Should Gig Platforms Decentralize Dispute Resolution?

1. Introduction

Online labor platforms have experienced tremendous growth over the years. According to a study by Katz and Krueger (2019), online labor platforms account for more than 90% of net employment growth in the United States between 2005 and 2015. These platforms host a wide range of tasks, such as website and graphic design (e.g., Upwork and Freelancer.com), programming (e.g., PeoplePerHour), home services (e.g., TaskRabbit), and personal assistance for daily or ad hoc needs (e.g., Fancy Hands). These online labor platforms are part of the broader “gig economy,” where they leverage the expertise of freelancers and provide them with flexible work opportunities. Freelancers can often work remotely and on their own schedule, increasing flexibility for businesses and facilitating connections between suitable tasks and available freelancers. The recent coronavirus disease 2019 pandemic has further amplified the use of online labor platforms, resulting in a 47% increase in employers hiring freelancers from these platforms as they adapt to remote work and seek more flexible workers to navigate evolving business landscapes (Upwork 2020).

However, despite the convenience of connecting clients with freelancers, online labor platforms are not without drawbacks. Concerns often arise regarding the quality of freelancers’ work, as their skill levels may be lower compared with professional employees, and many platforms lack stringent screening processes to ensure quality. To address this concern, online labor platforms often allow clients to reject payment to freelancers even after the work is completed (Aloisi 2016). However, this leads to a potential contention for payment disputes. Moreover, the lack of face-to-face interaction can result in more disagreements over the quality of work delivered. As a result, disputes can be a common occurrence on online labor platforms (DeVault et al. 2019). Traditionally, the resolution of disputes has relied on the platform acting as an arbitrator (Aloisi 2016). Although the platform retains more decision-making power in this case, its arbitration may be biased toward a particular side because of a conflict of interest. We term this dispute resolution method the “centralized dispute system.”

Because of concerns about the centralized nature of traditional platforms’ dispute arbitration, emerging platforms, such as LibertyLance and Ethearnal, are introducing a novel tribunal system to ensure fairer dispute settlements. Under this system, members of the platform are allowed to participate in the tribunal to assess and vote on dispute cases between clients and freelancers. When a dispute occurs, a tribunal is formed, and the side with the majority vote wins the case. This way, the platform outsources dispute resolution to the tribunal and crowdsources justice from its members with heterogeneous views. To implement the tribunal system, these emerging platforms have developed a self-funded mechanism where each voter is required to deposit a certain amount of money to participate in the tribunal. The total deposit forms a reward pool, which is then distributed to the winning side of voters based on the majority rule. Consequently, the voters who vote with the majority will earn a monetary reward, whereas the voters who vote with the minority will lose their stake to the winning side. We term this dispute resolution method the “decentralized dispute system.”

The decentralized dispute system offers several advantages. First, it is more cost-effective as the platform outsources dispute resolution to the tribunal, freeing up resources that would otherwise be invested in resolving disputes (CryptoTask 2022). Second, by crowdsourcing dispute resolution to the public, the dispute system becomes more scalable, as highlighted in an interview with the Chief Executive Officer of an emerging platform (Rasheed 2021). This means that the system can handle a potential increase in disputes from an expanding user base. Third, the tribunal system can resolve disputes in a more timely manner if the voting process is limited to a short time window, typically 24 hours, compared with traditional platforms that can take several weeks (Hyve 2023). In addition to these benefits, the emerging platforms also believe that the decentralized dispute system is fairer because disputes are resolved by a group of users rather than a single entity (Ethearnal 2021). Furthermore, these platforms randomly assign voters to individual dispute cases, so that the platform users do not know a priori which dispute case they will be assigned to. This helps minimize the possibility of collusion and ensure a fair voting outcome.

However, the decentralized dispute mechanism is not without risks. The monetary reward introduced by the self-funded mechanism may potentially incentivize the voters to vote strategically in order to win rather than achieve justice. Thus, it is not clear whether the existence of a monetary incentive can still elicit a fair judgment from the voters and whether the voters’ strategic voting decisions can still effectively incentivize quality improvement from the freelancers. Motivated by the recent industry practice, we aim to answer the following research questions in this paper. First, in light of the concern of strategic voting, can the decentralized dispute system achieve justice in dispute resolution, and under what conditions? Second, when can the decentralized dispute system be more profitable for the platform? Third, does the decentralized dispute system benefit the platform without hurting the social welfare?

We model a platform that intermediates the transaction between a client and a freelancer. The client offers a contract price to the freelancer, and the freelancer chooses the quality of work. Upon receiving the work, the client has the right to reject payment to the freelancer, in which case the freelancer can initiate a dispute by paying the dispute fee to the platform. If the centralized dispute system is adopted, the dispute outcome will be determined by the platform. On the other hand, if the decentralized dispute system is adopted, a tribunal consisting of independent platform users will be formed to vote on the dispute case. Each member of the tribunal deposits a participation fee, and those who vote for the winning side will share the total deposit. The voters can have different but correlated evaluations of the freelancer’s work quality and can vote strategically by taking into account the chance of winning, for which they need to form beliefs about the other voters’ votes without any communication among themselves. This makes the tribunal’s voting game a coordination game, which we model using the global games framework (Carlsson and van Damme 1993, Morris and Shin 1998). If the freelancer wins the dispute so that the client has to pay the freelancer or if the client accepts the freelancer’s work in the first place, the platform earns a percentage commission from the contract price. Thus, the platform can earn revenues from two sources: (1) extracting commissions and (2) charging dispute fees.

First, by examining the decentralized dispute system, we find that despite the voters’ strategic motive to coordinate with each other, the tribunal’s voting mechanism can ensure a just dispute resolution outcome if the degree of subjectivity in the voters’ judgments is sufficiently high. In this case, the voters can coordinate on a fair equilibrium where they evaluate the freelancer’s work quality according to a publicly recognized industry standard. However, if the degree of subjectivity is low, the voters’ judgments are likely to be similar, resulting in a situation where they follow each other’s decisions rather than adhering to a fair industry standard. This not only makes an unfair equilibrium possible but also creates multiple equilibria, resulting in a coordination failure in the voting game. Therefore, the tribunal’s ability to elicit heterogeneous assessments from its members is critical for the decentralized dispute system to function effectively, and the platform should be cautious about any policy that may homogenize voters’ judgments. Moreover, achieving the fair equilibrium in the tribunal indicates that the decentralized dispute system can effectively eliminate the platform’s decision-making bias that arises under the centralized dispute system. Because the platform will earn the commission if the freelancer wins the dispute, such a conflict of interest induces the platform to set a lower quality standard to rule in favor of the freelancer. By eliminating this bias, the decentralized dispute system requires a more stringent quality standard for the freelancer.

