OAMA were both steps in the right direction.
Subcommittee Chair Amy Klobuchar (D-Minn.), who sponsored AICOA in the last Congress, seems keen to reintroduce it without modification. In her introductory remarks, she lamented the power, wealth (if that’s different), and influence of Big Tech in helping to sink her bill last year.
Apparently, firms targeted by anticompetitive legislation would rather they weren’t. Folks outside the Beltway should sit down for this: it seems those firms hire people to help them explain, to Congress and the public, both the fact that they don’t like the bills and why. The people they hire are called “lobbyists.” It appears that, sometimes, that strategy works or is at least an input into a process that sometimes ends, more or less, as they prefer. Dirty pool, indeed.
There are, of course, other reasons why AICOA and OAMA might have stalled. Had they been enacted, it’s very likely that they would have chilled innovation, harmed consumers, and provided a level of regulatory discretion that would have been very hard, if not impossible, to dial back. If reintroduced and enacted, the bills would be more likely to “rein in” competition and innovation in the American digital sector and, specifically, targeted tech firms’ ability to deliver innovative products and services to tens of millions of (hitherto very satisfied) consumers.
Our colleagues at the International Center for Law & Economics (ICLE) and its affiliated scholars, among others, have explained why. For a selected bit of self-plagiarism, AICOA and OAMA received considerable attention in our symposium on Antitrust’s Uncertain Future; ICLE’s Dirk Auer had a Truth on the Market post on AICOA; and Lazar Radic wrote a piece on OAMA that’s currently up for a Concurrences award.
To revisit just a few critical points:
AICOA and OAMA both suppose that “self-preferencing” is generally harmful. Not so. A firm might invest in developing a successful platform and ecosystem because it expects to recoup some of that investment through, among other means, preferred treatment for some of its own products. Exercising a measure of control over downstream or adjacent products might drive the platform’s development in the first place (see here and here for some potential advantages). To cite just a few examples from the empirical literature, Li and Agarwal (2017) find that Facebook’s integration of Instagram led to a significant increase in user demand, not just for Instagram, but for the entire category of photography apps; Foerderer, et al. (2018) find that Google’s 2015 entry into the market for photography apps on Android created additional user attention and demand for such apps generally; and Cennamo, et al. (2018) find that video games offered by console firms often become blockbusters and expanded the consoles’ installed base. As a result, they increase the potential for independent game developers, even in the face of competition from first-party games.
AICOA and OAMA, in somewhat different ways, favor open systems, interoperability, and/or data portability. All of these have potential advantages but, equally, potential costs or disadvantages. Whether any is procompetitive or anticompetitive depends on particular facts and circumstances. In the abstract, each represents a business model that might well be procompetitive or benign, and that consumers might well favor or disfavor. For example, interoperability has potential benefits and costs, and, as Sam Bowman has observed, those costs sometimes exceed the benefits. For instance, interoperability can be exceedingly costly to implement or maintain, and it can generate vulnerabilities that challenge or undermine data security. Data portability can be handy, but it can also harm the interests of third parties—say, friends willing to be named, or depicted in certain photos on a certain platform, but not just anywhere. And while recent commentary suggests that the absence of “open” systems signals a competition problem, it’s hard to understand why. There are many reasons that consumers might prefer “closed” systems, even when they have to pay a premium for them.
AICOA and OAMA both embody dubious assumptions. For example, underlying AICOA is a supposition that vertical integration is generally (or at least typically) harmful. Critics of established antitrust law can point to a few recent studies that cast doubt on the ubiquity of benefits from vertical integration. And it is, in fact, possible for vertical mergers or other vertical conduct to harm competition. But that possibility, and the findings of these few studies, are routinely overstated. The weight of the empirical evidence shows that vertical integration tends to be competitively benign. For example, widely acclaimed meta-analysis by economists Francine Lafontaine (former director of the Federal Trade Commission’s Bureau of Economics under President Barack Obama) and Margaret Slade led them to conclude:
“[U]nder most circumstances, profit-maximizing vertical integration decisions are efficient, not just from the firms’ but also from the consumers’ points of view. Although there are isolated studies that contradict this claim, the vast majority support it. . . . We therefore conclude that, faced with a vertical arrangement, the burden of evidence should be placed on competition authorities to demonstrate that that arrangement is harmful before the practice is attacked.”
Network effects and data advantages are not insurmountable, nor even necessarily harmful. Advantages of scope and scale for data sets vary according to the data at issue; the context and analytic sophistication of those with access to the data and application; and are subject to diminishing returns, in any case. Simple measures of market share or other numerical thresholds may signal very little of competitive import. See, e.g., this on the contestable platform paradox; Carl Shapiro on the putative decline of competition and irrelevance of certain metrics; and, more generally, antitrust’s well-grounded and wholesale repudiation of the Structure-Conduct-Performance paradigm.
These points are not new. As we note above, they’ve been made more carefully, and in more detail, before. What’s new is that the failure of AICOA and OAMA to reach floor votes in the last Congress leaves their sponsors, and many of their advocates, unchastened.
Conclusion
At yesterday’s hearing, Sen. Klobuchar noted that nations around the world are adopting regulatory frameworks aimed at “reining in” American digital platforms. True enough, but that’s exactly what AICOA and OAMA promise; they will not foster competition or competitiveness.
Novel industries may pose novel challenges, not least to antitrust. But it does not follow that the EU’s Digital Markets Act (DMA), proposed policies in Australia and the United Kingdom, or AICOA and OAMA represent beneficial, much less optimal, policy reforms. As Francis noted, the central commitments of OAMA and AICOA, like the DMA and other proposals, aim to help certain firms at the expense of other firms and consumers. This is not procompetitive reform; it is rent-seeking by less-successful competitors.
AICOA and OAMA were laid to rest with the 117th Congress. They should be left to rest in peace.
The Senate Judiciary Committee’s Subcommittee on Privacy, Technology, and the Law will host a hearing this afternoon on Gonzalez v. Google, one of two terrorism-related cases currently before the U.S. Supreme Court that implicate Section 230 of the Communications Decency Act of 1996.
We’ve written before about how the Court might and should rule in Gonzalez (see here and here), but less attention has been devoted to the other Section 230 case on the docket: Twitter v. Taamneh. That’s unfortunate, as a thoughtful reading of the dispute at issue in Taamneh could highlight some of the law’s underlying principles. At first blush, alas, it does not appear that the Court is primed to apply that reading.
During the recent oral arguments, the Court considered whether Twitter (and other social-media companies) can be held liable under the Antiterrorism Act for providing a general communications platform that may be used by terrorists. The question under review by the Court is whether Twitter “‘knowingly’ provided substantial assistance [to terrorist groups] under [the statute] merely because it allegedly could have taken more ‘meaningful’ or ‘aggressive’ action to prevent such use.” Plaintiffs’ (respondents before the Court) theory is, essentially, that Twitter aided and abetted terrorism through its inaction.
The oral argument found the justices grappling with where to draw the line between aiding and abetting, and otherwise legal activity that happens to make it somewhat easier for bad actors to engage in illegal conduct. The nearly three-hour discussion between the justices and the attorneys yielded little in the way of a viable test. But a more concrete focus on the law & economics of collateral liability (which we also describe as “intermediary liability”) would have significantly aided the conversation.
Taamneh presents a complex question of intermediary liability generally that goes beyond the bounds of a (relatively) simpler Section 230 analysis. As we discussed in our amicus brief in Fleites v. Mindgeek (and as briefly described in this blog post), intermediary liability generally cannot be predicated on the mere existence of harmful or illegal content on an online platform that could, conceivably, have been prevented by some action by the platform or other intermediary.
The specific statute may impose other limits (like the “knowing” requirement in the Antiterrorism Act), but intermediary liability makes sense only when a particular intermediary defendant is positioned to control (and thus remedy) the bad conduct in question, and when imposing liability would cause the intermediary to act in such a way that the benefits of its conduct in deterring harm outweigh the costs of impeding its normal functioning as an intermediary.
Had the Court adopted such an approach in its questioning, it could have better honed in on the proper dividing line between parties whose normal conduct might reasonably give rise to liability (without some heightened effort to mitigate harm) and those whose conduct should not entail this sort of elevated responsibility.
Here, the plaintiffs have framed their case in a way that would essentially collapse this analysis into a strict liability standard by simply asking “did something bad happen on a platform that could have been prevented?” As we discuss below, the plaintiffs’ theory goes too far and would overextend intermediary liability to the point that the social costs would outweigh the benefits of deterrence.
The Law & Economics of Intermediary Liability: Who’s Best Positioned to Monitor and Control?
In our amicus brief in Fleites v. MindGeek (as well as our law review article on Section 230 and intermediary liability), we argued that, in limited circumstances, the law should (and does) place responsibility on intermediaries to monitor and control conduct. It is not always sufficient to aim legal sanctions solely at the parties who commit harms directly—e.g., where harms are committed by many pseudonymous individuals dispersed across large online services. In such cases, social costs may be minimized when legal responsibility is placed upon the least-cost avoider: the party in the best position to limit harm, even if it is not the party directly committing the harm.
Thus, in some circumstances, intermediaries (like Twitter) may be the least-cost avoider, such as when information costs are sufficiently low that effective monitoring and control of end users is possible, and when pseudonymity makes remedies against end users ineffective.
But there are costs to imposing such liability—including, importantly, “collateral censorship” of user-generated content by online social-media platforms. This manifests in platforms acting more defensively—taking down more speech, and generally moving in a direction that would make the Internet less amenable to open, public discussion—in an effort to avoid liability. Indeed, a core reason that Section 230 exists in the first place is to reduce these costs. (Whether Section 230 gets the balance correct is another matter, which we take up at length in our law review article linked above).
From an economic perspective, liability should be imposed on the party or parties best positioned to deter the harms in question, so long as the social costs incurred by, and as a result of, enforcement do not exceed the social gains realized. In other words, there is a delicate balance that must be struck to determine when intermediary liability makes sense in a given case. On the one hand, we want illicit content to be deterred, and on the other, we want to preserve the open nature of the Internet. The costs generated from the over-deterrence of legal, beneficial speech is why intermediary liability for user-generated content can’t be applied on a strict-liability basis, and why some bad content will always exist in the system.
The Spectrum of Properly Construed Intermediary Liability: Lessons from Fleites v. Mindgeek
Fleites v. Mindgeek illustrates well that proper application of liability to intermedium exists on a spectrum. Mindgeek—the owner/operator of the website Pornhub—was sued under Racketeer Influenced and Corrupt Organizations Act (RICO) and Victims of Trafficking and Violence Protection Act (TVPA) theories for promoting and profiting from nonconsensual pornography and human trafficking. But the plaintiffs also joined Visa as a defendant, claiming that Visa knowingly provided payment processing for some of Pornhub’s services, making it an aider/abettor.
The “best” defendants, obviously, would be the individuals actually producing the illicit content, but limiting enforcement to direct actors may be insufficient. The statute therefore contemplates bringing enforcement actions against certain intermediaries for aiding and abetting. But there are a host of intermediaries you could theoretically bring into a liability scheme. First, obviously, is Mindgeek, as the platform operator. Plaintiffs felt that Visa was also sufficiently connected to the harm by processing payments for MindGeek users and content posters, and that it should therefore bear liability, as well.
The problem, however, is that there is no limiting principle in the plaintiffs’ theory of the case against Visa. Theoretically, the group of intermediaries “facilitating” the illicit conduct is practically limitless. As we pointed out in our Fleites amicus:
In theory, any sufficiently large firm with a role in the commerce at issue could be deemed liable if all that is required is that its services “allow[]” the alleged principal actors to continue to do business. FedEx, for example, would be liable for continuing to deliver packages to MindGeek’s address. The local waste management company would be liable for continuing to service the building in which MindGeek’s offices are located. And every online search provider and Internet service provider would be liable for continuing to provide service to anyone searching for or viewing legal content on MindGeek’s sites.
