Campos & Laurent on A Definition of General-Purpose AI Systems

Simeon Campos (SaferAI) and Romain Laurent (same) have posted “A Definition of General-Purpose AI Systems: Mitigating Risks from the Most Generally Capable Models” on SSRN. Here is the abstract:

The European Union (EU) is currently going through the legislative process on the EU AI Act – the first bill intended to regulate Artificial Intelligence (AI) comprehensively in a major jurisdiction. The bill includes provisions to manage risks of generally capable AIs classified as “General Purpose AI Systems” (GPAIS). We believe that this crucial aspect of the act could be improved by focusing the definition more on the most generally capable systems, which bring very specific risks. The Future of Life Institute (FLI) proposed a definition of GPAIS to better target these models, a significant step in the right direction. Expanding on FLI’s proposal, this paper introduces a new definition of GPAIS, which serves to clearly differentiate between narrow and general systems, and cannot be easily exploited by GPAIS providers who may wish to avoid new regulatory constraints.

This paper consists of two sections. The first section discusses the specific risks of GPAIS, including unpredictability, adaptability, and the potential for emergent capabilities. The second section presents the new definition of GPAIS, and explains the changes made and how they address the risks presented in the first section. The EU AI Act could set a global standard for AI-related risk management. The aim of this document is to help inform AI Act draft reviews and improve the ability to mitigate risks from the most generally capable models to protect stakeholders in the EU and globally.

Gaske on Regulation Priorities for Artificial Intelligence Foundation Models

Matthew Gaske has posted “Regulation Priorities for Artificial Intelligence Foundation Models” (26 VAND. J. ENT. & TECH. L. 1, Forthcoming (2023)) on SSRN. Here is the abstract:

This Article responds to the call in techlaw literature for high-level frameworks to guide regulation of the development and use of Artificial Intelligence (AI) technologies. Accordingly, it adapts a generalized form of the fintech Innovation Trilemma framework to argue that a regulator can prioritize only two of three aims when considering AI oversight: 1) promoting innovation, 2) mitigating systemic risk, and 3) providing clear regulatory requirements. Specifically, this Article expressly connects legal scholarship to research in other fields focusing on “foundation model” AI and explores this kind of system’s implications for regulation priorities from the geopolitical and commercial competitive contexts. These models are so named because they have a novel ability to easily apply their resources across a broad variety of use cases, unlike prior AI technologies. These systems, such as OpenAI’s ChatGPT or Alphabet’s LaMDA, have recently rocketed to popularity and have the potential to fundamentally change many areas of life. Yet legal scholarship examining AI has insufficiently recognized the role of international and corporate competition in such a transformational field. Considering that competitive context and the Trilemma, this Article argues from a descriptive perspective that solely one policy prioritization choice is needed: whether to emphasize systemic risk mitigation or clear requirements, given that prioritizing innovation is effectively a given for many governmental and private actors. Next, regulation should prioritize systemic risk over clarity because foundation models present a substantive change in the potential for, and nature of, systemic disruption. Finally, the Article considers ways to mitigate regulators’ lack of legal clarity, examines the potential role of the public and private sectors in AI regulation under these assumptions, and argues instead for harm-based liability for AI providers when reasonably implementable, known technological advances could have prevented injury. This tradeoff thus promotes innovation and mitigates systemic risk from foundation AI models.

Wang on Can ChatGPT Personalize Index Funds’ Voting Decisions?

Chen Wang (UC Berkeley – School of Law) has posted “Can ChatGPT Personalize Index Funds’ Voting Decisions?” on SSRN. Here is the abstract:

ChatGPT has risen rapidly to prominence due to its unique features and generalization ability. This article proposes using ChatGPT to assist small investment funds, particularly small passive funds, in making more accurate and informed proxy voting decisions.

Passive funds adopt a low-cost business model. Small passive funds lack financial incentives to make informed proxy voting decisions that align with their shareholders’ interests. This article examines the implications of passive funds on corporate governance and the issues associated with outsourcing voting decisions to proxy advisors. The article finds that passive funds underspend on investment stewardship and outsource their voting proxy decisions to proxy advisors, which could lead to biased or erroneous recommendations.