Next, by comparing the platform’s equilibrium utilities under the two dispute systems, we find that the decentralized dispute system is more profitable as long as the freelancer’s skill level is sufficiently high. The higher quality standard under the decentralized dispute system has different implications for different types of freelancers. If the freelancer is sufficiently skilled to meet the higher quality standard, the client will be willing to offer a higher contract price, and the platform will be able to extract more surplus from the transaction. However, if the freelancer’s skill level is insufficient, the higher quality standard will make it even more difficult for the freelancer to participate. In this case, the centralized dispute system enables the platform to adapt to a lower quality standard and allow more lower-skilled freelancers to participate. Therefore, in order to reap any benefit from the decentralized dispute system, platforms should ensure the skill level of the freelancer pool, which can be achieved by using certification or providing training programs.

Finally, although the decentralized dispute system strengthens the freelancer’s incentive to improve quality by eliminating the platform’s decision-making bias, the induced equilibrium quality level is guaranteed to be more socially optimal only if the voters are not too strict. If the tribunal’s judging criterion is overly stringent, the freelancer will be forced to provide a quality level that is too high compared with the socially optimal level. Therefore, educating the general public to form a proper standard can ensure that the decentralized dispute system benefits the platform while inducing a more socially optimal equilibrium outcome, achieving a win-win solution for gig platforms and policy makers.

2. Literature Review

Our paper is related to three streams of literature: (1) voting games, (2) platform operations, and (3) dispute management. First, to model the voting decisions of the tribunal members, we utilize the global games framework, which was pioneered by Carlsson and van Damme (1993) and Morris and Shin (1998). “Global games” is a class of games where each player observes a private but correlated signal of the fundamental state and has to account for the other players’ beliefs when deciding his or her own action. This framework has been used in the economics literature to capture the strategic interactions between decisions makers in voting game settings and has recently been used by operations management researchers. For example, Wang et al. (2021) study the strategic interactions between firms in the adoption of green technology. By using the global games framework, each firm observes a private but correlated benefit of the new green technology when making the adoption decision, and their payoffs are influenced by the other firms’ decisions as the regulator will make green technology adoption mandatory based on the proportion of firms that voluntarily adopt it. In a similar spirit, we apply the global games framework to study the strategic interactions between the tribunal members under the decentralized dispute system. In our model, each voter observes a private but correlated sentiment of the freelancer’s work. Because a voter’s payoff is influenced by whether his vote belongs to the majority of votes, he needs to reason on the other voters’ perception of the freelancer’s work in his own decision-making process. In addition, because voting is a form of group-based decision making, our paper is related to other settings that involve group-based decision making, such as crowdfunding (e.g., Xu and Zhang 2018, Belavina et al. 2020, Chakraborty and Swinney 2021, Chakraborty et al. 2023), information acquisition (e.g., Marinesi and Girotra 2013, Tsoukalas and Falk 2020), and team coordination (e.g., Dawande et al. 2019, Roels and Corbett 2021).

The operations management literature has so far studied several issues in platform operations, such as wage schemes (e.g., Taylor 2018, Bai et al. 2019, Hu and Zhou 2020), surge pricing (e.g., Cachon et al. 2017, Guda and Subramanian 2019, Hu et al. 2022), platform competition (e.g., Bernstein et al. 2021, Chen et al. 2022), and labor welfare (e.g., Benjaafar et al. 2019, 2022). Papers in this stream have typically considered settings where the platform sets the wages for the freelancers. Our setting differs in that a freelancer’s wage is determined by a contracting process between the client and the freelancer, which is typical for online labor platforms. Although contracting has been extensively studied in supply chain settings (see Elmaghraby 2000 for a review), our platform setting is distinguished by the unique feature that the contract between the client and the freelancer is mediated by the platform. Moreover, because the contracts on online labor platforms are signed between individual users on small-scale short-term projects, they are less formal in nature compared with supply chain contracts, leading to a higher likelihood of user disputes.

Dispute management in informal contracting settings has recently gained interest in operations management research. One such study by Papanastasiou et al. (2023) considers an e-commerce seller’s dispute over a customer review. They examine when a semidecentralized dispute system that allows the seller to remove a customer’s review subject to potential checks by the platform can be more efficient than the centralized dispute system where the platform decides whether a customer review should be removed. Our study differs in two ways. First, we study disputes that arise from payment rejections rather than from untruthful customer reviews. As a result, the reasons for initiating a dispute and the underlying economic dynamics are different. Second, we study a fully decentralized mechanism of dispute resolution, where independent platform users are involved to vote on a dispute case, and the platform fully relegates the authority to arbitrate dispute cases. Another study by Kwan et al. (2023) considers a crowd judging-based dispute system, which is more similar to ours. They use a data set from Taobao to empirically demonstrate that the bias in crowd judging tends to decrease as jurors gain experience. Our paper, on the other hand, considers a different crowd-judging mechanism where voters can earn a monetary payoff based on the voting outcome and hence, can act strategically. We provide the analytical insight that ensuring the heterogeneity of voter assessments plays a critical role in achieving justice of the voting outcome. In addition, our paper compares the performance of the decentralized dispute system (judged by the public) with the centralized dispute system (judged by the platform) and prescribes when the decentralized dispute system should be preferred by the platform and the social planner.

3. Model Setup

We consider a platform that serves as an intermediary between a client (referred to as “her”) and a freelancer (referred to as “him”). The platform’s revenue consists of two components. First, the platform extracts a commission equal to γ fraction of the contract price. The commission is paid by the freelancer, who is the party receiving the payment. We assume that the commission rate is exogenously given by the industry norm, whereas the contract price is endogenously determined by the contracting process between the client and the freelancer. The commission rate of online labor platforms typically falls in a narrow range of 10%–20%. For example, the commission rate is 10% for Freelancer.com, 15% for TaskRabbit, and 20% for PeoplePerHour and Upwork. Second, the platform charges a dispute fee $f\ge 0$, which is paid by the freelancer if he decides to initiate a dispute. The dispute fee is endogenously chosen by the platform and is nonrefundable regardless of the dispute outcome (e.g., PeoplePerHour). The dispute fee charged by online labor platforms varies considerably. For example, PeoplePerHour charges either $8 or 10% of the contract price, whichever is higher, whereas Upwork charges a flat fee of $200 regardless of the contract price. Nevertheless, the magnitude of the dispute fee is comparable with the commission fee, making it a significant source of revenue for these platforms. In our main model, we assume the platform charges the dispute fee only to the freelancer, who is the party initiating the dispute. In E-Companion EC.5, we consider an alternative setting where the platform charges the dispute fee to both the client and the freelancer.¹

The freelancer chooses the work quality $q\ge 0$ and faces an effort cost $\alpha q^2$, which is a convexly increasing function of q, where α represents the efficiency of the freelancer (Ha et al. 2016). A freelancer with a smaller α can achieve the same quality q at a lower cost; hence, his skill level is higher (Banker et al. 1998). For example, an experienced graphic designer can easily design a logo because of familiarity with professional software, such as Adobe Photoshop, whereas a less experienced designer may need to invest time in learning the software before starting the project, resulting in increased costs. We assume that α is public information because it is common practice for online labor platforms to have a reputation system for the users, and a freelancer’s skill level can be inferred from his rating and historical reviews (Jin et al. 2022). Note that because we consider a single freelancer in our main model, α can be interpreted as the skill level of the freelancer pool. In E-Companion EC.7, we consider a model extension where the freelancer pool can have heterogeneous skill levels.