Twitter’s attorney in Taamneh, Seth Waxman, made much the same point in responding to Justice Sonia Sotomayor:
…the rule that the 9th Circuit has posited and that the plaintiffs embrace… means that as a matter of course, every time somebody is injured by an act of international terrorism committed, planned, or supported by a foreign terrorist organization, each one of these platforms will be liable in treble damages and so will the telephone companies that provided telephone service, the bus company or the taxi company that allowed the terrorists to move about freely. [emphasis added]
In our Fleites amicus, we argued that a more practical approach is needed; one that tries to draw a sensible line on this liability spectrum. Most importantly, we argued that Visa was not in a position to monitor and control what happened on MindGeek’s platform, and thus was a poor candidate for extending intermediary liability. In that case, because of the complexities of the payment-processing network, Visa had no visibility into what specific content was being purchased, what content was being uploaded to Pornhub, and which individuals may have been uploading illicit content. Worse, the only evidence—if it can be called that—that Visa was aware that anything illicit was happening consisted of news reports in the mainstream media, which may or may not have been accurate, and on which Visa was unable to do any meaningful follow-up investigation.
Our Fleites brief didn’t explicitly consider MindGeek’s potential liability. But MindGeek obviously is in a much better position to monitor and control illicit content. With that said, merely having the ability to monitor and control is not sufficient. Given that content moderation is necessarily an imperfect activity, there will always be some bad content that slips through. Thus, the relevant question is, under the circumstances, did the intermediary act reasonably—e.g., did it comply with best practices—in attempting to identify, remove, and deter illicit content?
In Visa’s case, the answer is not difficult. Given that it had no way to know about or single out transactions as likely to be illegal, its only recourse to reduce harm (and its liability risk) would be to cut off all payment services for Mindgeek. The constraints on perfectly legal conduct that this would entail certainly far outweigh the benefits of reducing illegal activity.
Moreover, such a theory could require Visa to stop processing payments for an enormous swath of legal activity outside of PornHub. For example, purveyors of illegal content on PornHub use ISP services to post their content. A theory of liability that held Visa responsible simply because it plays some small part in facilitating the illegal activity’s existence would presumably also require Visa to shut off payments to ISPs—certainly, that would also curtail the amount of illegal content.
With MindGeek, the answer is a bit more difficult. The anonymous or pseudonymous posting of pornographic content makes it extremely difficult to go after end users. But knowing that human trafficking and nonconsensual pornography are endemic problems, and knowing (as it arguably did) that such content was regularly posted on Pornhub, Mindgeek could be deemed to have acted unreasonably for not having exercised very strict control over its users (at minimum, say, by verifying users’ real identities to facilitate law enforcement against bad actors and deter the posting of illegal content). Indeed, it is worth noting that MindGeek/Pornhub did implement exactly this control, among others, following the public attention arising from news reports of nonconsensual and trafficking-related content on the site.
But liability for MindGeek is only even plausible given that it might be able to act in such a way that imposes greater burdens on illegal content providers without deterring excessive amounts of legal content. If its only reasonable means of acting would be, say, to shut down PornHub entirely, then just as with Visa, the cost of imposing liability in terms of this “collateral censorship” would surely outweigh the benefits.
Applying the Law & Economics of Collateral Liability to Twitter in Taamneh
Contrast the situation of MindGeek in Fleites with Twitter in Taamneh. Twitter may seem to be a good candidate for intermediary liability. It also has the ability to monitor and control what is posted on its platform. And it faces a similar problem of pseudonymous posting that may make it difficult to go after end users for terrorist activity. But this is not the end of the analysis.
Given that Twitter operates a platform that hosts the general—and overwhelmingly legal—discussions of hundreds of millions of users, posting billions of pieces of content, it would be reasonable to impose a heightened responsibility on Twitter only if it could exercise it without excessively deterring the copious legal content on its platform.
At the same time, Twitter does have active policies to police and remove terrorist content. The relevant question, then, is not whether it should do anything to police such content, but whether a failure to do some unspecified amount more was unreasonable, such that its conduct should constitute aiding and abetting terrorism.
Under the Antiterrorism Act, because the basis of liability is “knowingly providing substantial assistance” to a person who committed an act of international terrorism, “unreasonableness” here would have to mean that the failure to do more transforms its conduct from insubstantial to substantial assistance and/or that the failure to do more constitutes a sort of willful blindness.
The problem is that doing more—policing its site better and removing more illegal content—would do nothing to alter the extent of assistance it provides to the illegal content that remains. And by the same token, almost by definition, Twitter does not “know” about illegal content it fails to remove. In theory, there is always “more” it could do. But given the inherent imperfections of content moderation at scale, this will always be true, right up to the point that the platform is effectively forced to shut down its service entirely.
This doesn’t mean that reasonable content moderation couldn’t entail something more than Twitter was doing. But it does mean that the mere existence of illegal content that, in theory, Twitter could have stopped can’t be the basis of liability. And yet the Taamneh plaintiffs make no allegation that acts of terrorism were actually planned on Twitter’s platform, and offer no reasonable basis on which Twitter could have practical knowledge of such activity or practical opportunity to control it.
Nor did plaintiffs point out any examples where Twitter had actual knowledge of such content or users and failed to remove them. Most importantly, the plaintiffs did not demonstrate that any particular content-moderation activities (short of shutting down Twitter entirely) would have resulted in Twitter’s knowledge of or ability to control terrorist activity. Had they done so, it could conceivably constitute a basis for liability. But if the only practical action Twitter can take to avoid liability and prevent harm entails shutting down massive amounts of legal speech, the failure to do so cannot be deemed unreasonable or provide the basis for liability.
And, again, such a theory of liability would contain no viable limiting principle if it does not consider the practical ability to control harmful conduct without imposing excessively costly collateral damage. Indeed, what in principle would separate a search engine from Twitter, if the search engine linked to an alleged terrorist’s account? Both entities would have access to news reports, and could thus be assumed to have a generalized knowledge that terrorist content might exist on Twitter. The implication of this case, if the plaintiff’s theory is accepted, is that Google would be forced to delist Twitter whenever a news article appears alleging terrorist activity on the service. Obviously, that is untenable for the same reason it’s not tenable to impose an effective obligation on Twitter to remove all terrorist content: the costs of lost legal speech and activity.
Justice Ketanji Brown Jackson seemingly had the same thought when she pointedly asked whether the plaintiffs’ theory would mean that Linda Hamilton in the Halberstram v. Welch case could have been held liable for aiding and abetting, merely for taking care of Bernard Welch’s kids at home while Welch went out committing burglaries and the murder of Michael Halberstram (instead of the real reason she was held liable, which was for doing Welch’s bookkeeping and helping sell stolen items). As Jackson put it:
…[I]n the Welch case… her taking care of his children [was] assisting him so that he doesn’t have to be at home at night? He’s actually out committing robberies. She would be assisting his… illegal activities, but I understood that what made her liable in this situation is that the assistance that she was providing was… assistance that was directly aimed at the criminal activity. It was not sort of this indirect supporting him so that he can actually engage in the criminal activity.
In sum, the theory propounded by the plaintiffs (and accepted by the 9th U.S. Circuit Court of Appeals) is just too far afield for holding Twitter liable. As Twitter put it in its reply brief, the plaintiffs’ theory (and the 9th Circuit’s holding) is that:
…providers of generally available, generic services can be held responsible for terrorist attacks anywhere in the world that had no specific connection to their offerings, so long as a plaintiff alleges (a) general awareness that terrorist supporters were among the billions who used the services, (b) such use aided the organization’s broader enterprise, though not the specific attack that injured the plaintiffs, and (c) the defendant’s attempts to preclude that use could have been more effective.
Conclusion
If Section 230 immunity isn’t found to apply in Gonzalez v. Google, and the complaint in Taamneh is allowed to go forward, the most likely response of social-media companies will be to reduce the potential for liability by further restricting access to their platforms. This could mean review by some moderator or algorithm of messages or videos before they are posted to ensure that there is no terrorist content. Or it could mean review of users’ profiles before they are able to join the platform to try to ascertain their political leanings or associations with known terrorist groups. Such restrictions would entail copious false negatives, along with considerable costs to users and to open Internet speech.
And in the end, some amount of terrorist content would still get through. If the plaintiffs’ theory leads to liability in Taamneh, it’s hard to see how the same principle wouldn’t entail liability even under these theoretical, heightened practices. Absent a focus on an intermediary defendant’s ability to control harmful content or conduct, without imposing excessive costs on legal content or conduct, the theory of liability has no viable limit.
In sum, to hold Twitter (or other social-media platforms) liable under the facts presented in Taamneh would stretch intermediary liability far beyond its sensible bounds. While Twitter may seem a good candidate for intermediary liability in theory, it should not be held liable for, in effect, simply providing its services.
Perhaps Section 230’s blanket immunity is excessive. Perhaps there is a proper standard that could impose liability on online intermediaries for user-generated content in limited circumstances properly tied to their ability to control harmful actors and the costs of doing so. But liability in the circumstances suggested by the Taamneh plaintiffs—effectively amounting to strict liability—would be an even bigger mistake in the opposite direction.
It seems that large language models (LLMs) are all the rage right now, from Bing’s announcement that it plans to integrate the ChatGPT technology into its search engine to Google’s announcement of its own LLM called “Bard” to Meta’s recent introduction of its Large Language Model Meta AI, or “LLaMA.” Each of these LLMs use artificial intelligence (AI) to create text-based answers to questions.
But it certainly didn’t take long after these innovative new applications were introduced for reports to emerge of LLM models just plain getting facts wrong. Given this, it is worth asking: how will the law deal with AI-created misinformation?
Among the first questions courts will need to grapple with is whether Section 230 immunity applies to content produced by AI. Of course, the U.S. Supreme Court already has a major Section 230 case on its docket with Gonzalez v. Google. Indeed, during oral arguments for that case, Justice Neil Gorsuch appeared to suggest that AI-generated content would not receive Section 230 immunity. And existing case law would appear to support that conclusion, as LLM content is developed by the interactive computer service itself, not by its users.
Another question raised by the technology is what legal avenues would be available to those seeking to challenge the misinformation. Under the First Amendment, the government can only regulate false speech under very limited circumstances. One of those is defamation, which seems like the most logical cause of action to apply. But under defamation law, plaintiffs—especially public figures, who are the most likely litigants and who must prove “malice”—may have a difficult time proving the AI acted with the necessary state of mind to sustain a cause of action.
Section 230 Likely Does Not Apply to Information Developed by an LLM
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.
The law defines an interactive computer service as “any information service, system, or access software provider that provides or enables computer access by multiple users to a computer server, including specifically a service or system that provides access to the Internet and such systems operated or services offered by libraries or educational institutions.”
The access software provider portion of that definition includes any tool that can “filter, screen, allow, or disallow content; pick, choose, analyze, or digest content; or transmit, receive, display, forward, cache, search, subset, organize, reorganize, or translate content.”
And finally, an information content provider is “any person or entity that is responsible, in whole or in part, for the creation or development of information provided through the Internet or any other interactive computer service.”
Taken together, Section 230(c)(1) gives online platforms (“interactive computer services”) broad immunity for user-generated content (“information provided by another information content provider”). This even covers circumstances where the online platform (acting as an “access software provider”) engages in a great deal of curation of the user-generated content.