However, by leveraging advanced AI language models such as ChatGPT, small passive funds can improve their proxy voting accuracy and personalization, enabling them to better serve their shareholders and navigate the competitive market.

To test ChatGPT’s potential, this article conducted an experiment using its zero-shot GPT-4 model to generate detailed proxy voting guidelines and apply them to a real-world proxy statement. The model successfully identified conflicts of interest in the election of directors and generated comprehensive guidelines with weight for each variable. However, ChatGPT has some limitations, such as token limitations, long-range dependencies, and likely ESG inclination.

To enhance its abilities, ChatGPT can be fine-tuned using high-quality, domain-specific datasets. However, investment funds may face challenges when outsourcing voting decisions to AI, such as data and algorithm biases, cybersecurity and privacy concerns, and regulatory uncertainties.

Cyphert & Martin on Developing a Liability Framework for Social Media Algorithmic Amplification

Amy Cyphert (West Virginia University – College of Law) and Jena Martin (same) have posted “‘A Change is Gonna Come:’ Developing a Liability Framework for Social Media Algorithmic Amplification” (U.C. Irvine Law Review, Vol. 13 (2022)) on SSRN. Here is the abstract:

From the moment social media companies like Facebook were created, they have been largely immune to suit for the actions they take with respect to user content. This is thanks to Section 230 of the Communications Decency Act, 47 U.S.C. § 230, which offers broad immunity to sites for content posted by users. But seemingly the only thing a deeply divided legislature can agree on is that Section 230 must be amended, and soon. Once that immunity is altered, either by Congress or the courts, these companies may be liable for the decisions and actions of their algorithmic recommendation systems, artificial intelligence models that sometimes amplify the worst in our society, as Facebook whistleblower Frances Haugen explained to Congress in her testimony.

But what, exactly, will it look like to sue a company for the actions of an algorithm?

Whether through torts like defamation or under certain statutes, such as those aimed at curbing terrorism, the mechanics of bringing such a claim will surely occupy academics and practitioners in the wake of changes to Section 230. To that end, this Article is the first to examine how the issue of algorithmic amplification might be addressed by agency principles of direct and vicarious liability, specifically within the context of holding social media companies accountable. As such, this Article covers the basics of algorithmic recommendation systems, discussing them in layman’s terms and explaining why Section 230 reform may spur claims that have a profound impact on traditional tort law. The Article looks to sex trafficking claims made against social media companies—an area already exempted from Section 230’s shield—as an early model of how courts might address other claims against these companies. It also examines the potential hurdles, such as causation, that will remain even when Section 230 is amended. It concludes by offering certain policy considerations for both lawmakers and jurists.

Balkin on Free Speech Versus the First Amendment

Jack M. Balkin (Yale Law School) has posted “Free Speech Versus the First Amendment” (UCLA Law Review, Forthcoming) on SSRN. Here is the abstract:

The digital age has widened the gap between the judge-made doctrines of the First Amendment and the practical exercise of freedom of speech. Today speech is regulated not only by territorial governments but also by the owners of digital infrastructure — for example, broadband and cellular providers, caching services, app stores, search engines, and social media companies. This has made First Amendment law less central and the private governance of speech more central.

When the free speech interests of digital companies and their end-users conflict, the major beneficiaries of First Amendment rights are likely to be the former and not the latter. Digital companies will try to use the First Amendment to avoid government regulation, including regulation designed to protect the free speech and privacy interests of end-users.

In response, internet reformers on both the left and the right will attempt to de-constitutionalize internet regulation: They will offer legal theories designed to transform conflicts over online speech from First Amendment questions into technical, statutory and administrative questions. In the U.S., at least, de-constitutionalization is the most likely strategy for imposing public obligations on privately-owned digital companies. If successful, it will make the First Amendment even less important to online expression.

The speed and scale of digital speech have also transformed how speech is governed. To handle the enormous traffic, social media companies have developed algorithmic and administrative systems that do not view speech in terms of rights. Accompanying these changes in governance is a different way of thinking about speech. In place of the civil liberties model of individual speech rights that developed in the twentieth century, the emerging model views speech in hygienic, epidemiological, environmental, and probabilistic terms.

