ReputationSystemsT

he Internet offers vast new opportu-nities to interact with totalstrangers. These interactions can befun, informative, even profitable.But they also involve risk. Is theadvice of a self-proclaimed expert at

expertcentral.com reliable? Will an unknown dot-com site or eBay seller ship items promptly withappropriate packaging? Will the product be thesame one described online?

Prior to the Internet, such questions wereanswered, in part, through personal andcorporate reputations. Vendors

provided references, Better Business Bureaus talliedcomplaints, and past personal experience and per-son-to-person gossip told you on whom you couldrely and on whom you could not. Participants’standing in their communities, including their rolesin church and civic organizations, served as a valu-able hostage.

Internet services operate on a vastly larger scale

COMMUNICATIONS OF THE ACM December 2000/Vol. 43, No. 12 45

Paul Resnick, Richard Zeckhauser, Eric Friedman, and Ko Kuwabara

MA

RC

MO

NG

EAU

For buyers and sellers alike, there’s

no better way to earn one another’s

trust in onlineinteractions.

46 December 2000/Vol. 43, No. 12 COMMUNICATIONS OF THE ACM

than Main Street and permit virtually anonymousinteractions. Nevertheless, reputations still play a majorrole. Systems are emerging that respect anonymity andoperate on the Internet’s scale. A reputation system col-lects, distributes, and aggregates feedback about partic-ipants’ past behavior. Though few producers orconsumers of the ratings know one another, thesesystems help people decide whom to trust,encourage trustworthy behavior, and deterparticipation by those who are unskilledor dishonest.

For example, consider eBay, thelargest person-to-person online auc-tion site, with more than four mil-lion auctions active at a time: itprovides limited insurance, andbuyers and sellers both acceptsignificant risks. There are prob-lematic transactions to be sure.Nevertheless, the overall rate ofsuccessful transactions remainsastonishingly high for a market as“ripe with the possibility of large-scale fraud and deceit” as eBay [5].

The high rate of successful trans-actions is attributerd by eBay to itsreputation system, called the Feed-back Forum. After a transaction iscomplete, the buyer and seller havethe opportunity to rate each other(1, 0, or �1) and leave com-ments (such as “good transac-tion,” “nice person to do businesswith,” “would highly recom-mend”). Participants have runningtotals of feedback points attached(visibly) to their screen names, whichmight be pseudonyms. Yahoo! Auc-tion, Amazon, and other auction sitesfeature reputation systems like eBay’s,with variations, including a rating scale of1�5, several measures (such as friendliness, promptresponse, quality product), and averaging instead oftotal feedback score.

Reputation systems have also spread beyond auctionsites. For example, Bizrate.com rates registered retailersby asking consumers to complete a survey form aftereach purchase. So-called “expert sites” (www.expertcen-tral.com and www.askme.com) provide Q&A forumsin which self-proclaimed experts provide answers forquestions posted by other users in exchange for reputa-tion points and comments. Product review sites (suchas www.epinions.com) offer rating services for productreviewers (the better the review, the more points the

reviewer receives). iExchange.com tallies and displaysreputations for stock market analysts based on the per-formance of their picks.

Why are these explicit reputation systems soimportant for fostering trust among strangers? Toanswer, it helps to first examine how trust builds nat-

urally in long-term relationships.First, when people interact with

one another over time, thehistory of past interactions

informs them about theirabilities and dispositions.Second, the expectationof reciprocity or retalia-tion in future interac-tions creates an incentivefor good behavior. (Polit-ical scientist RobertAxelrod calls this the “shadow of the future”[2].) An expectation thatpeople will consider one another’s pasts in futureinteractions constrains

behavior in the present.Among strangers, trust is

understandably much more diffi-cult to build. Strangers lack known

past histories or the prospect of futureinteraction, and they are not subject to a net-work of informed individuals who wouldpunish bad and reward good behavior. In

some sense, a stranger’s good name is not atstake. Given these factors, the temptation to“hit and run” outweighs the incentive to coop-erate, since the future casts no shadow. Reputation systems seek to establish the

shadow of the future to each transaction by creat-ing an expectation that other people will look back

on it. The connections among such people may be sig-nificantly weaker than in transactions on a town’sMain Street, but their numbers are vast in comparison.At eBay, for example, a stream of buyers interacts withthe same seller. They may never buy an item from theseller again, but by sharing their opinions about theseller via the Feedback Forum, they construct a mean-ingful history of the seller. Future buyers, lacking per-sonal histories with particular sellers, may still basetheir buying decisions on a sufficiently extensive pub-lic history. If buyers do behave this way, the sellers’ rep-utations will affect their future sales. Hence, they seekto accumulate as many positive points and commentsas possible and avoid negative feedback. Through themediation of a reputation system—assuming buyers

auction sites

enable trash to be

shuttled across the

country and in the

process transmuted

into treasures.