Section 230(c)(1) does not, however, protect information created by the interactive computer service itself.
There is case law to help determine whether content is created or developed by the interactive computer service. Online platforms applying “neutral tools” to help organize information have not lost immunity. As the 9th U.S. Circuit Court of Appeals put it in Fair Housing Council v. Roommates.com:
Providing neutral tools for navigating websites is fully protected by CDA immunity, absent substantial affirmative conduct on the part of the website creator promoting the use of such tools for unlawful purposes.
On the other hand, online platforms are liable for content they create or develop, which does not include “augmenting the content generally,” but does include “materially contributing to its alleged unlawfulness.”
The question here is whether the text-based answers provided by LLM apps like Bing’s Sydney or Google’s Bard comprise content created or developed by those online platforms. One could argue that LLMs are neutral tools simply rearranging information from other sources for display. It seems clear, however, that the LLM is synthesizing information to create new content. The use of AI to answer a question, rather than a human agent of Google or Microsoft, doesn’t seem relevant to whether or not it was created or developed by those companies. (Though, as Matt Perault notes, how LLMs are integrated into a product matters. If an LLM just helps “determine which search results to prioritize or which text to highlight from underlying search results,” then it may receive Section 230 protection.)
The technology itself gives text-based answers based on inputs from the questioner. LLMs uses AI-trained engines to guess the next word based on troves of data from the internet. While the information may come from third parties, the creation of the content itself is due to the LLM. As ChatGPT put it in response to my query here:
Proving Defamation by AI
In the absence of Section 230 immunity, there is still the question of how one could hold Google’s Bard or Microsoft’s Sydney accountable for purveying misinformation. There are no laws against false speech in general, nor can there be, since the Supreme Court declared such speech was protected in United States v. Alvarez. There are, however, categories of false speech, like defamation and fraud, which have been found to lie outside of First Amendment protection.
Defamation is the most logical cause of action that could be brought for false information provided by an LLM app. But it is notable that it is highly unlikely that people who have not received significant public recognition will be known by these LLM apps (believe me, I tried to get ChatGPT to tell me something about myself—alas, I’m not famous enough). On top of that, those most likely to have significant damages from their reputations being harmed by falsehoods spread online are those who are in the public eye. This means that, for the purposes of the defamation suit, it is public figures who are most likely to sue.
As an example, if ChatGPT answers the question of whether Johnny Depp is a wife-beater by saying that he is, contrary to one court’s finding (but consistent with another’s), Depp could sue the creators of the service for defamation. He would have to prove that a false statement was publicized to a third party that resulted in damages to him. For the sake of argument, let’s say he can do both. The case still isn’t proven because, as a public figure, he would also have to prove “actual malice.”
Under New York Times v. Sullivan and its progeny, a public figure must prove the defendant acted with “actual malice” when publicizing false information about the plaintiff. Actual malice is defined as “knowledge that [the statement] was false or with reckless disregard of whether it was false or not.”
The question arises whether actual malice can be attributed to a LLM. It seems unlikely that it could be said that the AI’s creators trained it in a way that they “knew” the answers provided would be false. But it may be a more interesting question whether the LLM is giving answers with “reckless disregard” of their truth or falsity. One could argue that these early versions of the technology are exactly that, but the underlying AI is likely to improve over time with feedback. The best time for a plaintiff to sue may be now, when the LLMs are still in their infancy and giving off false answers more often.
It is possible that, given enough context in the query, LLM-empowered apps may be able to recognize private figures, and get things wrong. For instance, when I asked ChatGPT to give a biography of myself, I got no results:
When I added my workplace, I did get a biography, but none of the relevant information was about me. It was instead about my boss, Geoffrey Manne, the president of the International Center for Law & Economics:
While none of this biography is true, it doesn’t harm my reputation, nor does it give rise to damages. But it is at least theoretically possible that an LLM could make a defamatory statement against a private person. In such a case, a lower burden of proof would apply to the plaintiff, that of negligence, i.e., that the defendant published a false statement of fact that a reasonable person would have known was false. This burden would be much easier to meet if the AI had not been sufficiently trained before being released upon the public.
Conclusion
While it is unlikely that a service like ChapGPT would receive Section 230 immunity, it also seems unlikely that a plaintiff would be able to sustain a defamation suit against it for false statements. The most likely type of plaintiff (public figures) would encounter difficulty proving the necessary element of “actual malice.” The best chance for a lawsuit to proceed may be against the early versions of this service—rolled out quickly and to much fanfare, while still being in a beta stage in terms of accuracy—as a colorable argument can be made that they are giving false answers in “reckless disregard” of their truthfulness.
In a Feb. 14 column in the Wall Street Journal, Commissioner Christine Wilson announced her intent to resign her position on the Federal Trade Commission (FTC).For those curious to know why, she beat you to the punch in the title and subtitle of her column: “Why I’m Resigning as an FTC Commissioner: Lina Khan’s disregard for the rule of law and due process make it impossible for me to continue serving.”
This is the seventh FTC roundup I’ve posted to Truth on the Market since joining the International Center for Law & Economics (ICLE) last September, having left the FTC at the end of August. Relentlessly astute readers of this column may have observed that I cited (and linked to) Commissioner Wilson’s dissents in five of my six previous efforts—actually, to three of them in my Nov. 4 post alone.
As anyone might guess, I’ve linked to Wilson’s dissents (and concurrences, etc.) for the same reason I’ve linked to other sources: I found them instructive in some significant regard. Priors and particular conclusions of law aside, I generally found Wilson’s statements to be well-grounded in established principles of antitrust law and economics. I cannot say the same about statements from the current majority.
Commission dissents are not merely the bases for blog posts or venues for venting. They can provide a valuable window into agency matters for lawmakers and, especially, for the courts. And I would suggest that they serve an important institutional role at the FTC, whatever one thinks of the merits of any specific matter. There’s really no point to having a five-member commission if all its votes are unanimous and all its opinions uniform. Moreover, establishing the realistic possibility of dissent can lend credence to those commission opinions that are unanimous. And even in these fractious times, there are such opinions.
Wilson did not spring forth fully formed from the forehead of the U.S. Senate. She began her FTC career as a Georgetown student, serving as a law clerk in the Bureau of Competition; she returned some years later to serve as chief of staff to Chairman Tim Muris; and she returned again when confirmed as a commissioner in April 2018 (later sworn in in September 2018). In between stints at the FTC, she gained antitrust experience in private practice, both in law firms and as in-house counsel. I would suggest that her agency experience, combined with her work in the private sector, provided a firm foundation for the judgments required of a commissioner.
Daniel Kaufman, former acting director of the FTC’s Bureau of Consumer Protection, reflected on Wilson’s departure here. Personally, with apologies for the platitude, I would like to thank Commissioner Wilson for her service. And, not incidentally, for her consistent support for agency staff.
Her three Democratic colleagues on the commission also thanked her for her service, if only collectively, and tersely: “While we often disagreed with Commissioner Wilson, we respect her devotion to her beliefs and are grateful for her public service. We wish her well in her next endeavor.” That was that. No doubt heartfelt. Wilson’s departure column was a stern rebuke to the Commission, so there’s that. But then, stern rebukes fly in all directions nowadays.
While I’ve never been a commissioner, I recall a far nicer and more collegial sendoff when I departed from my lowly staff position. Come to think of it, I had a nicer sendoff when I left a large D.C. law firm as a third-year associate bound for a teaching position, way back when.
So, what else is new?
In January, I noted that “the big news at the FTC is all about noncompetes”; that is, about the FTC’s proposed rule to ban the use of noncompetes more-or-less across the board The rule would cover all occupations and all income levels, with a narrow exception for the sale of the business in which the “employee” has at least a 25% ownership stake (why 25%?), and a brief nod to statutory limits on the commission’s regulatory authority with regard to nonprofits, common carriers, and some other entities.
Colleagues Brian Albrecht (and here),Alden Abbott, Gus Hurwitz, and Corbin K. Barthold also have had things to say about it. I suggested that there were legitimate reasons to be concerned about noncompetes in certain contexts—sometimes on antitrust grounds, and sometimes for other reasons. But certain contexts are far from all contexts, and a mixed and developing body of economic literature, coupled with limited FTC experience in the subject, did not militate in favor of nearly so sweeping a regulatory proposal. This is true even before we ask practical questions about staffing for enforcement or, say, whether the FTC Act conferred the requisite jurisdiction on the agency.
This is the first or second FTC competition rulemaking ever, depending on how one counts, and it is the first this century, in any case. Here’s administrative scholar Thomas Merrill on FTC competition rulemaking. Given the Supreme Court’s recent articulation of the major questions doctrine in West Virginia v. EPA, a more modest and bipartisan proposal might have been far more prudent. A bad turn at the court can lose more than the matter at hand. Comments are due March 20, by the way.
Now comes a missive from the House Judiciary Committee, along with multiple subcommittees, about the noncompete NPRM. The letter opens by stating that “The Proposed Rule exceeds its delegated authority and imposes a top-down one-size-fits-all approach that violates basic American principles of federalism and free markets.” And “[t]he Biden FTC’s proposed rule on non-compete clauses shows the radicalness of the so-called ‘hipster’ antitrust movement that values progressive outcomes over long-held legal and economic principles.”
Ouch. Other than that Mr. Jordan, how did you like the play?
There are several single-spaced pages on the “FTC’s power grab” before the letter gets to a specific, and substantial, formal document request in the service of congressional oversight. That does not stop the rulemaking process, but it does not bode well either.
Part of why this matters is that there’s still solid, empirically grounded, pro-consumer work that’s at risk. In my first Truth on the Market post, I applauded FTC staff commentsurging New York State to reject a certificate of public advantage (COPA) application. As I noted there, COPAs are rent-seeking mechanisms chiefly aimed at insulating anticompetitive mergers (and sometimes conduct) from federal antitrust scrutiny. Commission and staff opposition to COPAs was developed across several administrations on well-established competition principles and a significant body of research regarding hospital consolidation, health care prices, and quality of care.
Office of Policy Planning (OPP) Director Elizabeth Wilkins has now announced that the parties in question have abandoned their proposed merger. Wilkins thanks the staff of OPP, the Bureau of Economics, and the Bureau of Competition for their work on the matter, and rightly so. There’s no new-fangled notion of Section 5 or mergers at play. The work has developed over decades and it’s the sort of work that should continue. Notwithstanding numerous (if not legion) departures, good and experienced staff and established methods remain, and ought not to be repudiated, much less put at risk.
I won’t recapitulate the much-discussed case, but on the somewhat-less-discussed matter of the withdrawal, I’ll consider why the FTC announced that the matter “is withdrawn from adjudication, and that all proceedings before the Administrative Law Judge be and they hereby are stayed.” While the matter was not litigated to its conclusion in federal court, the substantial and workmanlike opinion denying the preliminary injunction made it clear that the FTC had lost on the facts under both of the theories of harm to potential competition that they’d advanced.
“Having reviewed and considered the objective evidence of Meta’s capabilities and incentives, the Court is not persuaded that this evidence establishes that it was ‘reasonably probable’ Meta would enter the relevant market.”
An appeal in the 9th U.S. Circuit Court of Appeals likely seemed fruitless. Stopping short of a final judgment, the FTC could have tried for a do-over in its internal administrative Part 3 process, and might have fared well before itself, but that would have demanded considerable additional resources in a case that, in the long run, was bound to be a loser. Bloomberg had previously reported that the commission voted to proceed with the case against the merger contra the staff’s recommendation. Here, the commission noted that “Complaint Counsel [the Commission’s own staff] has not registered any objection” to Meta’s motion to withdraw proceedings from adjudication.