The rise of algorithmic decisionmaking and data science also affect how people think about free expression. Speech becomes less the circulation of ideas and opinions among autonomous individuals and more a collection of measurable data and network connections that companies and governments use to predict social behavior and nudge end-users. Conceived as a collection of data, speech is no longer special; it gets lumped together with other sources of measurable and analyzable data about human behavior that can be used to make predictions for influence and profit.

Meanwhile, the speed and scale of digital expression, the scarcity of audience attention, and social media’s facilitation of online propaganda and conspiracy theories have placed increasing pressure on the standard justifications for freedom of speech, including the pursuit of truth and the promotion of democracy. The gap between the values that justify freedom of speech and what the First Amendment actually protects grows ever wider.

In response, some scholars have argued that courts should change basic First Amendment doctrines about incitement, defamation, and false speech. But it is far more important to focus on regulating the new forms of informational capitalism that drive private speech governance and have had harmful effects on democracy around the globe.

The digital age has also undermined many professions and institutions for producing and disseminating knowledge. These professions and institutions are crucial to the health and vitality of the public sphere. Changing First Amendment doctrines will do little to fix them. Instead, the task of the next generation is to revive, reestablish and recreate professional and public-regarding institutions for knowledge production and dissemination that are appropriate to the digital age. That task will take many years to accomplish.


Issacharoff & McKenzie on Managerialism and its Discontents

Samuel Issacharoff (NYU Law) and Troy A. McKenzie (same) have posted “Managerialism and its Discontents”
(Review of Litigation, Fall 2023) on SSRN. Here is the abstract:

Managerialism has rooted itself in the American system of civil litigation in the 40 years since the amendment of Rule 16 to recognize a new form of judicial authority, and since Judith Resnik gave the phenomenon the name that serves as its shorthand moniker. Time has not perfectly tamed the inherent tensions between the mantle judges had to adopt in the face of increasingly complex, high-stakes, and multi-jurisdictional disputes and their traditional role as detached adjudicators. One ready manifestation of that tension is the back-and-forth between fixed and discretionary practices in federal courts. This essay examines the gyrations between formal rules of application and those understood to be contextual, and it presents three approaches to the familiar rules/standard divide in federal procedure: formal managerialism, algorithmic managerialism, and structural managerialism. The first is readily exemplified by reforms to social security cases, which received a carve-out from the Federal Rules of Civil Procedure in 2022 and a set of formal rules tailored to their unique issues. Algorithmic managerialism hopes to harness the growing power of Artificial Intelligence to craft custom sets of discovery, motion, and other practice rules at the outset of litigation to maximize judicial economy. Lastly, structural managerialism addresses how courts choose the most efficient fora, from multidistrict litigation to bankruptcy, for resolving polycentric disputes, most notably mass torts. We conclude our review of these trends with a simple reflection: Managerialism is not just an established feature of federal judicial practice, but a new expansion may be on the horizon.

Siebecker on The Incompatibility of Artificial Intelligence and Citizens United

Michael R. Siebecker (U Denver Law) has posted “The Incompatibility of Artificial Intelligence and Citizens United” (Ohio State Law Journal, Vol. 83, No. 6, pp. 1211-1273, 2022) on SSRN. Here is the abstract:

In Citizens United v. FEC, the Supreme Court granted corporations essentially the same political speech rights as human beings. But does the growing prevalence of artificial intelligence (“AI”) in directing the content and dissemination of political communications call into question the jurisprudential soundness of such a commitment? Would continuing to construe the corporation as a constitutional rights bearer make much sense if AI entities could wholly own and operate business entities without any human oversight? Those questions seem particularly important, because in the new era of AI, the nature and practices of the modern corporation are quickly evolving. The magnitude of that evolution will undoubtedly affect some of the most important aspects of our shared social, economic, and political lives. To the extent our conception of the corporation changes fundamentally in the AI era, it seems essential to assess the enduring soundness of prior jurisprudential commitments regarding corporate rights that might no longer seem compatible with sustaining our democratic values. The dramatic and swift evolution of corporate practices in the age of AI provides a clarion call for revisiting the jurisprudential sensibility of imbuing corporations with full constitutional personhood in general and robust political speech rights in particular. For if corporations can use AI data mining and predictive analytics to manipulate political preferences and election outcomes for greater profits, the basic viability and legitimacy of our democratic processes hang in the balance. Moreover, if AI technology itself plays an increasingly important, if not controlling, role in determining the content of corporate political communication, granting corporations the same political speech rights as humans effectively surrenders the political realm to algorithmic entities. In the end, although AI could help corporations act more humanely, the very notion of a corporation heavily influenced or controlled by non-human entities creates the need to cabin at least somewhat the commitment to corporations as full constitutional rights bearers. In particular, with respect to corporate political activity, the growing prevalence of AI in managerial (and possibly ownership) positions makes granting corporations the same political speech rights as humans incompatible with maintaining human sovereignty.

Witt on The Digital Markets Act 

Anne Witt (EDHEC Business School – Department of Legal Sciences) has posted “The Digital Markets Act – Regulating the Wild West” (Common Market Law Review, Forthcoming 2023) on SSRN. Here is the abstract:

This contribution critically assesses the European Union’s Digital Markets Act (DMA). The DMA is the first comprehensive legal regime to regulate digital gatekeepers in the aim of making platforms markets fairer and more contestable. To this end, the DMA establishes 22 per se conduct rules for designated platforms. It also precludes national gatekeeper regulation by EU Member States, thereby calling into question the legality of the pioneering German sec. 19a GWB. The analysis shows that the DMA’s rules are not as rigid as they may appear at first sight. While it is more accepting of false positives than of false negatives, the DMA contains several corrective mechanisms that could allow the Commission to finetune the rules to address both the danger of over- and under-inclusiveness. A further positive is that the new regulation incorporates key concepts of the GDPR, and requires coordination between the Commission and key EU data protection bodies. On the downside, the DMA does not contain any substantive principles for the assessment of gatekeeper acquisitions, leaving a worrying gap. While the DMA’s conduct rules outlaw specific leveraging strategies in digital ecosystems and may thereby indirectly address certain non-horizontal concerns arising from gatekeeper acquisitions, it remains that the European Union’s existing guidance on merger control is seriously out of date. The merger guidelines therefore urgently need updating to include (workable) theories of harm for concentrations in the digital economy.

Shope on GPT Performance on the Bar Exam in Taiwan

Mark Shope (National Yang Ming Chiao Tung University; Indiana University Robert H. McKinney School of Law) has posted “GPT Performance on the Bar Exam in Taiwan” on SSRN. Here is the abstract:

This paper reports the performance of the GPT-4 Model of ChatGPT Plus (“ChatGPT4”) on the multiple-choice section of the 2022 Lawyer’s Bar Exam in Taiwan. ChatGPT4 outperforms approximately half of human test-takers on the multiple-choice section with a score of 342. This score, however, would not advance a test taker to the second and final essay portion of the exam. Therefore, this paper will not include an evaluation of ChatGPT4’s performance on the essay portion of the exam.

Gallese on the AI Act and the Right to Technical Interpretability

Chiara Gallese (University of Trieste Dept of Engineering) has posted “The AI Act Proposal: a New Right to Technical Interpretability?” on SSRN. Here is the abstract:

The debate about the concept of the so called right to explanation in AI is the subject of a wealth of literature. It has focused, in the legal scholarship, on art. 22 GDPR and, in the technical scholarship, on techniques that help explain the output of a certain model (XAI). The purpose of this work is to investigate if the new provisions introduced by the proposal for a Regulation laying down harmonised rules on artificial intelligence (AI Act), in combination with Convention 108 plus and GDPR, are enough to indicate the existence of a right to technical explainability in the EU legal framework and, if not, whether the EU should include it in its current legislation. This is a preliminary work submitted to the online event organised by the Information Society Law Center and it will be later developed into a full paper.