The client’s valuation of the work is assumed to be equivalent to the work quality q. This assumption is made without loss of generality as our results remain valid even if the client’s valuation can be different from q but follows an increasing linear function of q. When the client offers a contract, she chooses the contract price $p\ge 0$. Figures EC.1.5 and EC.1.6 in E-Companion EC.1 provide examples of gig platforms where it is the client, rather than the freelancer, who sets the price when posting the task. Quality is not contractible because it is uncommon for gig work to have a quality expectation formally specified in a contract. Shevchuk and Strebkov (2018) find that only 11% of the tasks rely on formal contracts, whereas the majority rely on verbal or informal correspondence. Even if a formal contract is created, specifying a quality standard can be challenging as textual descriptions can be ambiguous and task objectives are subjective in nature. Because a quality standard is not specified, the client can reject the freelancer’s work and refuse to pay, in which case the freelancer has the option to file a dispute. The client obtains the valuation of work, regardless of whether she rejects it or not, to reflect the nature of gig work that it typically involves providing a service that cannot be returned, like a physical product. In our model, payment rejection and dispute apply to the entire contract price, corresponding to a fixed-price contract where disputes can only be filed on the entire contract. It is worth mentioning that online labor markets use fixed-price contracts for the majority of tasks (Liang et al. 2016), as they are short-term small-scale projects. Some platforms (e.g., Upwork and Freelancer.com) offer the option of milestone payments for larger-scale projects, which allow disputes on particular milestones. However, each milestone payment can be treated as a separate fixed-price contract, and our insights are also applicable.

The sequence of events is illustrated in Figure 1 and closely parallels the operations of well-established platforms, such as PeoplePerHour. The game consists of six decision-making stages, beginning with the platform choosing the dispute fee f (stage 1). The client then chooses the contract price p to offer to the freelancer (stage 2), and the freelancer subsequently chooses the quality level q (stage 3). After completing the task, the client decides whether to accept or reject the freelancer’s work (stage 4). If the client accepts the freelancer’s work, the transaction concludes, the platform receives the commission γp, and the freelancer receives the payment $(1-\gamma )p$. If the client rejects the work, the freelancer decides whether to initiate a dispute by paying the dispute fee f (stage 5). Finally, if the freelancer chooses to initiate a dispute, a dispute resolution process is invoked (stage 6). If the freelancer wins the dispute, he receives the payment, and the platform receives the commission. The platform, the client, and the freelancer are all forward-looking decision makers.

We consider two types of dispute systems: the centralized and decentralized dispute systems. The difference between the two systems lies only in stage 6 of the sequence of events. Specifically, under the centralized dispute system (Section 4), the platform is responsible for carrying out the dispute resolution, whereas under the decentralized dispute system (Section 5), a tribunal is responsible. We describe the models of the dispute resolution processes for both systems in subsequent sections, and E-Companion EC.1 provides examples of gig platforms’ dispute policies for each system. Because evaluations of gig work tend to be subjective, an external evaluator (such as the platform or the voters in a tribunal) may form a different evaluation of the freelancer’s work than the client. In addition, the true quality of work is only known to the client and the freelancer (i.e., the contracting parties) but not observable to an external evaluator (Baker et al. 1994, Levin 2003). As we will show, the outcome of dispute resolution yields the freelancer a probability of winning the dispute, which we denote as h(q). We use superscripts of $*$ and + to differentiate between the equilibria of the centralized and decentralized dispute systems, respectively. The proofs of results for the main model are presented in Online Appendices A–C.

4. Centralized Dispute System

In this section, we study the centralized dispute system, where disputes are resolved by the platform. In Section 4.1, we first delve into the platform’s decision-making process of dispute resolution. Then, building on the platform’s dispute decision, we proceed to analyze the overall game between the client, the freelancer, and the platform in Section 4.2.

4.1. Platform’s Dispute Decision

We first model and analyze the subgame of the platform’s dispute resolution given that a dispute has been initiated. In reality, a platform would make such a decision by evaluating whether the freelancer’s work quality is sufficiently high to justify payment. To mimic this decision-making process, we assume that the platform chooses a quality threshold k and compares the evaluated quality of the freelancer’s work with this threshold. The platform’s quality threshold is not announced up front but chosen when a dispute has occurred to reflect the reality that it could be difficult for the platform to credibly commit to a quality evaluation criterion. However, the platform’s dispute arbitration is binding. In E-Companion EC.3, we consider a model extension where the platform’s dispute arbitration is nonbinding, and the users can appeal the platform’s dispute decision, in which case a third-party arbitration will be involved to make a binding decision.

The outcome of dispute resolution can be influenced by the subjective judgment of the evaluator (Taylor and Yildirim 2011, Deb et al. 2016), who is typically a member of the platform’s dispute team. For example, even if two logos are designed by the same freelancer with the same amount of effort, they may receive different evaluations from the evaluator because of the evaluator’s personal aesthetic preferences. Because the exact preference of the evaluator is unknown until the evaluation is made, there is variability in the evaluated quality. We assume that the evaluator’s subjective judgment causes the evaluated quality to be a random variable around the freelancer’s true quality q. The platform’s evaluated quality is represented by the evaluator’s judgment given by a random variable $x=q+\sigma ϵ$, where ϵ is uniformly distributed over $[-1,1]$ and σ is a scaling factor that measures the degree of subjectivity in the platform’s evaluation. Consequently, the platform will rule in favor of the freelancer if $x\ge k$ and in favor of the client if x < k. Let h(q, k) denote the resulting probability for the freelancer to win the dispute following this decision-making rule.