There are novel approaches to antitrust. And there are the courts and the law. And, as noted above, many among the staff are well-versed in that law and experienced at investigations. You can’t always get what you want, but if you try sometimes, you get what you deserve.
In the world of video games, the process by which players train themselves or their characters in order to overcome a difficult “boss battle” is called “leveling up.” I find that the phrase also serves as a useful metaphor in the context of corporate mergers. Here, “leveling up” can be thought of as acquiring another firm in order to enter or reinforce one’s presence in an adjacent market where a larger and more successful incumbent is already active.
In video-game terminology, that incumbent would be the “boss.” Acquiring firms choose to level up when they recognize that building internal capacity to compete with the “boss” is too slow, too expensive, or is simply infeasible. An acquisition thus becomes the only way “to beat the boss” (or, at least, to maximize the odds of doing so).
Alas, this behavior is often mischaracterized as a “killer acquisition” or “reverse killer acquisition.” What separates leveling up from killer acquisitions is that the former serve to turn the merged entity into a more powerful competitor, while the latter attempt to weaken competition. In the case of “reverse killer acquisitions,” the assumption is that the acquiring firm would have entered the adjacent market regardless absent the merger, leaving even more firms competing in that market.
In other words, the distinction ultimately boils down to a simple (though hard to answer) question: could both the acquiring and target firms have effectively competed with the “boss” without a merger?
Because they are ubiquitous in the tech sector, these mergers—sometimes also referred to as acquisitions of nascent competitors—have drawn tremendous attention from antitrust authorities and policymakers. All too often, policymakers fail to adequately consider the realistic counterfactual to a merger and mistake leveling up for a killer acquisition. The most recent high-profile example is Meta’s acquisition of the virtual-reality fitness app Within. But in what may be a hopeful sign of a turning of the tide, a federal court appears set to clear that deal over objections from the Federal Trade Commission (FTC).
Some Recent ‘Boss Battles’
The canonical example of leveling up in tech markets is likely Google’s acquisition of Android back in 2005. While Apple had not yet launched the iPhone, it was already clear by 2005 that mobile would become an important way to access the internet (including Google’s search services). Rumors were swirling that Apple, following its tremendously successful iPod, had started developing a phone, and Microsoft had been working on Windows Mobile for a long time.
In short, there was a serious risk that Google would be reliant on a single mobile gatekeeper (i.e., Apple) if it did not move quickly into mobile. Purchasing Android was seen as the best way to do so. (Indeed, averting an analogous sort of threat appears to be driving Meta’s move into virtual reality today.)
The natural next question is whether Google or Android could have succeeded in the mobile market absent the merger. My guess is that the answer is no. In 2005, Google did not produce any consumer hardware. Quickly and successfully making the leap would have been daunting. As for Android:
Google had significant advantages that helped it to make demands from carriers and OEMs that Android would not have been able to make. In other words, Google was uniquely situated to solve the collective action problem stemming from OEMs’ desire to modify Android according to their own idiosyncratic preferences. It used the appeal of its app bundle as leverage to get OEMs and carriers to commit to support Android devices for longer with OS updates. The popularity of its apps meant that OEMs and carriers would have great difficulty in going it alone without them, and so had to engage in some contractual arrangements with Google to sell Android phones that customers wanted. Google was better resourced than Android likely would have been and may have been able to hold out for better terms with a more recognizable and desirable brand name than a hypothetical Google-less Android. In short, though it is of course possible that Android could have succeeded despite the deal having been blocked, it is also plausible that Android became so successful only because of its combination with Google. (citations omitted)
In short, everything suggests that Google’s purchase of Android was a good example of leveling up. Note that much the same could be said about the company’s decision to purchase Fitbit in order to compete against Apple and its Apple Watch (which quickly dominated the market after its launch in 2015).
A more recent example of leveling up is Microsoft’s planned acquisition of Activision Blizzard. In this case, the merger appears to be about improving Microsoft’s competitive position in the platform market for game consoles, rather than in the adjacent market for games.
At the time of writing, Microsoft is staring down the barrel of a gun: Sony is on the cusp of becoming the runaway winner of yet another console generation. Microsoft’s executives appear to have concluded that this is partly due to a lack of exclusive titles on the Xbox platform. Hence, they are seeking to purchase Activision Blizzard, one of the most successful game studios, known among other things for its acclaimed Call of Duty series.
Again, the question is whether Microsoft could challenge Sony by improving its internal game-publishing branch (known as Xbox Game Studios) or whether it needs to acquire a whole new division. This is obviously a hard question to answer, but a cursory glance at the titles shipped by Microsoft’s publishing studio suggest that the issues it faces could not simply be resolved by throwing more money at its existing capacities. Indeed, Microsoft Game Studios seems to be plagued by organizational failings that might only be solved by creating more competition within the Microsoft company. As one gaming journalist summarized:
The current predicament of these titles goes beyond the amount of money invested or the buzzwords used to market them – it’s about Microsoft’s plan to effectively manage its studios. Encouraging independence isn’t an excuse for such a blatantly hands-off approach which allows titles to fester for years in development hell, with some fostering mistreatment to occur. On the surface, it’s just baffling how a company that’s been ranked as one of the top 10 most reputable companies eight times in 11 years (as per RepTrak) could have such problems with its gaming division.
The upshot is that Microsoft appears to have recognized that its own game-development branch is failing, and that acquiring a well-functioning rival is the only way to rapidly compete with Sony. There is thus a strong case to be made that competition authorities and courts should approach the merger with caution, as it has at least the potential to significantly increase competition in the game-console industry.
Finally, leveling up is sometimes a way for smaller firms to try and move faster than incumbents into a burgeoning and promising segment. The best example of this is arguably Meta’s effort to acquire Within, a developer of VR fitness apps. Rather than being an attempt to thwart competition from a competitor in the VR app market, the goal of the merger appears to be to compete with the likes of Google, Apple, and Sony at the platform level. As Mark Zuckerberg wrote back in 2015, when Meta’s VR/AR strategy was still in its infancy:
Our vision is that VR/AR will be the next major computing platform after mobile in about 10 years… The strategic goal is clearest. We are vulnerable on mobile to Google and Apple because they make major mobile platforms. We would like a stronger strategic position in the next wave of computing….
Over the next few years, we’re going to need to make major new investments in apps, platform services, development / graphics and AR. Some of these will be acquisitions and some can be built in house. If we try to build them all in house from scratch, then we risk that several will take too long or fail and put our overall strategy at serious risk. To derisk this, we should acquire some of these pieces from leading companies.
In short, many of the tech mergers that critics portray as killer acquisitions are just as likely to be attempts by firms to compete head-on with incumbents. This “leveling up” is precisely the sort of beneficial outcome that antitrust laws were designed to promote.
Building Products Is Hard
Critics are often quick to apply the “killer acquisition” label to any merger where a large platform is seeking to enter or reinforce its presence in an adjacent market. The preceding paragraphs demonstrate that it’s not that simple, as these mergers often enable firms to improve their competitive position in the adjacent market. For obvious reasons, antitrust authorities and policymakers should be careful not to thwart this competition.
The harder part is how to separate the wheat from the chaff. While I don’t have a definitive answer, an easy first step would be for authorities to more seriously consider the supply side of the equation.
Building a new product is incredibly hard, even for the most successful tech firms. Microsoft famously failed with its Zune music player and Windows Phone. The Google+ social network never gained any traction. Meta’s foray into the cryptocurrency industry was a sobering experience. Amazon’s Fire Phone bombed. Even Apple, which usually epitomizes Silicon Valley firms’ ability to enter new markets, has had its share of dramatic failures: Apple Maps, its Ping social network, and the first Home Pod, to name a few.
To put it differently, policymakers should not assume that internal growth is always a realistic alternative to a merger. Instead, they should carefully examine whether such a strategy is timely, cost-effective, and likely to succeed.
This is obviously a daunting task. Firms will struggle to dispositively show that they need to acquire the target firm in order to effectively compete against an incumbent. The question essentially hinges on the quality of the firm’s existing management, engineers, and capabilities. All of these are difficult—perhaps even impossible—to measure. At the very least, policymakers can improve the odds of reaching a correct decision by approaching these mergers with an open mind.
Under Chair Lina Khan’s tenure, the FTC has opted for the opposite approach and taken a decidedly hostile view of tech acquisitions. The commission sued to block both Meta’s purchase of Within and Microsoft’s acquisition of Activision Blizzard. Likewise, several economists—notably Tommasso Valletti—have called for policymakers to reverse the burden of proof in merger proceedings, and opined that all mergers should be viewed with suspicion because, absent efficiencies, they always reduce competition.
Unfortunately, this skeptical approach is something of a self-fulfilling prophecy: when authorities view mergers with suspicion, they are likely to be dismissive of the benefits discussed above. Mergers will be blocked and entry into adjacent markets will occur via internal growth.
Large tech companies’ many failed attempts to enter adjacent markets via internal growth suggest that such an outcome would ultimately harm the digital economy. Too many “boss battles” will needlessly be lost, depriving consumers of precious competition and destroying startup companies’ exit strategies.
Output of the LG Research AI to the prompt: “a system of copyright for artificial intelligence”
Not only have digital-image generators like Stable Diffusion, DALL-E, and Midjourney—which make use of deep-learning models and other artificial-intelligence (AI) systems—created some incredible (and sometimes creepy – see above) visual art, but they’ve engendered a good deal of controversy, as well. Human artists have banded together as part of a fledgling anti-AI campaign; lawsuits have been filed; and policy experts have been trying to think through how these machine-learning systems interact with various facets of the law.
Debates about the future of AI have particular salience for intellectual-property rights. Copyright is notoriously difficult to protect online, and these expert systems add an additional wrinkle: it can at least argued that their outputs can be unique creations. There are also, of course, moral and philosophical objections to those arguments, with many grounded in the supposition that only a human (or something with a brain, like humans) can be creative.
Leaving aside for the moment a potentially pitched battle over the definition of “creation,” we should be able to find consensus that at least some of these systems produce unique outputs and are not merely cutting and pasting other pieces of visual imagery into a new whole. That is, at some level, the machines are engaging in a rudimentary sort of “learning” about how humans arrange colors and lines when generating images of certain subjects. The machines then reconstruct this process and produce a new set of lines and colors that conform to the patterns they found in the human art.
But that isn’t the end of the story. Even if some of these systems’ outputs are unique and noninfringing, the way the machines learn—by ingesting existing artwork—can raise a number of thorny issues. Indeed, these systems are arguably infringing copyright during the learning phase, and such use may not survive a fair-use analysis.
We are still in the early days of thinking through how this new technology maps onto the law. Answers will inevitably come, but for now, there are some very interesting questions about the intellectual-property implications of AI-generated art, which I consider below.
The Points of Collision Between Intellectual Property Law and AI-Generated Art
AI-generated art is not a single thing. It is, rather, a collection of differing processes, each with different implications for the law. For the purposes of this post, I am going to deal with image-generation systems that use “generated adversarial networks” (GANs) and diffusion models. The various implementations of each will differ in some respects, but from what I understand, the ways that these techniques can be used generate all sorts of media are sufficiently similar that we can begin to sketch out some of their legal implications.
A (very) brief technical description
This is a very high-level overview of how these systems work; for a more detailed (but very readable) description, see here.
A GAN is a type of machine-learning model that consists of two parts: a generator and a discriminator. The generator is trained to create new images that look like they come from a particular dataset, while the discriminator is trained to distinguish the generated images from real images in the dataset. The two parts are trained together in an adversarial manner, with the generator trying to produce images that can fool the discriminator and the discriminator trying to correctly identify the generated images.
