2024 Decentralization Conference

2024 NSF/CEME Decentralization Conference, Vanderbilt, April 2024

Mechanism Design with AI and Distributed Ledgers

Friday, April 19, 2024

Panel: Blockchains & Mechanism Design

Trust in Blockchains

Daniel Obermeier, Decentralization vs. Blockchain Neutrality: The Unequal Burden of Ethereum’s Market Mechanism on dApps

John Conley, AI needs blockchain: Trustless solutions to failures in machine to colloidal markets

Saturday, April 20, 2023

Decentralization in Blockchains

Hanna Halaburda, Permissioned vs Permissionless Blockchain Platforms: Tradeoffs in Trust and Performance

Harang Ju, The Virtualization Hypothesis: Explaining Sustained Blockchain Decentralization with Quasi-Experiments

Algorithm Design

Shota Ichihashi, Buyer-Optimal Algorithmic Consumption

Discrimination and Manipulation

John Zhu, Interventions Against Machine-Assisted Statistical Discrimination

Matheus Xavier Ferreira, I See You! Robust Measurement of Adversarial Behavior

Saturday, April 20, 2023

Collusion

In Koo Cho, Collusion through Algorithms: Fact or Myth?

Ran Shorrer, Algorithmic Collusion by Large Language Models

PANEL: AI and Mechanism Design

Recent advances in AI have already begun to disrupt economic, social, and political processes.    That will continue.  In this panel, we will discuss where and how mechanism design can be helpful in explaining and anticipating those effects, and, most importantly, how it might be used to design mechanisms with humans and AI.

This panel is organized around one big question

What if we were to begin the field of mechanism design now?

Our discussion will focus on four questions:

1. Do “amplified humans” require rethinking any of the core assumptions or findings from mechanism design?  In particularly, what results from experimental economics might be rethought?

2. How do we model the capabilities and attributes of AI and Hybrid agents?  Do they have preferences, beliefs, and information?   How do we model their abilities?   Is the alignment problem something that merits the status of individual rationality or incentive compatibility? Ide and Talamas Artificial Intelligence in the Knowledge Economy assume comparability between AI and people and one-dimensional knowledge. They then show that the implications of AI agents has markedly different implications depending on whether its abilities rise to managerial level?

3. How does the possibility of a second type of actor change mechanism design, market design and organizational theory?   Does machine learning and the capability of AI allow for mechanisms and organizational forms heretofore not possible?  If Arrow were to rewrite The Limits of Organization,  if Williamson were to rewrite The Economics of Organizations, or if Debreu were to rewrite The Theory of Value how would they differ?

4. Does the opaqueness of AI and the fact that incentives and punishments operate only in a mathematical space as opposed to in the human domain have implications for Mechanism Design? 

To begin,we have asked John Ledyard to provide us with 10 minutes on what the original goal of this conference was.

Call for Papers

We seek theoretical, experimental, and empirical papers that apply tools and insights from mechanism design, game theory, and formal theory that fall into two broad categories:

1. Formal Models of Mechanism, Markets, Democracies with AI.
2. Blockchains and Distributed Ledgers.

At the end of the document, we have added some context for how this conference aligns with the conference’s historical agenda. As always, we welcome theoretical, experimental, and empirical papers of general interest to the mechanism design community.

Formal Models of Mechanism, Markets, Democracies with AI:

The rapid progress of AI will cause disruptions across the economy and society. Labor markets, job classifications, organizational structures, democratic and educational institutions, and the nature of work will all confront disruptions.

These disruptions raise any number of theoretical questions: how should humans allocate their limited capacity in light of the capacities of AI? Which increases in human capacity created by AI will produce zero-sum red queen games, and which produce positive sum outcomes? How should market and democratic mechanism be redesigned given humans can be assisted by AI and the potential for Autonomous AI actors? How will organizational structures change with AI agents? How should AI be regulated? How can hybrid teams of humans and AI best make decisions? The list goes on and on. We do not want to limit what people might find theoretically of interest.

We present here a tentative set of categories that could be of interest to theorists.