Moreover, the platform’s dispute decision can be affected by its concern about how closely its judgment aligns with the industry norm. A decision that deviates significantly from the industry norm may invite heavy criticism and damage the platform’s reputation, which can, in turn, negatively affect its future revenue. For example, setting a standard that is too high may cause the platform to lose freelancers, whereas setting it too low may discourage client participation. To account for this, we introduce a disutility term that represents the platform’s cost of deviating from the industry norm. Specifically, we model this term as $-\theta {(k-y)}^2$, which is a quadratic function of the distance between the platform’s threshold k and the industry standard y. The industry standard y represents the quality standard expected by the general public and serves as an anchoring point for a “fair” judgment in our model (Chen 2022). In other words, a quality evaluation criterion is deemed fair if it conforms to the industry norm. It is easy to see that a higher value of y indicates that the public holds a stricter standard, which means that even under fair judgment, the freelancer would have to put in more effort. In addition, $\theta >0$ represents the degree of penalization when the platform deviates from the industry norm. The quadratic term captures the two-sided impact that the platform’s long-term reputation can be negatively affected if it unfairly rules in favor of either side of the market (Tsoukalas and Falk 2020).

Therefore, the platform’s utility comprises two components:

As shown by Equation (1), the first component is the monetary payoff from earning commissions and dispute fees, and the second component is the disutility if the platform deviates from the industry norm in its dispute resolution. Lemma 1 characterizes the platform’s equilibrium quality threshold and the resulting winning probability of the freelancer for the dispute resolution subgame.

Lemma 1 shows that the platform’s quality threshold $k^{*}$ is always lower than the industry standard y and is dependent on the magnitude of the commission fee γp. If dispute occurs and the platform rules in favor of the freelancer, it is able to earn the commission fee in addition to the dispute fee. Thus, the platform intentionally sets a lower quality threshold than the industry standard (i.e., $k^{*}\le y$) to increase the winning probability of the freelancer. As a result, the probability of the freelancer winning the dispute comprises two components: (1) the probability that the platform’s evaluated quality is above the industry standard y $($i.e., $\frac{q-y+\sigma }{2\sigma })$ and (2) a positive bias term that increases his chance of winning the dispute $($i.e., $\frac{\gamma p}{8\theta {\sigma }^2})$. Furthermore, the bias term increases if the commission fee is higher and decreases if the degree of penalization θ is higher.

Thus, serving as both the intermediary of transactions and the arbitrator of disputes, centralized dispute resolution will inevitably lead to a conflict of interest in the platform’s decision-making process. The platform’s inclination to let the freelancer win has been observed by the clients who have gone through the dispute process (Tzezana 2015, Chen 2017, Chris 2021). These clients warn that arbitration rarely rules in favor of the client. Moreover, they complain that the unfair dispute resolution gives rise to freelancers relying on dispute to earn their profit instead of putting in effort to improve their work quality. As our model highlights, such an issue is rooted in the platform’s decision-making bias caused by the centralized mechanism of dispute resolution.

4.2. Contracting Equilibrium Under the Centralized Dispute System

Given the equilibrium of the dispute resolution subgame obtained from Section 4.1, we continue the backward induction process of the overall game and solve for the contracting equilibrium between the client and the freelancer as well as the platform’s equilibrium dispute fee. Using the equilibrium winning probability of the freelancer if dispute occurs, $h^{*}(q)$, we can express the utilities of all decision makers as follows. In particular, under the centralized dispute system, if both the client and the freelancer participate, the freelancer’s utility is given by

As the client is forward looking, her decision to accept or reject the freelancer’s work takes into account the platform’s dispute judging criterion, which results in the client rejecting the freelancer’s work in equilibrium if and only if the freelancer’s quality is below a threshold.² Similarly, given that the client has rejected the freelancer’s work, the freelancer’s decision to initiate a dispute or not also takes into account the platform’s dispute judging criterion, which results in him initiating a dispute if and only if his quality is above a threshold. Finally, the platform’s utility is given by

Proposition 1 characterizes the equilibrium under the centralized dispute system.

Proposition 1(i) shows that the equilibrium can fall into three regimes depending on the skill level of the freelancer. First, if the freelancer’s skill level is high (i.e., $\alpha \le {\underset{¯}{\alpha }}_c$), contracting occurs without dispute. Second, if the freelancer’s skill level is medium (i.e., ${\underset{¯}{\alpha }}_c<\alpha \le {\overline{\alpha }}_c$), contracting occurs with dispute. Third, if the freelancer’s skill level is low (i.e., $\alpha >{\overline{\alpha }}_c$), contracting does not occur. For the first two regimes where contracting occurs, Proposition 1, (ii) and (iii) characterizes the equilibrium decisions of all decision makers, including the platform, the client, and the freelancer.

If the freelancer’s skill level is sufficiently high (i.e., $\alpha \le {\underset{¯}{\alpha }}_c$), he is able to choose a quality level that is sufficiently high to guarantee winning the dispute. This indicates that the client does not have any incentive to reject the freelancer’s work and run into a dispute. As a result, the client will accept the freelancer’s work, and dispute does not occur in equilibrium. In this case, the freelancer does not pay any dispute fee in equilibrium. However, the dispute fee influences the equilibrium contract price. In order to incentivize the freelancer to participate, the client’s contract price needs to cover both the commission fee and the dispute fee (i.e., $(1-\gamma )p\ge f$). The freelancer will not participate if the client’s contract price net of commission is less than the dispute fee because if the client deviates from the equilibrium by rejecting to pay, the freelancer will not be able to afford the dispute. Thus, the client’s equilibrium contract price satisfies $p^{*}=\frac{f^{*}}{1-\gamma }$, which indicates that the platform can use the dispute fee to nudge the client to offer a higher price to the freelancer. In this case, the platform’s equilibrium utility comprises solely the commission revenue.

If the freelancer’s skill level is medium (i.e., ${\underset{¯}{\alpha }}_c<\alpha \le {\overline{\alpha }}_c$), although contracting still occurs, it will lead to a dispute in equilibrium. Dispute occurs in this case because it is too costly for a freelancer with an insufficient skill level to choose a sufficiently high quality level to guarantee winning the dispute. Anticipating this, the client will reject the freelancer’s work so that she will not have to pay if the freelancer ends up losing the dispute. Correspondingly, the freelancer will respond by initiating the dispute because he can only earn the contract price by winning the dispute. In this case, the platform’s utility comprises both the commission fee and the dispute fee, as well as the disutility because of its deviation from the industry norm in dispute resolution. Moreover, it is increasingly costly for a lower-skilled freelancer to participate because of the higher quality cost and lower probability of winning the dispute. Therefore, if the freelancer’s skill level is too low (i.e., $\alpha >{\overline{\alpha }}_c$), he is not able to participate.

5. Decentralized Dispute System

In this section, we study the decentralized dispute system, where disputes are resolved by a tribunal comprising a separate group of platform users. Each member of the tribunal votes between the freelancer and the client, and the decision of the tribunal is based on the majority rule. In Section 5.1, we first introduce the model of the voting game. Then, in Section 5.2, we derive the equilibrium of the voting game. Finally, building on the tribunal’s voting equilibrium, we proceed to analyze the overall game between the client, the freelancer, and the platform in Section 5.3.