A diffusion model, by contrast, analyzes the distribution of information in an image, as noise is progressively added to it. This kind of algorithm analyzes characteristics of sample images—like the distribution of colors or lines—in order to “understand” what counts as an accurate representation of a subject (i.e., what makes a picture of a cat look like a cat and not like a dog).
For example, in the generation phase, systems like Stable Diffusion start with randomly generated noise, and work backward in “denoising” steps to essentially “see” shapes:
The sampled noise is predicted so that if we subtract it from the image, we get an image that’s closer to the images the model was trained on (not the exact images themselves, but the distribution – the world of pixel arrangements where the sky is usually blue and above the ground, people have two eyes, cats look a certain way – pointy ears and clearly unimpressed).
It is relevant here that, once networks using these techniques are trained, they do not need to rely on saved copies of the training images in order to generate new images. Of course, it’s possible that some implementations might be designed in a way that does save copies of those images, but for the purposes of this post, I will assume we are talking about systems that save known works only during the training phase. The models that are produced during training are, in essence, instructions to a different piece of software about how to start with a text prompt from a user—a palette of pure noise—and progressively “discover” signal in that image until some new image emerges.
Input-stage use of intellectual property
The creators of OpenAI, one of the most popular AI tools, are not shy about their use of protected works in the training phase of AI algorithms. In comments to the U.S. Patent and Trademark Office (PTO), they note that:
…[m]odern AI systems require large amounts of data. For certain tasks, that data is derived from existing publicly accessible “corpora”… of data that include copyrighted works. By analyzing large corpora (which necessarily involves first making copies of the data to be analyzed), AI systems can learn patterns inherent in human-generated data and then use those patterns to synthesize similar data which yield increasingly compelling novel media in modalities as diverse as text, image, and audio. (emphasis added).
Thus, at the training stage, the most popular forms of machine-learning systems require making copies of existing works. And where the material being used is either not in the public domain or is not licensed, an infringement occurs (as Getty Images notes in a suit against Stability AI that it recently filed). Thus, some affirmative defense is needed to excuse the infringement.
Toward this end, OpenAI believes that its algorithmic training should qualify as a fair use. Other major services that use these AI techniques to “learn” from existing media would likely make similar arguments. But, at least in the way that OpenAI has framed the fair-use analysis (that these uses are sufficiently “transformative”), it’s not clear that they should qualify.
The purpose and character of the use
In brief, fair use—found in 17 USC § 107—provides for an affirmative defense against infringement when the use is “for purposes such as criticism, comment, news reporting, teaching…, scholarship, or research.” When weighing a fair-use defense, a court must balance a number of factors:
the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes;
the nature of the copyrighted work;
the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and
the effect of the use upon the potential market for or value of the copyrighted work.
OpenAI’s fair-use claim is rooted in the first factor: the nature and character of the use. I should note, then, that what follows is solely a consideration of Factor 1, with special attention paid to whether these uses are “transformative.” But it is important to stipulate fair-use analysis is a multi-factor test and that, even within the first factor, it’s not mandatory that a use be “transformative.” It is entirely possible that a court balancing all of the factors could, indeed, find that OpenAI is engaged in fair use, even if it does not agree that it is “transformative.”
Whether the use of copyrighted works to train an AI is “transformative” is certainly a novel question, but it is likely answered through an observation that the U.S. Supreme Court made in Campbell v. Acuff Rose Music:
[W]hat Sony said simply makes common sense: when a commercial use amounts to mere duplication of the entirety of an original, it clearly “supersede[s] the objects,”… of the original and serves as a market replacement for it, making it likely that cognizable market harm to the original will occur… But when, on the contrary, the second use is transformative, market substitution is at least less certain, and market harm may not be so readily inferred.
A key question, then, is whether training an AI on copyrighted works amounts to mere “duplication of the entirety of an original” or is sufficiently “transformative” to support a fair-use finding. Open AI, as noted above, believes its use is highly transformative. According to its comments:
Training of AI systems is clearly highly transformative. Works in training corpora were meant primarily for human consumption for their standalone entertainment value. The “object of the original creation,” in other words, is direct human consumption of the author’s expression. Intermediate copying of works in training AI systems is, by contrast, “non-expressive” the copying helps computer programs learn the patterns inherent in human-generated media. The aim of this process—creation of a useful generative AI system—is quite different than the original object of human consumption. The output is different too: nobody looking to read a specific webpage contained in the corpus used to train an AI system can do so by studying the AI system or its outputs. The new purpose and expression are thus both highly transformative.
But the way that Open AI frames its system works against its interests in this argument. As noted above, and reinforced in the immediately preceding quote, an AI system like DALL-E or Stable Diffusion is actually made of at least two distinct pieces. The first is a piece of software that ingests existing works and creates a file that can serve as instructions to the second piece of software. The second piece of software then takes the output of the first part and can produce independent results. Thus, there is a clear discontinuity in the process, whereby the ultimate work created by the system is disconnected from the creative inputs used to train the software.
Therefore, contrary to what Open AI asserts, the protected works are indeed ingested into the first part of the system “for their standalone entertainment value.” That is to say, the software is learning what counts as “standalone entertainment value” and therefore, the works mustbe used in those terms.
Surely, a computer is not sitting on a couch and surfing for its own entertainment. But it is solely for the very “standalone entertainment value” that the first piece of software is being shown copyrighted material. By contrast, parody or “remixing” uses incorporate the work into some secondary expression that transforms the input. The way these systems work is to learn what makes a piece entertaining and then to discard that piece altogether. Moreover, this use of art qua art most certainly interferes with the existing market insofar as this use is in lieu of reaching a licensing agreement with rightsholders.
The 2nd U.S. Circuit Court of Appeals dealt with an analogous case. In American Geophysical Union v. Texaco, the 2nd Circuit considered whether Texaco’s photocopying of scientific articles produced by the plaintiffs qualified for a fair-use defense. Texaco employed between 400 and 500 research scientists and, as part of supporting their work, maintained subscriptions to a number of scientific journals. It was common practice for Texaco’s scientists to photocopy entire articles and save them in a file.
The plaintiffs sued for copyright infringement. Texaco asserted that photocopying by its scientists for the purposes of furthering scientific research—that is to train the scientists on the content of the journal articles—should count as a fair use, at least in part because it was sufficiently “transformative.” The 2nd Circuit disagreed:
The “transformative use” concept is pertinent to a court’s investigation under the first factor because it assesses the value generated by the secondary use and the means by which such value is generated. To the extent that the secondary use involves merely an untransformed duplication, the value generated by the secondary use is little or nothing more than the value that inheres in the original. Rather than making some contribution of new intellectual value and thereby fostering the advancement of the arts and sciences, an untransformed copy is likely to be used simply for the same intrinsic purpose as the original, thereby providing limited justification for a finding of fair use… (emphasis added).
As in the case at hand, the 2nd Circuit observed that making full copies of the scientific articles was solely for the consumption of the material itself. A rejoinder, of course, is that training these AI systems surely advances scientific research and, thus, does foster the “advancement of the arts and sciences.” But in American Geophysical Union, where the secondary use was explicitly for the creation of new and different scientific outputs, the court still held that making copies of one scientific article in order to learn and produce new scientific innovations did not count as “transformative.”
What this case represents is that one cannot merely state that some social goal will be advanced in the future by permitting an exception to copyright protection today. As the 2nd Circuit put it:
…the dominant purpose of the use is a systematic institutional policy of multiplying the available number of copies of pertinent copyrighted articles by circulating the journals among employed scientists for them to make copies, thereby serving the same purpose for which additional subscriptions are normally sold, or… for which photocopying licenses may be obtained.
The secondary use itself must be transformative and different. Where an AI system ingests copyrighted works, that use is simply not transformative; it is using the works in their original sense in order to train a system to be able to make other original works. As in American Geophysical Union, the AI creators are completely free to seek licenses from rightsholders in order to train their systems.
Finally, there is a sense in which this machine learning might not infringe on copyrights at all. To my knowledge, the technology does not itself exist, but if it were possible for a machine to somehow “see” in the way that humans do—without using stored copies of copyrighted works—merely “learning” from those works, such as we can call it learning, probably would not violate copyright laws.
Do the outputs of these systems violate intellectual property laws?
The outputs of GANs and diffusion models may or may not violate IP laws, but there is nothing inherent in the processes described above to dictate that they must. As noted, the most common AI systems do not save copies of existing works, but merely “instructions” (more or less) on how to create new works that conform to patterns they found by examining existing work. If we assume that a system isn’t violating copyright at the input stage, it’s entirely possible that it can produce completely new pieces of art that have never before existed and do not violate copyright.
They can, however, be made to violate IP rights. For example, trademark violations appear to be one of the most popular uses of these AI systems by end users. To take but one example, a quick search of Google Images for “midjourney iron man” returns a slew of images that almost certainly violate trademarks for the character Iron Man. Similarly, these systems can be instructed to generate art that is not just “in the style” of a particular artist, but that very closely resembles existing pieces. In this sense, the system would be making a copy that theoretically infringes.
There is a common bug in such systems that leads to outputs that are more likely to violate copyright in this way. Known as “overfitting,” the training leg of these AI systems can be presented with samples that contain too many instances of a particular image. This leads to a data set that contains too much information about the specific image, such that when the AI generates a new image, it is constrained to producing something very close to the original.
An argument can also be made that generating art “in the style of” a famous artist violates moral rights (in jurisdictions where such rights exist).
At least in the copyright space, cases like Sony are going to become crucial. Does the user side of these AI systems have substantial noninfringing uses? If so, the firms that host software for end users could avoid secondary-infringement liability, and the onus would fall on users to avoid violating copyright laws. At the same time, it seems plausible that legislatures could place some obligation on these providers to implement filters to mitigate infringement by end users.
Opportunities for New IP Commercialization with AI
There are a number of ways that AI systems may inexcusably infringe on intellectual-property rights. As a best practice, I would encourage the firms that operate these services to seek licenses from rightsholders. While this would surely be an expense, it also opens new opportunities for both sides to generate revenue.
For example, an AI firm could develop its own version of YouTube’s ContentID that allows creators to opt their work into training. For some well-known artists this could be negotiated with an upfront licensing fee. On the user-side, any artist who has opted in could then be selected as a “style” for the AI to emulate. When users generate an image, a royalty payment to the artist would be created. Creators would also have the option to remove their influence from the system if they so desired.
Undoubtedly, there are other ways to monetize the relationship between creators and the use of their work in AI systems. Ultimately, the firms that run these systems will not be able to simply wish away IP laws. There are going to be opportunities for creators and AI firms to both succeed, and the law should help to generate that result.
In our previous post on Gonzalez v. Google LLC, which will come before the U.S. Supreme Court for oral arguments Feb. 21, Kristian Stout and I argued that, while the U.S. Justice Department (DOJ) got the general analysis right (looking to Roommates.com as the framework for exceptions to the general protections of Section 230), they got the application wrong (saying that algorithmic recommendations should be excepted from immunity).
Now, after reading Google’s brief, as well as the briefs of amici on their side, it is even more clear to me that:
algorithmic recommendations are protected by Section 230 immunity; and
creating an exception for such algorithms would severely damage the internet as we know it.
I address these points in reverse order below.