MA

RC

MO

NG

EAU

provide and rely on feedback—isolated interactionstake on attributes of long-term relationships. In termsof building trust, a boost in the quantity of informa-tion compensates for a significant reduction in its quality.

For people trying to sell off, say, their old LP-recordcollections, reputation systems might seem like a nui-sance. But consider such an effort in a market with nosuch system, and hence no obvious distinctionbetween sellers in terms of, say, quality of goods andshipping service. Buyers would be reluctant to pay fullprices given their uncertainty about the sellers’ quality(such as whether they reveal scratches in the records atthe time of sale). However, high-quality sellers wouldbe reluctant to accept discounted prices. Over time,high-quality sellers would desert the market. Eventu-ally, only the lowest-quality sellers would remain, adynamic economist George Akerlof memorialized asthe “market for lemons” [1].

Reputation systems can reverse this flow and“unsqueeze” a bitter lemon. With clear reputationmarkers, low-quality sellers get lower prices, leaving ahealthier market with a variety of prices and quality ofservice. For example, sellers with stellar reputationsmay enjoy a premium on their services; some usersmay be willing to pay for the security and comfort ofhigh-quality services. Such premiums are observed inauctions in auctions of coins and computer chips oneBay [3, 6, 7]. The benefits of informative reputationsystems return to buyers and to sellers, enabling theold LPs to spin out the door.

Ratings are not the only way to convey reputations.When agreeing to be rated is optional (such as whenregistering as a retailer at bizrate.com), doing so islikely an indication of higher-quality services, evenbefore ratings are available. Other ways to indicatequality are to use one’s real name, rather than a pseu-donym, and to indicate on a Web site that one alsohas a physical store with its attendant overhead costs.

To operate effectively, reputation systems require atleast three properties:

• Long-lived entities that inspire an expectation offuture interaction;

• Capture and distribution of feedback about cur-rent interactions (such information must be visi-ble in the future); and

• Use of feedback to guide trust decisions.

In the offline world, capturing and distributingfeedback is costly. Businesses often collect feedbackfrom consumers but tend not to publicize the com-plaints. A few independent services, such as Zagat’s for

restaurants and Consumer Reports magazine for appli-ance repair histories, systematically capture and dis-seminate feedback. For the most part, however,reputations travel haphazardly through word ofmouth, rumor, or the mass media.

The Internet can vastly accelerate and add structureto the process of capturing and distributing informa-tion. To post feedback, users need only fill out anonline form; a mere mouse click is often enough.Where interactions are mediated electronically, objec-tive information about performance may be capturedautomatically (such as delay from question to responseat an expertise site). The same technology facilitatingmarket-style interaction among strangers also facili-tates the sharing of reputations that maintain trust.

Despite this promise, significant challenges remainin the operating phases of such systems: eliciting, dis-tributing, and aggregating feedback.

Eliciting feedback encounters three related prob-lems. The first is that people may not bother to providefeedback at all. For example, when a trade is completedat eBay, there is little incentive to spend another fewminutes filling out a form. That many people do so isa testament to their community spirit, or perhaps theirgratitude or desire to exact revenge. People could bepaid for providing feedback, but more refined schemes,such as paying on the basis of concurrence with futureevaluations by others, would be required to assure thattheir evaluations are thorough.

Second, it is especially difficult to elicit negativefeedback. For example, at eBay, it is common practiceto negotiate first before resorting to negative feedback.Therefore, only really bad performance is reported.Even then, fear of retaliatory negative feedback orsimply a desire to avoid further unpleasant interac-tions may keep a dissatisfied buyer quiet. In the end,because information about patterns of moderate dis-content may remain invisible, buyers cannot shun thesellers who foster such discontent.

Third is the difficulty of ensuring honest reports.One party could blackmail another, threatening topost negative feedback unrelated to actual perfor-mance. At the other extreme, in order to accumulatepositive feedback, a group of sellers might collaborateand rate one another positively, artificially inflatingtheir individual reputations.

Distributing feedback, the second phase, poses itsown challenges. One is name changes. At many sites,people choose pseudonyms when registering. If theyregister again, they might choose another pseudonym,effectively erasing prior feedback. Reputations can stillhave effects, since newcomers want to accrue positivefeedback, and those with established reputations want

COMMUNICATIONS OF THE ACM December 2000/Vol. 43, No. 12 47

to avoid negative feedback. Game-theory analysisdemonstrates that there are inherent limitations to theeffectiveness of reputation systems when participantsare allowed to start over with new names [4]. In par-ticular, newcomers (those with no feedback) shouldalways be distrusted until they have somehow paidtheir dues, either through an entry fee or by acceptingmore risk or worse prices while developing their repu-tations. Another alternative is to prevent namechanges—either by using real names or preventingpeople from acquiring multiple pseudonyms, a tech-nique called “once-in-a-lifetime pseudonyms” [4].