5.1. Tribunal’s Voting Game

We first consider the subgame of the tribunal’s dispute resolution given that a dispute has been initiated. Recall that an important feature introduced by the emerging platforms is a self-funded mechanism where the voters who vote for the winning side will share the total deposit from all voters. This can create an incentive for voters to prioritize earning more monetary reward in their decisions rather than voting solely based on the evaluation of the freelancer’s work quality. Furthermore, in order to earn the monetary reward, a voter must ensure that he votes for the winning side, which means that his vote coincides with the majority of votes. Thus, the voters’ decisions become strategic complements of each other (i.e., a voter’s relative gain from voting for a particular side over the other side increases when the proportion of voters voting for that side increases), leading to a coordination game. Such coordination would require a voter to strategically anticipate the decisions of other voters when making his own decision, without knowing who the other voters are.

To model the strategic interaction among the voters in the tribunal, we adopt the global games framework commonly used for studying coordination games (e.g., Morris and Shin 1998, Edmond 2013, Rundlett and Svolik 2016). This framework assumes that players receive private but correlated signals of an underlying fundamental, and their payoffs are jointly determined by the decisions of all players. Consistent with the global games framework, we model the voters in the tribunal as a continuum with total mass normalized to one, approximating a large but finite number of voters and invoking the law of large numbers. In our setting, the signal that voter i privately receives is his own evaluation of the freelancer’s work quality, which is modeled as $x_i=q+\sigma {ϵ}_i$, where q is the true quality unobservable to the voters, ϵ_i is a random variable uniformly distributed over $[-1,1]$, and σ is the scaling factor that measures the degree of subjectivity in the voters’ evaluations. We assume that the degree of subjectivity is the same for any external evaluator, including the platform’s dispute team members and the voters in a tribunal, as it is a characteristic of the general population (Rowe and Wright 2001). In E-Companion EC.6, we explore a model extension where the degree of subjectivity can be different under the two systems.

Different from the platform’s dispute decision, a voter’s decision needs to factor in his inferences about other voters’ decisions, which hinge on his belief about the true quality of the freelancer’s work. Following the global games framework, we assume that the voters initially have a uniform prior on the true quality of work, and after observing their own evaluated quality, the voters update their belief about the true quality of work and draw inferences about the decisions of other voters. As a result, if a voter receives a high signal, he is likely to believe that the other voters receive high signals as well. It is worth mentioning that the global games framework requires that the voters cannot communicate with each other so that they cannot collude in their voting decisions. The emerging platforms have adopted several measures to prevent communication among the voters. (1) The voters are randomly assigned to dispute cases, so they do not know a priori which case they will be assigned to; (2) the platform does not establish any communication channels for the voters to communicate with each other; and (3) the time window for the voters to submit their votes is typically limited to 24 hours, which further reduces the chance of finding and communicating with other voters in the same tribunal. These measures lead to an opacity in the composition of the tribunal, which helps to prevent collusion among the voters and also makes the global games framework applicable.

To gain the right to vote in the tribunal, each voter must pay a participation fee of t. The total participation fees are pooled together to form the reward pool for the dispute case. Voters who are part of the majority votes will win an even share of the total reward. Consistent with the platform’s decision under the centralized dispute system, voter i chooses a quality threshold k_i such that he votes for the freelancer if $x_i\ge k_i$ and votes for the client if x_i < k_i. Given that l proportion of voters votes for the freelancer, the utility of voter i, $u_i(x_i,k_i,l)$, is as follows:

To determine his threshold strategy k_i, voter i needs to consider all possible values of x_i and his voting decision given that each x_i should maximize his expected utility by taking expectation of Equation (5) based on his belief about l.

As Equation (5) shows, similar to the platform’s utility, a voter’s utility comprises two components. The first component is the voter’s monetary payoff if voting for the winning side, which is the difference between the share of the total reward and the participation fee. For example, when the majority of the voters vote for the freelancer (i.e., $l\ge 0.5$), a voter will receive a reward of $t/l$ if he also votes for the freelancer (i.e., if $x_i\ge k_i$) and zero if he votes for the client (i.e., if x_i < k_i). The second component corresponds to a disutility of “guilt,” which may weigh on the voter’s conscience if he does not let a worthy freelancer win and is weighted by a factor of $\xi >0$. This disutility term mirrors the platform’s disutility of deviating from the industry norm and is modeled consistently as a quadratic function of the distance between voter i’s threshold k_i and the industry standard y. The inclusion of such a nonmonetary component in a voter’s utility follows the convention of voting game models, where the voters receive a disutility for convicting an innocent party or acquitting a guilty party (e.g., Feddersen and Pesendorfer 1998, Kojima and Takagi 2010). Moreover, a utility model with both monetary and nonmonetary components is commonly used in behavioral modeling (e.g., Falk and Fischbacher 2006; Battigalli and Dufwenberg 2007, 2009).

5.2. When Can Tribunal Voting Achieve Justice?

Given the voting game described in Section 5.1, we now analyze the equilibrium voting strategy of tribunal members and examine under what conditions the tribunal voting mechanism can achieve justice of dispute resolution. As we have seen, the self-funded mechanism creates an incentive for the voters to coordinate their votes, as a voter only receives a monetary reward by being part of the majority. However, it remains unclear if such coordination can be achieved in equilibrium and if all voters can reach agreement on the same quality threshold to follow in their voting decisions. If there exist multiple equilibria corresponding to different thresholds, the voting game may result in a coordination failure (Van Huyck et al. 2002). Moreover, even if the voters can coordinate their decisions by agreeing on the same threshold to follow, it is still unclear if the outcome of their coordination can be a fair equilibrium where all voters follow the industry standard (i.e., k_i = y).

To ensure that the decentralized dispute system achieves justice, two conditions must be satisfied. First, a fair equilibrium must exist and arise as an equilibrium of the voting game. Second, it must be the unique equilibrium of the voting game. These conditions together guarantee that the tribunal voting will follow the industry standard and achieve a fair resolution of disputes. Therefore, we derive conditions for the existence and uniqueness of the fair equilibrium. Lemma 2 first shows the existence of the fair equilibrium and derives the resulting winning probability of the freelancer under this equilibrium.

Lemma 2 shows that the fair equilibrium always exists in the tribunal’s voting game. In this case, each voter will vote for the freelancer if the evaluated quality is higher than y and will vote for the client otherwise. Collectively, the voters’ voting strategies form a probability distribution, such that as the freelancer’s quality q increases, his winning probably $h^{+}(q)$ also increases.³ Moreover, by comparing $h^{+}(q)$ to the freelancer’s winning probability $h^{*}(q)$ under the centralized dispute system (Lemma 1), we observe that given the same quality level, the freelancer has a lower chance of winning the dispute under the decentralized dispute system. The decentralized dispute system removes the platform’s bias of ruling in favor of the freelancer in order to earn more commissions, and the industry norm serves as a focal point for the tribunal’s voting. This results in a higher quality standard that is expected of a freelancer under the decentralized dispute system, and the freelancer has to offer a higher quality level in order to achieve the same winning probability.