Google on the Death of the Internet Without Algorithms
The central point that Google makes throughout its brief is that a finding that Section 230’s immunity does not extend to the use of algorithmic recommendations would have potentially catastrophic implications for the internet economy. Google and amici for respondents emphasize the ubiquity of recommendation algorithms:
Recommendation algorithms are what make it possible to find the needles in humanity’s largest haystack. The result of these algorithms is unprecedented access to knowledge, from the lifesaving (“how to perform CPR”) to the mundane (“best pizza near me”). Google Search uses algorithms to recommend top search results. YouTube uses algorithms to share everything from cat videos to Heimlich-maneuver tutorials, algebra problem-solving guides, and opera performances. Services from Yelp to Etsy use algorithms to organize millions of user reviews and ratings, fueling global commerce. And individual users “like” and “share” content millions of times every day. – Brief for Respondent Google, LLC at 2.
The “recommendations” they challenge are implicit, based simply on the manner in which YouTube organizes and displays the multitude of third-party content on its site to help users identify content that is of likely interest to them. But it is impossible to operate an online service without “recommending” content in that sense, just as it is impossible to edit an anthology without “recommending” the story that comes first in the volume. Indeed, since the dawn of the internet, virtually every online service—from news, e-commerce, travel, weather, finance, politics, entertainment, cooking, and sports sites, to government, reference, and educational sites, along with search engines—has had to highlight certain content among the thousands or millions of articles, photographs, videos, reviews, or comments it hosts to help users identify what may be most relevant. Given the sheer volume of content on the internet, efforts to organize, rank, and display content in ways that are useful and attractive to users are indispensable. As a result, exposing online services to liability for the “recommendations” inherent in those organizational choices would expose them to liability for third-party content virtually all the time. – Amicus Brief for Meta Platforms at 3-4.
In other words, if Section 230 were limited in the way that the plaintiffs (and the DOJ) seek, internet platforms’ ability to offer users useful information would be strongly attenuated, if not completely impaired. The resulting legal exposure would lead inexorably to far less of the kinds of algorithmic recommendations upon which the modern internet is built.
This is, in part, why we weren’t able to fully endorse the DOJ’s brief in our previous post. The DOJ’s brief simply goes too far. It would be unreasonable to establish as a categorical rule that use of the ubiquitous auto-discovery algorithms that power so much of the internet would strip a platform of Section 230 protection. The general rule advanced by the DOJ’s brief would have detrimental and far-ranging implications.
Amici on Publishing and Section 230(f)(4)
Google and the amici also make a strong case that algorithmic recommendations are inseparable from publishing. They have a strong textual hook in Section 230(f)(4), which explicitly protects “enabling tools that… filter, screen, allow, or disallow content; pick, choose, analyze or disallow content; or transmit, receive, display, forward, cache, search, subset, organize, reorganize, or translate content.”
As the amicus brief from a group of internet-law scholars—including my International Center for Law & Economics colleagues Geoffrey Manne and Gus Hurwitz—put it:
Section 230’s text should decide this case. Section 230(c)(1) immunizes the user or provider of an “interactive computer service” from being “treated as the publisher or speaker” of information “provided by another information content provider.” And, as Section 230(f)’s definitions make clear, Congress understood the term “interactive computer service” to include services that “filter,” “screen,” “pick, choose, analyze,” “display, search, subset, organize,” or “reorganize” third-party content. Automated recommendations perform exactly those functions, and are therefore within the express scope of Section 230’s text. – Amicus Brief of Internet Law Scholars at 3-4.
In other words, Section 230 protects not just the conveyance of information, but how that information is displayed. Algorithmic recommendations are a subset of those display tools that allow users to find what they are looking for with ease. Section 230 can’t be reasonably read to exclude them.
Why This Isn’t Really (Just) a Roommates.com Case
This is where the DOJ’s amicus brief (and our previous analysis) misses the point. This is not strictly a Roomates.com case. The case actually turns on whether algorithmic recommendations are separable from publication of third-party content, rather than whether they are design choices akin to what was occurring in that case.
For instance, in our previous post, we argued that:
[T]he DOJ argument then moves onto thinner ice. The DOJ believes that the 230 liability shield in Gonzalez depends on whether an automated “recommendation” rises to the level of development or creation, as the design of filtering criteria in Roommates.com did.
While we thought the DOJ went too far in differentiating algorithmic recommendations from other uses of algorithms, we gave them too much credit in applying the Roomates.com analysis. Section 230 was meant to immunize filtering tools, so long as the information provided is from third parties. Algorithmic recommendations—like the type at issue with YouTube’s “Up Next” feature—are less like the conduct in Roommates.com and much more like a search engine.
The DOJ did, however, have a point regarding algorithmic tools in that they may—like any other tool a platform might use—be employed in a way that transforms the automated promotion into a direct endorsement or original publication. For instance, it’s possible to use algorithms to intentionally amplify certain kinds of content in such a way as to cultivate more of that content.
That’s, after all, what was at the heart of Roommates.com. The site was designed to elicit responses from users that violated the law. Algorithms can do that, but as we observed previously, and as the many amici in Gonzalez observe, there is nothing inherent to the operation of algorithms that match users with content that makes their use categorically incompatible with Section 230’s protections.
Conclusion
After looking at the textual and policy arguments forwarded by both sides in Gonzalez, it appears that Google and amici for respondents have the better of it. As several amici argued, to the extent there are good reasons to reform Section 230, Congress should take the lead. The Supreme Court shouldn’t take this case as an opportunity to significantly change the consensus of the appellate courts on the broad protections of Section 230 immunity.
Later next month, the U.S. Supreme Court will hear oral arguments in Gonzalez v. Google LLC, a case that has drawn significant attention and many bad takes regarding how Section 230 of the Communications Decency Act should be interpreted. Enacted in the mid-1990s, when the Internet as we know it was still in its infancy, Section 230 has grown into a law that offers online platforms a fairly comprehensive shield against liability for the content that third parties post to their services. But the law has also come increasingly under fire, from both the political left and the right.
At issue in Gonzalez is whether Section 230(c)(1) immunizes Google from a set of claims brought under the Antiterrorism Act of 1990 (ATA). The petitioners are relatives of Nohemi Gonzalez, an American citizen murdered in a 2015 terrorist attack in Paris. They allege that Google, through YouTube, is liable under the ATA for providing assistance to ISIS for four main reasons. They allege that:
Google allowed ISIS to use YouTube to disseminate videos and messages, thereby recruiting and radicalizing terrorists responsible for the murder.
Google failed to take adequate steps to take down videos and accounts and keep them down.
Google recommends videos of others, both through subscriptions and algorithms.
Google monetizes this content through its AdSense service, with ISIS-affiliated users receiving revenue.
The 9th U.S. Circuit Court of Appeals dismissed all of the non-revenue-sharing claims as barred by Section 230(c)(1), but allowed the revenue-sharing claim to go forward.
Highlights of DOJ’s Brief
In an amicus brief, the U.S. Justice Department (DOJ) ultimately asks the Court to vacate the 9th Circuit’s judgment regarding those claims that are based on YouTube’s alleged targeted recommendations of ISIS content. But the DOJ also rejects much of the petitioner’s brief, arguing that Section 230 does rightfully apply to the rest of the claims.
As the DOJ notes, radical theories advanced by the plaintiffs and other amici would go too far in restricting Section 230 immunity based on a platform’s decisions on whether or not to block or remove user content (see, e.g., its discussion on pp. 17-21 of the merits and demerits of Justice Clarence Thomas’s Malwarebytes concurrence).
At the same time, the DOJ’s brief notes that there is room for a reasonable interpretation of Section 230 that allows for liability to attach when online platforms behave unreasonably in their promotion of users’ content. Applying essentially the 9th Circuit’s Roommates.com standard, the DOJ argues that YouTube’s choice to amplify certain terrorist content through its recommendations algorithm is a design choice, rather than simply the hosting of third-party content, thereby removing it from the scope of Section 230 immunity.
While there is much to be said in favor of this approach, it’s important to point out that, although directionally correct, it’s not at all clear that a Roommates.com analysis should ultimately come down as the DOJ recommends in Gonzalez. More broadly, the way the DOJ structures its analysis has important implications for how we should think about the scope of Section 230 reform that attempts to balance accountability for intermediaries with avoiding undue collateral censorship.
Charting a Middle Course on Immunity
The important point on which the DOJ relies from Roommates.com is that intermediaries can be held accountable when their own conduct creates violations of the law, even if it involves third–party content. As the DOJ brief puts it:
Section 230(c)(1) protects an online platform from claims premised on its dissemination of third-party speech, but the statute does not immunize a platform’s other conduct, even if that conduct involves the solicitation or presentation of third-party content. The Ninth Circuit’s Roommates.com decision illustrates the point in the context of a website offering a roommate-matching service… As a condition of using the service, Roommates.com “require[d] each subscriber to disclose his sex, sexual orientation and whether he would bring children to a household,” and to “describe his preferences in roommates with respect to the same three criteria.” Ibid. The plaintiffs alleged that asking those questions violated housing-discrimination laws, and the court of appeals agreed that Section 230(c)(1) did not shield Roommates.com from liability for its “own acts” of “posting the questionnaire and requiring answers to it.” Id. at 1165.
Imposing liability in such circumstances does not treat online platforms as the publishers or speakers of content provided by others. Nor does it obligate them to monitor their platforms to detect objectionable postings, or compel them to choose between “suppressing controversial speech or sustaining prohibitive liability.”… Illustrating that distinction, the Roommates.com court held that although Section 230(c)(1) did not apply to the website’s discriminatory questions, it did shield the website from liability for any discriminatory third-party content that users unilaterally chose to post on the site’s “generic” “Additional Comments” section…
The DOJ proceeds from this basis to analyze what it would take for Google (via YouTube) to no longer benefit from Section 230 immunity by virtue of its own editorial actions, as opposed to its actions as a publisher (which 230 would still protect). For instance, are the algorithmic suggestions of videos simply neutral tools that allow for users to get more of the content they desire, akin to search results? Or are the algorithmic suggestions of new videos a design choice that makes it akin to Roommates?
The DOJ argues that taking steps to better display pre-existing content is not content development or creation, in and of itself. Similarly, it would be a mistake to make intermediaries liable for creating tools that can then be deployed by users:
Interactive websites invariably provide tools that enable users to create, and other users to find and engage with, information. A chatroom might supply topic headings to organize posts; a photo-sharing site might offer a feature for users to signal that they like or dislike a post; a classifieds website might enable users to add photos or maps to their listings. If such features rendered the website a co-developer of all users’ content, Section 230(c)(1) would be a dead letter.
At a high level, this is correct. Unfortunately, the DOJ argument then moves onto thinner ice. The DOJ believes that the 230 liability shield in Gonzalez depends on whether an automated “recommendation” rises to the level of development or creation, as the design of filtering criteria in Roommates.com did. Toward this end, the brief notes that:
The distinction between a recommendation and the recommended content is particularly clear when the recommendation is explicit. If YouTube had placed a selected ISIS video on a user’s homepage alongside a message stating, “You should watch this,” that message would fall outside Section 230(c)(1). Encouraging a user to watch a selected video is conduct distinct from the video’s publication (i.e., hosting). And while YouTube would be the “publisher” of the recommendation message itself, that message would not be “information provided by another information content provider.” 47 U.S.C. 230(c)(1).
