payment of a fixed amount does not motivate the expert to be either truthful or careful, let alone to actively seek out new information. A “scoring rule” is a payment function that depends on the expert’s prediction and the actual outcome in such a way as to motivate truthful participation. ³² Shared scoring rules can form the basis of self-financ-ing (budget-balanced) wagering mechanisms to obtain multiple individual forecasts. ²³ Indeed, the line between scoring rules and markets becomes blurred. For example, the most common automated market maker used for prediction markets can be viewed as a sequential shared scoring rule. ¹⁸

Rating systems are forums for gathering subjective opinions on a variety of things, such as movies, restaurants, or trading partners. Unlike forecasts, rating and reputation systems have no fundamental truths on which to base rewards. Although rating systems may provide personalized recommendations or advice, the incentives of raters may not align with these goals. Real rating systems do provide considerable value, despite the persistence of spam and pollution. Designing rating systems that are resistant to manipulation is an important challenge, requiring both good algorithms and good economics. ²⁸

Peer Production and interaction

“Peer production,” a term coined by Benkler, ⁶ refers to large-scale collaboration that is not based on price signals and occurs outside of the typical hierarchies and reward structures provided by firms. Salient examples of successful artifacts of peer production include Wikipedia and Linux. Social production, a more general phenomenon, describes the output of social relations—for example, the videos that people upload to You Tube and the content on social-networking sites such as Facebook. Taken together, these activities make up an energetic swath of the e-commerce landscape, both because of the opportunities to promote non-traditional production through appropriate feedback, trust, and accounting mechanisms and because of opportunities for targeted advertising and monetization efforts.

How can it be that peer production seems to succeed despite the widely accepted economic model of people

as selfish and rational actors? Nontraditional motivations, such as hedonic pleasure from doing useful work and civic pride in observing community and societal norms, seem to be a part of the story. Indeed, the established field of behavioral economics seeks to explain people’s social or “ other-regarding” preferences. ¹⁶ Monetary rewards can actually lead to the crowding out of social motivations and to increasingly selfish behavior. In one noted example, when a country moved from a system of voluntary blood donations to one with small payments it found that donation rates went down instead of up. ¹⁷

Despite positive examples of peer production, there is little formal knowledge about the design of successful peer-production systems. Pertinent methodologies seem necessarily more indirect than those of economic mechanism design, ²⁰ given that actions are intrinsically voluntary. One challenge is to design systems that observe behaviors with a view to learning (social) preferences. Can environments be usefully modulated through appropriate constraints and affordances (for example, with moderator rights and other tiers), the assignment of rewards (say, gold stars), ³⁵ and the ability to aggregate and disseminate information about peers (such as through scoreboards and social context)?

Lessons learned from peer-to-peer file sharing suggest that incentive considerations will remain important even in the presence of other-regarding preferences. Early protocols failed to provide appropriate incentives for the uploading of files, and systems such as Gnutella suffered from a large amount of free-riding. ³ The Bit Torrent protocol addressed this problem by limiting a user’s download rate according to upload history, thus mitigating incentives to free-ride. But such tit-for-tat during bilateral peering introduces its own market inefficiencies; for example, users cannot contribute upload resources in return for credit for later downloads unless centralized trackers are employed. Hybrid systems that allow for accounting and some form of currency may be of interest, though they have associated challenges with regard to dynamic stability. ²² There are interesting directions as well in the introduction of social

structure into peering societies. ²⁷

Reputation and trust metrics have an important role to play in Internet commerce, ²⁹ yet finding the right design can be quite a subtle problem. For example, one study suggests that good reputations of sellers on eBay are sometimes the result of a badly designed feedback protocol. A seller, who has the “last move,” can punish a buyer who leaves negative feedback; the seller may respond with negative feedback of his own about the buyer. ⁸

Recent work has formalized the challenges of providing provably non-manipulable trust metrics in graph-theoretic terms. Suppose that players are nodes and that player j can choose to lay down an edge (j, k) to another player k (possibly weighted), indicating a trust relationship; and suppose further that nodes can misrepresent trust information and create new (“fake”) nodes. Various algorithms can be defined to compute pairwise trust between nodes, and their informativeness and manipulability can be compared. For instance, the Eigen Trust algorithm²¹ is vulnerable to Sybil attacks; one player can lay down multiple fake nodes that pump reputation flow in its direction. Other algorithms are more robust. Consider, for instance, defining pairwise trust i-j (i’s trust of j) as the number of hops on the shortest path from i to j in an unweighted directed graph. Player j cannot reduce the i-j path length and improve the i-j trust by adding fake nodes and ( directed) edges, as these nodes and edges can only affect shortest paths that flow through j and thus must leave the i-j path unaffected. ⁴ A similar argument establishes that node j cannot reduce the i-k trust for any node k that i trusts more than j. The only paths affected would be those that go through j and therefore ultimately reach nodes less trusted by i than j.

One outstanding challenge in the area of trust metrics is to find a satisfactory definition of informativeness; current axiomatic approaches appear unsatisfactory in this regard. With such a definition in hand, tensions between robustness and informativeness could be explored and perhaps also allow for mitigating factors such as the presence of other-regarding and altruistic actors within peering systems.