Although the voters’ strategic decision making to earn more monetary rewards does not necessarily prevent the fair equilibrium from being an equilibrium of the voting game, its mere existence is not sufficient to guarantee a just dispute resolution outcome by the tribunal. If an unfair equilibrium, corresponding to a different threshold than the industry standard, exists simultaneously, a coordination failure may occur, making it impossible to ensure the tribunal selects the fair equilibrium. Therefore, we next establish a condition for the fair equilibrium identified in Lemma 2 to become the unique equilibrium of the tribunal’s voting game, which is given by Proposition 2.

The condition established in Proposition 2 requires a sufficiently high degree of subjectivity in the voters’ evaluations, as indicated by the requirement for a sufficiently large σ. To prevent any unfair equilibrium, the voters must have an incentive to deviate to a fairer equilibrium by moving their quality threshold closer to the industry standard, thereby reducing the disutility of guilt they incur. However, changing the threshold from that of the majority of voters also reduces their probability of being on the winning side, which in turn, decreases their monetary payoff. Whether the voters have an incentive to deviate to a fairer equilibrium is determined by the trade-off between the incentive to reduce guilt (i.e., the gain of being fair) and the incentive to lose less monetary reward (i.e., the cost of being fair).

If the degree of subjectivity is high, voters tend to have more heterogeneous judgments for a given dispute case. If a voter changes his threshold, the probability of flipping from the winning to the losing side is small because the majority of voters are unlikely to fall between the old and new thresholds of that voter. Consequently, a voter’s loss in monetary reward is relatively insensitive to the quality threshold he chooses, and he is less concerned about not aligning his threshold with that of other voters when other voters are being unfair. Therefore, a higher degree of subjectivity reduces the cost of being fair. When the degree of subjectivity is sufficiently high, the incentive to deviate toward the industry standard is guaranteed to exist for any threshold unequal to the industry standard, making it the only threshold that can sustain as an equilibrium in the voting game.

However, if the degree of subjectivity is low, voters are likely to have similar evaluations for a given dispute case, making their winning probability sensitive to deviations from the threshold chosen by the majority of voters. If the majority of voters have coordinated on a threshold that is not too different from the industry standard, the remaining voters will agree to follow the same threshold and compromise on fairness because it is too costly to deviate toward the industry standard. This creates a greater incentive to conform to other voters’ decisions, making a threshold unequal to the industry standard a possible equilibrium. Additionally, a lower degree of subjectivity enlarges the range of thresholds that the voters can agree on, leading to multiple equilibria and a coordination failure in the voting game. Therefore, a lower degree of subjectivity increases the cost of being fair and reduces the tribunal’s ability to achieve a just dispute resolution outcome.

The condition in Proposition 2 has important implications for how the platform can ensure that the fair equilibrium is the unique equilibrium of the voting game. When adopting the decentralized dispute system, the platform should be cautious about any policy that may reduce the degree of subjectivity in the voters’ judgments, and the value of the decentralized dispute system lies critically in its ability to elicit heterogeneous assessments from the tribunal members. Dispute evaluation guidelines provided by the platform should not aim to homogenize the voters’ evaluations, as a more homogenized voter judgment may incentivize voters to prioritize conforming to each other and forgo their objective to achieve justice.

Moreover, because the degree of subjectivity can be influenced by the nature of the task being evaluated, the condition in Proposition 2 also implies that the tribunal voting mechanism may not work equally well for all types of tasks. For tasks such as home cleaning and programming, objective evaluation criteria are easily defined, and the evaluators’ judgments tend to be less subjective. Platforms should exercise caution when decentralizing dispute resolution in this case, as a lower degree of subjectivity limits the ability of the decentralized disputes system to elicit heterogeneous assessments from the voters. However, for tasks such as design works, objective evaluation criteria are difficult to define, and the evaluators’ judgments rely heavily on their personal preferences, leading to a higher degree of subjectivity. In this case, the decentralized dispute system can maximize its potential to elicit heterogeneous assessments from the voters and ensure a just dispute resolution outcome.

In addition, the condition in Proposition 2 also depends on two other parameters: t, the voters’ participation fee, and ξ, the weight that voters place on fairness consideration. Holding the tribunal’s degree of subjectivity constant, the fair equilibrium is guaranteed to be the unique equilibrium if the voters’ participation fee is sufficiently small or if their weight on fairness consideration is sufficiently high. A lower participation fee reduces the reward pool and the voters’ incentive to conform to each other, whereas a higher weight on fairness consideration strengthens their incentive to deviate toward the industry standard from any threshold unequal to the industry standard. Both changes increase the likelihood of achieving a just dispute resolution outcome. Therefore, the platform should avoid setting a participation fee that is too high and consider educating the voters to prevent them from gaming the system.

Having obtained the measures that the platform can take to induce the fair equilibrium as the unique equilibrium of the tribunal’s voting game, we proceed to analyze the overall game between the client, the freelancer, and the platform, with the fair equilibrium in Lemma 2 as the dispute resolution outcome.

5.3. Contracting Equilibrium Under the Decentralized Dispute System

Given the equilibrium of the tribunal voting subgame obtained from Section 5.2, we continue the backward induction process of the overall game to solve for the contracting equilibrium under the decentralized dispute system. If both the client and the freelancer participate, their utilities are similar to Equations (2) and (3), respectively, with $h^{*}(q)$ replaced by $h^{+}(q)$ (i.e., the freelancer’s probability of winning the dispute under the decentralized dispute system). The platform’s utility is given by

Notice that if dispute occurs, the platform does not incur any disutility from the outcome of dispute resolution when the tribunal’s quality threshold resulting from the voting game is equal to the industry standard (Lemma 2). Proposition 3 characterizes the equilibrium under the decentralized dispute system.

Proposition 3 shows that the equilibrium under the decentralized dispute system can fall into three regimes depending on the skill level of the freelancer, similar to the centralized dispute system; contracting occurs without dispute if $\alpha \le {\underset{¯}{\alpha }}_d$, contracting occurs with dispute if ${\underset{¯}{\alpha }}_d<\alpha \le {\overline{\alpha }}_d$, and contracting does not occur if $\alpha >{\overline{\alpha }}_d$. However, there are notable differences. Most importantly, as we have seen from Section 5.2, the tribunal’s dispute decision eliminates the platform’s decision-making bias that arises under centralized decision making. The resulting differences in the equilibrium can be clearly seen by comparing Proposition 3(iii) with Proposition 1(iii). For example, the platform no longer incurs the disutility, $\frac{{\gamma }^2{(p^{*})}^2}{16\theta {\sigma }^2}$, because of the deviation from the industry norm.