An Absence of Immunity Does Not Mean a Presence of Liability
Importantly, the DOJ brief emphasizes throughout that remanding the ATA claims is not the end of the analysis—i.e., it does not mean that the plaintiffs can prove the elements. Moreover, other background law—notably, the First Amendment—can limit the application of liability to intermediaries, as well. As we put it in our paper on Section 230 reform:
It is important to again note that our reasonableness proposal doesn’t change the fact that the underlying elements in any cause of action still need to be proven. It is those underlying laws, whether civil or criminal, that would possibly hold intermediaries liable without Section 230 immunity. Thus, for example, those who complain that FOSTA/SESTA harmed sex workers by foreclosing a safe way for them to transact (illegal) business should really be focused on the underlying laws that make sex work illegal, not the exception to Section 230 immunity that FOSTA/SESTA represents. By the same token, those who assert that Section 230 improperly immunizes “conservative bias” or “misinformation” fail to recognize that, because neither of those is actually illegal (nor could they be under First Amendment law), Section 230 offers no additional immunity from liability for such conduct: There is no underlying liability from which to provide immunity in the first place.
There’s a strong likelihood that, on remand, the court will find there is no violation of the ATA at all. Section 230 immunity need not be stretched beyond all reasonable limits to protect intermediaries from hypothetical harms when underlying laws often don’t apply.
Conclusion
To date, the contours of Section 230 reform largely have been determined by how courts interpret the statute. There is an emerging consensus that some courts have gone too far in extending Section 230 immunity to intermediaries. The DOJ’s brief is directionally correct, but the Court should not adopt it wholesale. More needs to be done to ensure that the particular facts of Gonzalez are not used to completely gut Section 230 more generally.
The €390 million fine that the Irish Data Protection Commission (DPC) levied last week against Meta marks both the latest skirmish in the ongoing regulatory war on the use of data by private firms, as well as a major blow to the ad-driven business model that underlies most online services.
More specifically, the DPC was forced by the European Data Protection Board (EDPB) to find that Meta violated the General Data Protection Regulation (GDPR) when it relied on its contractual relationship with Facebook and Instagram users as the basis to employ user data in personalized advertising.
Meta still has other bases on which it can argue it relies in order to make use of user data, but a larger issue is at-play: the decision’s findings both that making use of user data for personalized advertising is not “necessary” between a service and its users and that privacy regulators are in a position to make such an assessment.
More broadly, the case also underscores that there is no consensus within the European Union on the broad interpretation of the GDPR preferred by some national regulators and the EDPB.
The DPC Decision
The core disagreement between the DPC and Meta, on the one hand, and some other EU privacy regulators, on the other, is whether it is lawful for Meta to treat the use of user data for personalized advertising as “necessary for the performance of” the contract between Meta and its users. The Irish DPC accepted Meta’s arguments that the nature of Facebook and Instagram is such that it is necessary to process personal data this way. The EDPB took the opposite approach and used its powers under the GDPR to direct the DPC to issue a decision contrary to DPC’s own determination. Notably, the DPC announced that it is considering challenging the EDPB’s involvement before the EU Court of Justice as an unlawful overreach of the board’s powers.
In the EDPB’s view, it is possible for Meta to offer Facebook and Instagram without personalized advertising. And to the extent that this is possible, Meta cannot rely on the “necessity for the performance of a contract” basis for data processing under Article 6 of the GDPR. Instead, Meta in most cases should rely on the “consent” basis, involving an explicit “yes/no” choice. In other words, Facebook and Instagram users should be explicitly asked if they consent to their data being used for personalized advertising. If they decline, then under this rationale, they would be free to continue using the service without personalized advertising (but with, e.g., contextual advertising).
Notably, the decision does not mandate a particular contractual basis for processing, but only invalidates “contractual necessity” for personalized advertising. Indeed, Meta believes it has other avenues for continuing to process user data for personalized advertising while not depending on a “consent” basis. Of course, only time will tell if this reasoning is accepted. Nonetheless, the EDBP’s underlying animus toward the “necessity” of personalized advertising remains concerning.
What Is ‘Necessary’ for a Service?
The EDPB’s position is of a piece with a growing campaign against firms’ use of data more generally. But as in similar complaints against data use, the demonstrated harms here are overstated, while the possibility that benefits might flow from the use of data is assumed to be zero.
How does the EDPB know that it is not necessary for Meta to rely on personalized advertising? And what does “necessity” mean in this context? According to the EDPB’s own guidelines, a business “should be able to demonstrate how the main subject-matter of the specific contract with the data subject cannot, as a matter of fact, be performed if the specific processing of the personal data in question does not occur.” Therefore, if it is possible to distinguish various “elements of a service that can in fact reasonably be performed independently of one another,” then even if some processing of personal data is necessary for some elements, this cannot be used to bundle those with other elements and create a “take it or leave it” situation for users. The EDPB stressed that:
This assessment may reveal that certain processing activities are not necessary for the individual services requested by the data subject, but rather necessary for the controller’s wider business model.
This stilted view of what counts as a “service” completely fails to acknowledge that “necessary” must mean more than merely technologically possible. Any service offering faces both technical limitations as well as economic limitations. What is technically possible to offer can also be so uneconomic in some forms as to be practically impossible. Surely, there are alternatives to personalized advertising as a means to monetize social media, but determining what those are requires a great deal of careful analysis and experimentation. Moreover, the EDPB’s suggested “contextual advertising” alternative is not obviously superior to the status quo, nor has it been demonstrated to be economically viable at scale.
Thus, even though it does not strictly follow from the guidelines, the decision in the Meta case suggests that, in practice, the EDPB pays little attention to the economic reality of a contractual relationship between service providers and their users, instead trying to carve out an artificial, formalistic approach. It is doubtful whether the EDPB engaged in the kind of robust economic analysis of Facebook and Instagram that would allow it to reach a conclusion as to whether those services are economically viable without the use of personalized advertising.
However, there is a key institutional point to be made here. Privacy regulators are likely to be eminently unprepared to conduct this kind of analysis, which arguably should lead to significant deference to the observed choices of businesses and their customers.
Conclusion
A service’s use of its users’ personal data—whether for personalized advertising or other purposes—can be a problem, but it can also generate benefits. There is no shortcut to determine, in any given situation, whether the costs of a particular business model outweigh its benefits. Critically, the balance of costs and benefits from a business model’s technological and economic components is what truly determines whether any specific component is “necessary.” In the Meta decision, the EDPB got it wrong by refusing to incorporate the full economic and technological components of the company’s business model.
The blistering pace at which the European Union put forward and adopted the Digital Markets Act (DMA) has attracted the attention of legislators across the globe. In its wake, countries such as South Africa, India, Brazil, and Turkey have all contemplated digital-market regulations inspired by the DMA (and other models of regulation, such as the United Kingdom’s Digital Markets Unit and Australia’s sectoral codes of conduct).
Racing to be among the first jurisdictions to regulate might intuitively seem like a good idea. By emulating the EU, countries could hope to be perceived as on the cutting edge of competition policy, and hopefully earn a seat at the table when the future direction of such regulations is discussed.
There are, however, tradeoffs involved in regulating digital markets, which are arguably even more salient in the case of emerging markets. Indeed, as we will explain here, these jurisdictions often face challenges that significantly alter the ratio of costs and benefits when it comes to enacting regulation.
Drawing from a paper we wrote with Sam Bowman about competition policy in the Association of Southeast Asian Nations (ASEAN) zone, we highlight below three of the biggest issues these initiatives face.
To Regulate Competition, You First Need to Attract Competition
Perhaps the biggest factor cautioning emerging markets against adoption of DMA-inspired regulations is that such rules would impose heavy compliance costs to doing business in markets that are often anything but mature. It is probably fair to say that, in many (maybe most) emerging markets, the most pressing challenge is to attract investment from international tech firms in the first place, not how to regulate their conduct.
The most salient example comes from South Africa, which has sketched out plans to regulate digital markets. The Competition Commission has announced that Amazon, which is not yet available in the country, would fall under these new rules should it decide to enter—essentially on the presumption that Amazon would overthrow South Africa’s incumbent firms.
It goes without saying that, at the margin, such plans reduce either the likelihood that Amazon will enter the South African market at all, or the extent of its entry should it choose to do so. South African consumers thus risk losing the vast benefits such entry would bring—benefits that dwarf those from whatever marginal increase in competition might be gained from subjecting Amazon to onerous digital-market regulations.
While other tech firms—such as Alphabet, Meta, and Apple—are already active in most emerging jurisdictions, regulation might still have a similar deterrent effect to their further investment. Indeed, the infrastructure deployed by big tech firms in these jurisdictions is nowhere near as extensive as in Western countries. To put it mildly, emerging-market consumers typically only have access to slower versions of these firms’ services. A quick glimpse at Google Cloud’s global content-delivery network illustrates this point well (i.e., that there is far less infrastructure in developing markets):
Ultimately, emerging markets remain relatively underserved compared to those in the West. In such markets, the priority should be to attract tech investment, not to impose regulations that may further slow the deployment of critical internet infrastructure.
Growth Is Key
The potential to boost growth is the most persuasive argument for emerging markets to favor a more restrained approach to competition law and regulation, such as that currently employed in the United States.
Emerging nations may not have the means (or the inclination) to equip digital-market enforcers with resources similar to those of the European Commission. Given these resource constraints, it is essential that such jurisdictions focus their enforcement efforts on those areas that provide the highest return on investment, notably in terms of increased innovation.
This raises an important point. A recent empirical study by Ross Levine, Chen Lin, Lai Wei, and Wensi Xie finds that competition enforcement does, indeed, promote innovation. But among the study’s more surprising findings is that, unlike other areas of competition enforcement, the strength of a jurisdiction’s enforcement of “abuse of dominance” rules does not correlate with increased innovation. Furthermore, jurisdictions that allow for so-called “efficiency defenses” in unilateral-conduct cases also tend to produce more innovation. The authors thus conclude that:
From the perspective of maximizing patent-based innovation, therefore, a legal system that allows firms to exploit their dominant positions based on efficiency considerations could boost innovation.
These findings should give pause to policymakers who seek to emulate the European Union’s DMA—which, among other things, does not allow gatekeepers to put forward so-called “efficiency defenses” that would allow them to demonstrate that their behavior benefits consumers. If growth and innovation are harmed by overinclusive abuse-of-dominance regimes and rules that preclude firms from offering efficiency-based defenses, then this is probably even more true of digital-market regulations that replace case-by-case competition enforcement with per se prohibitions.
In short, the available evidence suggests that, faced with limited enforcement resources, emerging-market jurisdictions should prioritize other areas of competition policy, such as breaking up or mitigating the harmful effects of cartels and exercising appropriate merger controls.
These findings also cut in favor of emphasizing the traditional antitrust goal of maximizing consumer welfare—or, at least, protecting the competitive process. Many of the more recent digital-market regulations—such as the DMA, the UK DMU, and the ACCC sectoral codes of conduct—are instead focused on distributional issues. They seek to ensure that platform users earn a “fair share” of the benefits generated on a platform. In light of Levine et al.’s findings, this approach could be undesirable, as using competition policy to reduce monopoly rents may lead to less innovation.
In short, traditional antitrust law’s focus on consumer welfare and relatively limited enforcement in the area of unilateral conduct may be a good match for emerging nations that want competition regimes that maximize innovation under important resource constraints.
Consider Local Economic and Political Conditions
Emerging jurisdictions have diverse economic and political profiles. These features, in turn, affect the respective costs and benefits of digital-market regulations.
For example, digital-market regulations generally offer very broad discretion to competition enforcers. The DMA details dozens of open-ended prohibitions upon which enforcers can base infringement proceedings. Furthermore, because they are designed to make enforcers’ task easier, these regulations often remove protections traditionally afforded to defendants, such as appeals to the consumer welfare standard or efficiency defenses. The UK’s DMU initiative, for example, would lower the standard of proof that enforcers must meet.
Giving authorities broad powers with limited judicial oversight might be less problematic in jurisdictions where the state has a track record of self-restraint. The consequences of regulatory discretion might, however, be far more problematic in jurisdictions where authorities routinely overstep the mark and where the threat of corruption is very real.