A second difficulty in distributing feedback stemsfrom the lack of portability from system to system.Amazon.com initially allowed users to import their rat-ings from eBay. But when eBay protested vigorously,claiming its user ratings were proprietary, Amazon dis-continued its rating-import service. Limited distribu-tion of feedback limits its effectiveness; the future castsa shadow on only a single online arena, not on many.Efforts are under way to construct a more universalframework. For example, virtualfeedback.com providesa rating service for users across different systems, but ithas yet to gain wide public acceptance.

Finally, there is also potential difficulty in aggre-gating and displaying feedback, so it is useful in influ-encing future decisions about whom to trust. Netfeedback (positives minus negatives) is displayed ateBay; other sites, including Amazon.com, display anaverage. These simple numerical ratings fail to conveyimportant subtleties of online interactions; for exam-ple, Did the feedback come from low-value transac-tions? What were the reputations of the peopleproviding the feedback?

As a solution to the ubiquitous problem of trust innew short-term relationships on the Internet, reputa-tion systems have immediate appeal; the participantsthemselves create a safe community. Unfortunately,these systems face complex challenges, many of whichyield no easy solutions. Efforts are under way toaddress these problems; for example, the ReputationsResearch Network (see databases.si.umich.edu/reputa-tions) represents a first step toward recognizing repu-tation systems as a subject of study and as a vital assetfor the safety of online interaction environments.

Internet-based reputation systems, like traditionalmarkets, aggregate vast amounts of information,which then significantly influences choices made bybusinesses, as well as by individuals. The parallel mayend there. The theoretical underpinnings of the effec-tive operation of markets are well understood, and theaggregation to a brief set of statistics, namely a singleprice for each item, proceeds automatically.

Today’s reputation systems, by contrast, shouldn’t

work in theory. Individuals shouldn’t be expected tomake the effort to provide evaluations; negative eval-uations should be avoided completely; and vendorsshould be expected to develop sophisticated ways tomanipulate and trick the system. Even if all reportingwere complete and honest, users would find it virtu-ally impossible to utilize the torrents of informationthey receive on other participants, given the lack ofsatisfactory summary statistics.

Despite their theoretical and practical difficulties,it is reassuring that reputation systems appear to per-form reasonably well. Systems that rely on the partic-ipation of large numbers of individuals accumulatetrust simply by operating effectively over time.Already, Internet-based reputation systems performcommercial alchemy. On auction sites, for example,they enable trash to be shuttled across the countryand in the process transmuted into treasures.

We conclude with an allusion to democracy,another theoretically flawed and practically chal-lenged system that nonetheless appears to performmiracles. Were Winston Churchill, the World War II-era British prime minister, to comment on reputationsystems and building trust as he did on democracyand government, he might say: “Reputation systemsare the worst way of building trust on the Internet,except for all those other ways that have been triedfrom time-to-time.”

References1. Akerlof, G. The market for ‘lemons’: Quality uncertainty and the market

mechanism. Quart. J. Econom. 84 (1970), 488–500. 2. Axelrod, R. The Evolution of Cooperation. Basic Books, New York, 1984. 3. Bajari, P. and Hortacsu, A. Winner’s curse, reserve prices, and endogenous

entry: Empirical insights from eBay auctions; see www.stanford.edu/~bajari/wp/auction/ebay.pdf.

4. Friedman, E. and Resnick, P. The social cost of cheap pseudonyms; seewww.si.umich.edu/~presnick/papers/identifiers/index.html.

5. Kollock, P. The production of trust in online markets. In Advances inGroup Processes, vol. 16, E. Lawler, M. Macy, S. Thyne, and H. Walker,Eds. JAI Press, Greenwich, CT, 1999; see also www.sscnet.ucla.edu/soc/faculty/kollock/papers/online_trust.htm.

6. Lucking-Reiley, D., Bryan, D., Prasad, N., and Reeves, D. Pennies fromeBay: The determinants of price in online auctions; seewww.vanderbilt.edu/econ/reiley/papers/PenniesFromEBay.pdf.

7. Houser, D. and Wooders, J. Reputation in auctions: Theory and evidencefrom eBay; see bpa.arizona.edu/~jwooders/ebay.pdf.

Paul Resnick (presnick@umich.edu) is an associate professor atthe University of Michigan School of Information in Ann Arbor.Richard Zeckhauser (richard_zeckhauser@harvard.edu) is theFrank P. Ramsey Professor of Political Economy at Harvard University's John F. Kennedy School of Government.Eric Friedman (friedman@economics.rutgers.edu) is an assistantprofessor in the Department of Economics at Rutgers University inNew Brunswick, NJ.Ko Kuwabara (kokopuff@umich.edu) is a Ph.D. student at theUniversity of Michigan School of Information in Ann Arbor.

This work is supported in part by the National Science Foundation under grant IIS-9977999.

© 2000 ACM 0002-0782/00/1200 $5.00

c

48 December 2000/Vol. 43, No. 12 COMMUNICATIONS OF THE ACM