Under the centralized dispute system, the platform’s bias in favor of the freelancer can cause the client to have reservations in offering a higher contract price to the freelancer. This is because an increased contract price can further amplify the platform’s bias, which in turn, creates a counterforce in incentivizing the freelancer to choose a higher quality level. In contrast, the decentralized dispute system eliminates the platform’s bias and raises the quality threshold required for the freelancer to win the dispute. This can make the client more willing to offer a higher contract price. As a result, the incentives of the client and the freelancer are more aligned under the decentralized dispute system.

6. Value of Decentralization

In this section, we compare the centralized and decentralized dispute systems and derive insights regarding the type of markets they each cater to as well as the value of decentralizing dispute resolution. In Section 6.1, we first compare the equilibrium decisions of the client, the freelancer, and the platform. In Section 6.2, we examine when the platform should adopt the decentralized dispute system. In Section 6.3, we examine when the decentralized dispute system can improve social welfare.

6.1. Equilibrium Comparison

As seen in Sections 4 and 5, under both systems, the equilibrium is characterized by three regimes depending on the skill level of the freelancer. We start by comparing the thresholds on the freelancer’s skill level that define the three equilibrium regimes to gain a first understanding of how the decentralized dispute system can change the equilibrium structure.

Theorem 1 shows that the decentralized dispute system reduces the range of the freelancer’s skill level α for contracting to occur and also reduces the range where dispute can be prevented. Recall from Section 4 that the centralized dispute system leads to a decision-making bias of the platform to let the freelancer win the dispute with a higher probability because of its interest to earn the commission. Thus, a lower quality is expected of the freelancer, and it is less costly for the lower-skilled freelancers to participate under the centralized dispute system (i.e., ${\overline{\alpha }}_d\le {\overline{\alpha }}_c$). Moreover, the platform’s bias in favor of the freelancer also causes the client to be more willing to compromise on quality. This pushes the client to accept the freelancer’s work and prevents dispute in more cases (i.e., ${\underset{¯}{\alpha }}_d\le {\underset{¯}{\alpha }}_c$).

The comparisons of the α thresholds are illustrated in Figure 2. To ease our subsequent analyses, we define case 1 as the “high-skill” case where dispute does not occur under either system (i.e., $\alpha \le {\underset{¯}{\alpha }}_d$), case 2 as the “medium-skill” case where dispute occurs under the decentralized dispute system but not under the centralized dispute system (i.e., ${\underset{¯}{\alpha }}_d<\alpha \le {\underset{¯}{\alpha }}_c$), and case 3 as the “low-skill” case where dispute occurs under both systems (i.e., ${\underset{¯}{\alpha }}_c<\alpha \le {\overline{\alpha }}_d$).⁴ We focus on these three cases where contracting occurs under both systems (i.e., $\alpha \le {\overline{\alpha }}_d$) in our subsequent analyses and make references to them in our discussions. Theorem 2 summarizes the comparison of the equilibrium decisions of the freelancer, the client, and the platform between the two systems.

Theorem 2, (i) and (ii) shows that under the same equilibrium regime, such that either dispute occurs under both systems (i.e., case 3) or dispute does not occur under either system (i.e., case 1), the decentralized dispute system induces a higher quality level and contract price compared with the centralized dispute system. As we have seen, the decentralized dispute system eliminates the platform’s bias of ruling in favor of the freelancer and raises the quality threshold for the freelancer to win the dispute. The freelancer needs to factor in how stringent the quality standard is in order to win the dispute (i.e., case 3) or prevent the dispute (i.e., case 1). In both cases, the higher standard of dispute resolution increases the freelancer’s incentive to improve quality under the decentralized dispute system, and hence, the client is willing to offer a higher contract price. However, it is possible for the decentralized dispute system to induce a lower quality level and contract price when the equilibrium regimes are different under the two systems. In case 2, dispute only occurs under the decentralized dispute system but not under the centralized dispute system. This means that the freelancer has to pay the dispute fee under the decentralized dispute system, which would constrain how much effort he can expend to improve quality. We further find that when the freelancer’s skill level is relatively low within case 2 (i.e., ${\alpha }_q<\alpha <{\underset{¯}{\alpha }}_c$), this constraint will outweigh the quality-improving incentive because of the elimination of the platform’s bias, leading to a lower equilibrium quality level under the decentralized dispute system. Correspondingly, the client would offer a lower contract price.

We next turn to the platform’s dispute fee. Theorem 2(iii) shows that when dispute occurs under both systems (i.e., case 3), the equilibrium dispute fee is higher under the decentralized dispute system if the freelancer’s skill level is relatively high (i.e., $\alpha \le {\alpha }_f$) and is lower if the freelancer’s skill level is relatively low (i.e., $\alpha >{\alpha }_f$). As the freelancer’s skill level decreases (i.e., α increases), his equilibrium quality level decreases, and it is increasingly difficult for the him to win the dispute. Nevertheless, the platform can reduce its quality threshold under the centralized dispute system, whereas it does not decide how the tribunal would judge the freelancer’s work under the decentralized dispute system. Thus, as the freelancer’s skill level decreases, his equilibrium probability of winning the dispute would decrease at a slower rate under the centralized dispute system. This indicates that a lower-skilled freelancer will receive a greater advantage under the centralized dispute system and is hence willing to pay a higher dispute fee relative to the decentralized dispute system. Thus, when the freelancer’s skill level decreases below a certain threshold (i.e., α_f), the platform would be able to charge a higher dispute fee under the centralized dispute system.

6.2. When Is the Decentralized Dispute System More Profitable?

To gain an insight into how the decentralized dispute system performs with respect to the centralized dispute system, we next compare the platforms’ equilibrium utilities under the two dispute systems to uncover which system the platform should choose.

Theorem 3 characterizes, under the conditions of $\gamma \le 1/3$ and $\sigma \le y$, that the platform can achieve a higher utility under the decentralized dispute system only when the freelancer’s skill level is sufficiently high (i.e., $\alpha \le \overline{\alpha }$). We first note that the conditions in Theorem 3 are unlikely to eliminate scenarios that are practically relevant. First, a commission rate higher than $1/3$ is uncommon for online labor platforms. Second, given that the evaluated quality (by both the platform and the voters) follows a uniform distribution over $[q-\sigma ,q+\sigma ],\hspace{0.17em}\sigma \le y$ ensures that if the freelancer’s quality level is equal to the industry standard y, the evaluated quality is nonnegative. In addition, in E-Companion EC.2, we show numerically that Theorem 3 still holds even when these conditions are not imposed. Moreover, if ${\overline{\alpha }}_d<\alpha \le {\overline{\alpha }}_c$ (i.e., the region beyond case 3), contracting only occurs under the centralized dispute system; hence, the centralized dispute system dominates the decentralized dispute system. Thus, the result in Theorem 3 extends to the region beyond the three cases of interest (i.e., $\alpha \le {\overline{\alpha }}_d$).