To name but two, countries like South Africa and India rank relatively low in the World Bank’s “ease of doing business index” (84th and 62nd, respectively). They also rank relatively low on the Cato Institute’s “human freedom index” (77th and 119th, respectively—and both score particularly badly in terms of economic freedom). This suggests strongly that authorities in those jurisdictions are prone to misapply powers derived from digital-market regulations in ways that hurt growth and consumers.
To make matters worse, outright corruption is also a real problem in several emerging nations. Returning to South Africa and India, both jurisdictions face significant corruption issues (they rank 70th and 85th, respectively, on Transparency International’s “Corruption Perception Index”).
At a more granular level, an inquiry in South Africa revealed rampant corruption under former President Jacob Zuma, while current President Cyril Ramaphosa also faces significant corruption allegations. Writing in the Financial Times in 2018, Gaurav Dalmia—chair of Delhi-based Dalmia Group Holdings—opined that “India’s anti-corruption battle will take decades to win.”
This specter of corruption thus counsels in favor of establishing competition regimes with sufficient checks and balances, so as to prevent competition authorities from being captured by industry or political forces. But most digital-market regulations are designed precisely to remove those protections in order to streamline enforcement. The risk that they could be mobilized toward nefarious ends are thus anything but trivial. This is of particular concern, given that such regulations are typically mobilized against global firms in order to shield inefficient local firms—raising serious risks of protectionist enforcement that would harm local consumers.
Conclusion
The bottom line is that emerging markets would do well to reconsider the value of regulating digital markets that have yet to reach full maturity. Recent proposals threaten to deter tech investments in these jurisdictions, while raising significant risks of reduced growth, corruption, and consumer-harming protectionism.
Twitter has seen a lot of ups and downs since Elon Musk closed on his acquisition of the company in late October and almost immediately set about his initiatives to “reform” the platform’s operations.
One of the stories that has gotten somewhat lost in the ensuing chaos is that, in the short time under Musk, Twitter has made significant inroads—on at least some margins—against the visibility of child sexual abuse material (CSAM) by removing major hashtags that were used to share it, creating a direct reporting option, and removing major purveyors. On the other hand, due to the large reductions in Twitter’s workforce—both voluntary and involuntary—there are now very few human reviewers left to deal with the issue.
Section 230 immunity currently protects online intermediaries from most civil suits for CSAM (a narrow carveout is made under Section 1595 of the Trafficking Victims Protection Act). While the federal government could bring criminal charges if it believes online intermediaries are violating federal CSAM laws, and certain narrow state criminal claims could be brought consistent with federal law, private litigants are largely left without the ability to find redress on their own in the courts.
This, among other reasons, is why there has been a push to amend Section 230 immunity. Our proposal (along with co-author Geoffrey Manne) suggests online intermediaries should have a reasonable duty of care to remove illegal content. But this still requires thinking carefully about what a reasonable duty of care entails.
For instance, one of the big splash moves made by Twitter after Musk’s acquisition was to remove major CSAM distribution hashtags. While this did limit visibility of CSAM for a time, some experts say it doesn’t really solve the problem, as new hashtags will arise. So, would a reasonableness standard require the periodic removal of major hashtags? Perhaps it would. It appears to have been a relatively low-cost way to reduce access to such material, and could theoretically be incorporated into a larger program that uses automated discovery to find and remove future hashtags.
Of course it won’t be perfect, and will be subject to something of a Whac-A-Mole dynamic. But the relevant question isn’t whether it’s a perfect solution, but whether it yields significant benefit relative to its cost, such that it should be regarded as a legally reasonable measure that platforms should broadly implement.
On the flip side, Twitter has lost such a large amount of its workforce that it potentially no longer has enough staff to do the important review of CSAM. As long as Twitter allows adult nudity, and algorithms are unable to effectively distinguish between different types of nudity, human reviewers remain essential. A reasonableness standard might also require sufficient staff and funding dedicated to reviewing posts for CSAM.
But what does it mean for a platform to behave “reasonably”?
Platforms Should Behave ‘Reasonably’
Rethinking platforms’ safe harbor from liability as governed by a “reasonableness” standard offers a way to more effectively navigate the complexities of these tradeoffs without resorting to the binary of immunity or total liability that typically characterizes discussions of Section 230 reform.
It could be the case that, given the reality that machines can’t distinguish between “good” and “bad” nudity, it is patently unreasonable for an open platform to allow any nudity at all if it is run with the level of staffing that Musk seems to prefer for Twitter.
Consider the situation that MindGeek faced a couple of years ago. It was pressured by financial providers, including PayPal and Visa, to clean up the CSAM and nonconsenual pornography that appeared on its websites. In response, they removed more than 80% of suspected illicit content and required greater authentication for posting.
Notwithstanding efforts to clean up the service, a lawsuit was filed against MindGeek and Visa by victims who asserted that the credit-card company was a knowing conspirator for processing payments to MindGeek’s sites when they were purveying child pornography. Notably, Section 230 issues were dismissed early on in the case, but the remaining claims—rooted in the Racketeer Influenced and Corrupt Organizations Act (RICO) and the Trafficking Victims Protection Act (TVPA)—contained elements that support evaluating the conduct of online intermediaries, including payment providers who support online services, through a reasonableness lens.
In our amicus, we stressed the broader policy implications of failing to appropriately demarcate the bounds of liability. In short, we stressed that deterrence is best encouraged by placing responsibility for control on the party most closely able to monitor the situation—i.e., MindGeek, and not Visa. Underlying this, we believe that an appropriately tuned reasonableness standard should be able to foreclose these sorts of inquiries at early stages of litigation if there is good evidence that an intermediary behaved reasonably under the circumstances.
In this case, we believed the court should have taken seriously the fact that a payment processor needs to balance a number of competing demands— legally, economically, and morally—in a way that enables them to serve their necessary prosocial roles. Here, Visa had to balance its role, on the one hand, as a neutral intermediary responsible for handling millions of daily transactions, with its interests to ensure that it did not facilitate illegal behavior. But it also was operating, essentially, under a veil of ignorance: all of the information it had was derived from news reports, as it was not directly involved in, nor did it have special insight into, the operation of MindGeek’s businesses.
As we stressed in our intermediary-liability paper, there is indeed a valid concern that changes to intermediary-liability policy not invite a flood of ruinous litigation. Instead, there needs to be some ability to determine at the early stages of litigation whether a defendant behaved reasonably under the circumstances. In the MindGeek case, we believed that Visa did.
In essence, much of this approach to intermediary liability boils down to finding socially and economically efficient dividing lines that can broadly demarcate when liability should attach. For example, if Visa is liable as a co-conspirator in MindGeek’s allegedly illegal enterprise for providing a payment network that MindGeek uses by virtue of its relationship with yet other intermediaries (i.e., the banks that actually accept and process the credit-card payments), why isn’t the U.S. Post Office also liable for providing package-delivery services that allow MindGeek to operate? Or its maintenance contractor for cleaning and maintaining its offices?
Twitter implicitly engaged in this sort of analysis when it considered becoming an OnlyFans competitor. Despite having considerable resources—both algorithmic and human—Twitter’s internal team determined they could not “accurately detect child sexual exploitation and non-consensual nudity at scale.” As a result, they abandoned the project. Similarly, Tumblr tried to make many changes, including taking down CSAM hashtags, before finally giving up and removing all pornographic material in order to remain in the App Store for iOS. At root, these firms demonstrated the ability to weigh costs and benefits in ways entirely consistent with a reasonableness analysis.
Thinking about the MindGeek situation again, it could also be the case that MindGeek did not behave reasonably. Some of MindGeek’s sites encouraged the upload of user-generated pornography. If MindGeek experienced the same limitations in detecting “good” and “bad” pornography (which is likely), it could be that the company behaved recklessly for many years, and only tightened its verification procedures once it was caught. If true, that is behavior that should not be protected by the law with a liability shield, as it is patently unreasonable.
Apple is sometimes derided as an unfair gatekeeper of speech through its App Store. But, ironically, Apple itself has made complex tradeoffs between data security and privacy—through use of encryption, on the one hand, and checking devices for CSAM material, on the other. Prioritizing encryption over scanning devices (especially photos and messages) for CSAM is a choice that could allow for more CSAM to proliferate. But the choice is, again, a difficult one: how much moderation is needed and how do you balance such costs against other values important to users, such as privacy for the vast majority of nonoffending users?
As always, these issues are complex and involve tradeoffs. But it is obvious that more can and needs to be done by online intermediaries to remove CSAM.
But What Is ‘Reasonable’? And How Do We Get There?
The million-dollar legal question is what counts as “reasonable?” We are not unaware of the fact that, particularly when dealing with online platforms that deal with millions of users a day, there is a great deal of surface area exposed to litigation by potentially illicit user-generated conduct. Thus, it is not the case, at least for the foreseeable future, that we need to throw open gates of a full-blown common-law process to determine questions of intermediary liability. What is needed, instead, is a phased-in approach that gets courts in the business of parsing these hard questions and building up a body of principles that, on the one hand, encourage platforms to do more to control illicit content on their services, and on the other, discourages unmeritorious lawsuits by the plaintiffs’ bar.
One of our proposals for Section 230 reform is for a multistakeholder body, overseen by an expert agency like the Federal Trade Commission or National Institute of Standards and Technology, to create certified moderation policies. This would involve online intermediaries working together with a convening federal expert agency to develop a set of best practices for removing CSAM, including thinking through the cost-benefit analysis of more moderation—human or algorithmic—or even wholesale removal of nudity and pornographic content.
Compliance with these standards should, in most cases, operate to foreclose litigation against online service providers at an early stage. If such best practices are followed, a defendant could point to its moderation policies as a “certified answer” to any complaint alleging a cause of action arising out of user-generated content. Compliant practices will merit dismissal of the case, effecting a safe harbor similar to the one currently in place in Section 230.
In litigation, after a defendant answers a complaint with its certified moderation policies, the burden would shift to the plaintiff to adduce sufficient evidence to show that the certified standards were not actually adhered to. Such evidence should be more than mere res ipsa loquitur; it must be sufficient to demonstrate that the online service provider should have been aware of a harm or potential harm, that it had the opportunity to cure or prevent it, and that it failed to do so. Such a claim would need to meet a heightened pleading requirement, as for fraud, requiring particularity. And, periodically, the body overseeing the development of this process would incorporate changes to the best practices standards based on the cases being brought in front of courts.
Online service providers don’t need to be perfect in their content-moderation decisions, but they should behave reasonably. A properly designed duty-of-care standard should be flexible and account for a platform’s scale, the nature and size of its user base, and the costs of compliance, among other considerations. What is appropriate for YouTube, Facebook, or Twitter may not be the same as what’s appropriate for a startup social-media site, a web-infrastructure provider, or an e-commerce platform.
Indeed, this sort of flexibility is a benefit of adopting a “reasonableness” standard, such as is found in common-law negligence. Allowing courts to apply the flexible common-law duty of reasonable care would also enable jurisprudence to evolve with the changing nature of online intermediaries, the problems they pose, and the moderating technologies that become available.
Conclusion
Twitter and other online intermediaries continue to struggle with the best approach to removing CSAM, nonconsensual pornography, and a whole host of other illicit content. There are no easy answers, but there are strong ethical reasons, as well as legal and market pressures, to do more. Section 230 reform is just one part of a complete regulatory framework, but it is an important part of getting intermediary liability incentives right. A reasonableness approach that would hold online platforms accountable in a cost-beneficial way is likely to be a key part of a positive reform agenda for Section 230.
RSS - Posts