• AI-Assisted Human Actors: In an ever-growing number of contexts, human actions will be informed by AI. AI-Assisted Human Actors have different cognitive capacities than human actors. They have greater working memory, more bandwidth, and more computational power. Assumptions about human behavior, particularly those drawn from behavioral economics may require rethinking.
• Autonomous AI Actors: Mechanisms will include Autonomous AI Actors as well as human actors. How should formal models represent AI agents? How are their actions and capacities meaningfully different from those of human actors? Does the fact that AI relies on objective functions and not on underlying preferences make their behavior less robust? AI algorithms often exhibit alignment problems where the algorithm discovers an action that performs well according to the mathematical objective function assigned to an algorithm but which does not align with the underlying preferences of the designer of the algorithm.
• Human + AI Mechanisms: In many decision domains including hiring decisions, college admissions, and parole decisions, humans rely on AI as assistants and also as autonomous actors. How should such decision protocols be designed? Humans and autonomous AI actors also interact in markets and on social media platforms. How should such markets and platforms be designed when they include both human and autonomous AI actors?
• AI Mechanisms: AI makes possible new types and forms of mechanism by changing the structure of communication. For example, AI tools can absorb and categorize parallel communication from large numbers of human actors. AI can also, instantaneously, and strategically assign different information to different actors.
• AI-Design: Autonomous AI agents and AI assisted human agents both rely on large language model (LLMs). LLMs are capable of categorizing enormous amounts of unstructured data and then making predictions about the consequences of those actions. LLMs rely on an objective function which influences how the raw data becomes structured into categories and, ultimately, predictions. The distinction between autonomous AI actors and AI assisted human actors influences the design of an AI objective function. AI actors act independently. Their objective functions should ensure robust aggregate behavior. When AI acts as an advisor or a “signal” to a human actor, the objective function should be designed to best inform an agent. Is the AI agent part of the mechanism design or is it an independent actor (and, if so, who deploys it?) Do recent advances in AI change how we model hostile agents?

Blockchains and Distributed Ledgers:

Blockchains incorporate mechanisms and incentives as part of their protocols. Each chain’s approach to arriving at a consensus view of the chain state and making honest behavior by nodes in the validation network an equilibrium strategy, can be seen as a mechanism. Protocols also include more directed rewards to encourage beneficial behaviors, such as nodes making chain data available to users and the network, and punishments to discourage harmful ones, such as transaction censorship, front-running, Sybiling, and griefing the network.

Blockchains have several features that make it challenging to approach them as a conventional mechanism or game:

• Endogenous Players: The player set is endogenous: There are thousands of blockchain projects competing for widespread adoption. Users choose which project or projects to join and may choose not to participate at all. Moreover, even these endogenous set of players can change in real time with every block committed.
• Diverse Types: Blockchains have many player types: simple account holders, smart contract creators, stake delegators, voting stakeholders, light nodes following the chain state, full nodes participating in block building and transaction processing, and an array of governance, developer, and auditing roles. The distribution of individual players over these types is also endogenous and dynamic.
• Unbounded Strategy Spaces: Blockchain mechanism designers do not have control over the strategy space. There are many significant, out-of-band, or off-chain, actions that a designer cannot incorporate into an on-chain mechanism. For example, getting a token onto a popular exchange, pumping a token’s value, using different hardware to create “work”, or spreading o fear, uncertainty, and doubt (FUD). In some protocols, governance can be used to upgrade, fix bugs, or even fundamentally change, the mechanism. Such protocols effectively incorporate a meta-mechanism for mechanism choice The continuation game after a governance action might have a very different set of strategies, payoffs, and even players.
• Endogenous Payoffs: Payoffs can be endogenous: The designer might control the internal coin or token payoffs a mechanism provides, but he cannot control the value of those tokens. Players often get payoffs from off-chain sources. For example, revenues from selling tokens in ICOs, or from allocations to initial stakeholders, funders, or principles, are not modeled as payoffs resulting from choosing strategies within blockchain mechanisms. Building and deploying smart contracts with exploits, and payoffs to nation-states from forcing discloser of user identities tax reasons, are not, and cannot be incorporated into mechanisms.

There also exist a set of core research questions that are more traditional concerns for mechanism design.

