Algorithmic fairness has emerged as a central issue in ML, and it has become standard practice to adjust ML algorithms so that they will satisfy fairness requirements such as Equal Opportunity. In this paper we consider the effects of adopting such fair classifiers on the overall level of _ecosystem fairness_. Specifically, we introduce the study of fairness with competing firms, and demonstrate the failure of fair classifiers in yielding fair ecosystems. Our results quantify the loss of fairness in systems, under a variety of conditions, based on classifiers' correlation and the level of their data overlap. We show that even if competing classifiers are individually fair, the ecosystem's outcome may be unfair; and that adjusting biased algorithms to improve their individual fairness may lead to an overall decline in ecosystem fairness. In addition to these theoretical results, we also provide supporting experimental evidence. Together, our model and results provide a novel and essential call for action.

We study the robustness of cheap-talk equilibria to infinitesimal private information of the receiver in a model with a binary state-space and state-independent sender-preferences.

This paper analyzes cooperative data-sharing between competitors vying to predict a consumer's tastes. We design optimal data-sharing schemes both for when they compete only with each other, and for when they additionally compete with an Amazon – a company with more, better data. We show that simple schemes – threshold rules that probabilistically induce either full data-sharing between competitors, or the full transfer of data from one competitor to another – are either optimal or approximately optimal, depending on properties of the information structure. We also provide conditions under which firms share more data when they face stronger outside competition, and describe situations in which this conclusion is reversed.

The classic herding model examines the asymptotic behavior of agents who observe their predecessors' actions as well as a private signal from an exogenous information structure. In this paper we introduce a self-interested sender into the model, and study the sender's problem of designing this information structure. If agents cannot observe each other the model reduces to Bayesian persuasion. However, when agents observe predecessors' actions, they may learn from each other, potentially harming the sender. We identify necessary and sufficient conditions under which the sender can nevertheless obtain the same utility as when the agents are unable to observe each other.

Firms engaged in electronic commerce increasingly rely on predictive analytics via machine-learning algorithms to drive a wide array of managerial decisions. The tuning of many standard machine learning algorithms can be understood as trading off bias (i.e., accuracy) with variance (i.e., precision) in the algorithm's predictions. The goal of this paper is to understand how competition between firms affects their strategic choice of such algorithms. To this end, we model the interaction of two firms choosing learning algorithms as a game and analyze its equilibria. Absent competition, players care only about the magnitude of predictive error and not its source. In contrast, our main result is that with competition, players prefer to incur error due to variance rather than due to bias, even at the cost of higher total error. In addition, we show that competition can have counterintuitive implications---for example, reducing the error incurred by a firm's algorithm can be harmful to that firm---but we provide conditions under which such phenomena do not occur. In addition to our theoretical analysis, we also validate our insights by applying our metrics to a publicly available data set.

The Bayesian persuasion paradigm of strategic communication models interaction between a privately-informed agent, called the sender, and an ignorant but rational agent, called the receiver. The goal is typically to design a (near-)optimal communication (or signaling) scheme for the sender. It enables the sender to disclose information to the receiver in a way as to incentivize her to take an action that is preferred by the sender. Finding the optimal signaling scheme is known to be computationally difficult in general. This hardness is further exacerbated when there is also a constraint on the size of the message space, leading to NP-hardness of approximating the optimal sender utility within any constant factor. In this paper, we show that in several natural and prominent cases the optimization problem is tractable even when the message space is limited. In particular, we study signaling under a symmetry or an independence assumption on the distribution of utility values for the actions. For symmetric distributions, we provide a novel characterization of the optimal signaling scheme. It results in a polynomial-time algorithm to compute an optimal scheme for many compactly represented symmetric distributions. In the independent case, we design a constant-factor approximation algorithm, which stands in marked contrast to the hardness of approximation in the general case.

When committees make decisions, voting rules are coupled with one of three disclosure rules: open voting, in which each committee member's individual vote is revealed; anonymous voting, in which only an anonymized tally is publicized; and secret voting, in which only the outcome is disclosed. I focus on strategic voters who have a preference for strategic ambiguity, and show that the amount of disclosure may have a non-monotonic effect on both the accuracy of the decision and the welfare of the voters. In particular, anonymous voting can yield both lower accuracy and higher welfare than both open and secret voting.

Users of social, economic, or medical networks share personal information in exchange for tangible benefits, but may be harmed by leakage and misuse of the shared information. I analyze the effect of enhancing privacy in the presence of two opposing forces: network effects and informational interdependencies. I show that two privacy enhancements---reducing the likelihood of leakage and decreasing the level of informational interdependence---have opposite effects on the volume of information sharing, and that although they always seem beneficial to non-strategic users, both privacy enhancements may backfire when users are strategic.

Much of the literature on rational cryptography focuses on analyzing the strategic properties of cryptographic protocols. However, due to the presence of computationally-bounded players and the asymptotic nature of cryptographic security, a definition of sequential rationality for this setting has thus far eluded researchers. We propose a new framework for overcoming these obstacles, and provide the first definitions of computational solution concepts that guarantee sequential rationality. We argue that natural computational variants of sub game perfection are too strong for cryptographic protocols. As an alternative, we introduce a weakening called threat-free Nash equilibrium that is more permissive but still eliminates the undesirable "empty threats'' of non-sequential solution concepts. To demonstrate the applicability of our framework, we revisit the problem of implementing a mediator for correlated equilibria (Dodis-Halevi-Rabin, Crypto'00), and propose a variant of their protocol that is sequentially rational for a non-trivial class of correlated equilibria. Our treatment provides a better understanding of the conditions under which mediators in a correlated equilibrium can be replaced by a stable protocol.

A Nash equilibrium is an optimal strategy for each player under the assumption that others play according to their respective Nash strategies. In the presence of irrational