• Implementability: Algorithmic mechanism design adds computability to the list of desiderata, making implementation even more difficult. Players are often modeled as decision theoretic, rather than strategic players. The question of what can and cannot be implemented on a blockchain is open. Where are the limits, and how can protocols either bring more of the key levers within a designer’s control, or make them less relevant to outcomes or payoffs?

• Human and AI Interactions: Blockchain is a domain where humans with conventional preferences and limitations, interact with AIs with pseudo-preferences, created or influenced by their creators. Does their superior computational skill and capacity give AIs insurmountable advantages against human players? Can mechanisms be designed using blockchain or other technologies to level to playing field?
• Identification: Humans and human institutions have a physical existence which makes it possible to connect them to their historical actions. It is costly for human-based entities to change preferences, abilities, or habitual behaviors, rapidly, and the cost of building a reputation built on historical actions incentivizes consistency. It is not clear how one could even identify a specific, individual AI, and even if we could, it is quick and to change any AIs programing/preferences. Individual AI can also be perfectly replicated any number of times at low cost. Blockchain, however, can create a difficult to mutate, ordered, cryptographically bound, attestations, records of actions and transactions, and contingent, enforceable commitments. Can blockchain be used to as a data layer for mechanisms that make it possible to create AI histories, provable reputations, credible identities, and for humans and AI to transact and interact without the need for trust, and to their mutual benefit?

Background: Though AI might appear a “new” topic for the Decentralization conference, we believe that aligns with the core themes and questions that have animated the NSF/CEME Decentralization Conference from its beginning more than fifty years ago.

Hurwicz, Marschak, Reiter, Radner, and Arrow’s original goal was to derive theoretical foundations for the design and analysis of mechanisms (systems) that produce, allocate, and recombine goods, information, and services. The conference has long emphasized fundamental questions such as how to best design markets, organizational structures, and voting rules in light of informational and computational constraints.

Early papers in mechanism design paid close attention to the informational and computational costs of institutional structures. Market economies, for example, were shown to require lower dimensional messages spaces than centrally planned economies. And, for a period of time, the conference had regular interactions with computer scientists to gain insights into modeling of computation. The research spawned by those conversations laid the groundwork for the field of algorithmic mechanism design.

A combination of trends, the rise of game theory and the emphasis on incentive compatibility and the derivation of the revelation principle, meant that computational and informational concerns became less central. (Only to theorists. Those concerns were always issues for practical mechanism design.) The limits of human cognition re-emerged as a concern but more from a behavioral economics perspective.

Scholars of mechanism design now find themselves at an important moment. Advances computational power, computer science, and in machine learning have led to the development of blockchains and artificial general intelligence. Individually and jointly these present deep challenges to our current institutional architecture, the mechanisms, institutions, and organizations that we used to produce, allocate, and decide.

The advances in AI can be seen from two perspectives. First, they can be modeled as enhancements to human abilities. This approach obliges a return to the consideration of computational and informational constraints and costs as those have shifted. It also suggests a rethinking of the behavioral turn. How will human biases be affected by the presence of AIs? Will participation in “real” mechanisms become more rational – i.e. more like the behavior assumed in game theory. What does the access to AI do to Common Knowledge? And if Common Knowledge is based on AIs, is that now part of the mechanism? And, if so, can a mechanism be designed to be nonmanipulable?

Alternatively, as noted in the call for papers, AIs can be seen as a second type of agent that operates within existing mechanisms. These agents have assigned objective functions and possess capacities and limitations that differ from those of humans. Optimal mechanisms with humans and AI agents likely differ from optimal mechanisms that rely only on humans. In addition, AI agents cost less than humans and can be created almost instantaneously. These features present opportunities and challenges for institutional and mechanism design.

From a theorists point of view, mechanisms with AI agents within them (designed by the mechanism designer) could be thought of as another class of mechanism. That class of mechanisms may have practical applications if mechanisms are currently limited by the computational capability of people or computers. The impact of AI on combinatorial auctions would be one example worth considering. More generally, one might ask if increases in power move us closer to a world consisting primarily of direct revelation mechanisms.

Finally, can blockchain technology allow and incentivize AI agents to be identifiable, and behave honestly in commercial and other exchanges? Can blockchain be part of larger mechanisms that allow humans and AIs to usefully communicate and cooperate despite their extremely different motivations and capacities?