Because ${\underset{¯}{\alpha }}_d\le \overline{\alpha }\le {\overline{\alpha }}_d$, the threshold $\overline{\alpha }$ can be achieved in case 2 or case 3 but not in case 1. This immediately indicates that when dispute does not occur under either system (i.e., case 1), the platform is better off with the decentralized dispute system. If the freelancer’s skill level is sufficiently high, dispute does not occur, and the platform only earns the commission fee. As discussed previously in Theorem 2, the elimination of the platform’s bias under the decentralized dispute system induces the freelancer to provide a higher quality level and the client to offer a higher contract price. Thus, the platform can extract more commission under the decentralized dispute system.

If the freelancer’s skill level is not sufficiently high, dispute can occur (i.e., cases 2 and 3). When dispute occurs, the platform’s revenue structure changes to one that depends on both the dispute fee and the commission fee. The platform is guaranteed to earn the dispute fee as long as dispute occurs but only earns the commission fee if the freelancer wins the dispute. Depending on the skill level of the freelancer, the platform can extract more surplus from the participants through different means. When the freelancer’s skill level is relatively high, he would be able to win the dispute with a high probability. This indicates that the platform will earn the commission fee with a high probability; hence, its revenue is more dependent on the commission fee. Because the decentralized dispute system can induce the client to offer a higher contract price, the platform can earn a higher commission under the decentralized dispute system, and hence, the decentralized dispute system would make the platform better off. In contrast, when the freelancer’s skill level is relatively low, he can only win the dispute with a low probability. This indicates that the platform has to rely more on the dispute fee. As we have seen in Theorem 2, the platform’s bias creates a greater advantage for lower-skilled freelancers, which enables the platform to charge a higher dispute fee under the centralized dispute system. Thus, the centralized dispute system would make the platform better off in this case.

Therefore, the decentralized dispute system can only benefit the platform when the freelancers’ skill levels are sufficiently high. With higher-skilled freelancers, the platform would be able to utilize the tribunal to improve the incentive structure of participants and extract more commissions. However, with lower-skilled freelancers, the platform would benefit from retaining the decision-making power to arbitrate disputes to itself. By doing so, the platform can set a lower quality standard for lower-skilled freelancers and extract more dispute fees from them; however, under the decentralized dispute system, the more stringent quality standard set by the tribunal would make it disproportionately more costly for the lower-skilled freelancers to participate, and more of them can be weeded out.

Our results suggest that different types of dispute resolution systems can cater to different market segments. The decentralized dispute system is more suitable when the freelancer pool is higher skilled, whereas the centralized dispute system is more suitable when the freelancer pool is lower skilled. Therefore, for the emerging platforms to succeed with the decentralized dispute system, it is important to ensure the skill level of their freelancer pool. This can be achieved by providing better training to the freelancers, such as by partnering with online learning platforms (e.g., Coursera or Udemy) or adopting a stricter screening and certification process. For example, Upwork encourages its freelancers to obtain verifiable certifications, such as Adobe or Oracle, and Freelancer.com has internal programming and language tests for the freelancers to take.

Moreover, the freelancers’ skill levels may improve over time for the traditional platforms. The majority of the tasks in companies have been shifting to a project-based structure, under which companies can utilize an external workforce that is able to work remotely (Claussen et al. 2018). Thus, the freelancer market is expected to grow, and more professional employees will utilize the online labor platforms (Katz and Krueger 2019). Such an influx of professional employees may increase the overall skill level of the freelancer pool. As a result, it may become optimal for these platforms to switch to the decentralized dispute system in order to reap greater benefits.

6.3. Social Impact of the Decentralized Dispute System

Lastly, we investigate the impact of the decentralized dispute system on the social welfare. We have previously seen from Section 5 that the decentralized dispute system can be designed to deliver its promise of attaining a fairer dispute resolution than the centralized dispute system. As a result of having a fairer dispute resolution, the decentralized dispute system is able to induce a higher quality level in most cases (Theorem 2). However, a higher quality level does not always improve the social welfare. If the quality level already exceeds the socially optimal level, an even higher quality standard would cause the quality improvement incentive to deviate from that of a social planner and reduce the social welfare.

The social welfare in our model is $q-\alpha q^2$, and the socially optimal quality level is $\frac{1}{2\alpha }$. We are particularly interested in whether a higher quality level automatically translates into a more socially optimal outcome. To this end, Theorem 4 characterizes a condition such that the decentralized dispute system is able to attain a higher social welfare given that it induces a higher equilibrium quality level than the centralized dispute system.

As shown in Theorem 4, if the public is overly demanding on quality (i.e., $y>\overline{y}(\alpha )$), the freelancer’s effort may be overly exerted, and the decentralized dispute system will result in a quality level that deviates further away from the socially optimal level than that under the centralized dispute system $($i.e., $|q^{+}-\frac{1}{2\alpha }|>|q^{*}-\frac{1}{2\alpha }|)$. Thus, the industry standard y cannot be too high for the equilibrium outcome to be more socially optimal under the decentralized dispute system. Moreover, the threshold of how stringent the industry standard needs to be, $\overline{y}(\alpha )$, is dependent on the freelancer’s skill level, α. Theorem 4 further shows that under the same equilibrium regime, a lower-skilled freelancer would need a less stringent industry standard in order for the decentralized dispute system to result in a more socially optimal outcome. Therefore, to make sure that gig platforms’ adoption of the decentralized dispute system does not hurt the social welfare, a social planner should be mindful of the industry norm and take necessary measures to prevent the public from forming a standard that is overly stringent.

In addition, other studies have shown that most people tend to lose empathy and deviate from the norm as their influence and authority grow (Schaarschmidt 2017). Thus, it is important to manage the norm of the people with authority to ensure that the standard does not go exceedingly stringent over time. Such management of the norm is especially crucial under the decentralized dispute system, as the authority of the dispute decision has been surrendered by the platform to the voters. This may result in the standard becoming overly stringent if the voters’ authority goes unchecked. Thus, a totally hands-off approach toward the decentralized dispute system may not be advisable. One possible solution is for the platform to set guidelines or principles for the voters to follow when they are evaluating the dispute case. Such recommended guidelines can serve as a focal point for the voters to form their decisions, so that the voting mechanism of the tribunal can induce a more socially optimal quality level. This helps to ensure that the decentralized dispute system not only can be more profitable for the platform but also, can improve the social welfare at the same time.

7. Extensions

In addition to the main model presented in previous sections, we also generalize the model in several directions to test the robustness of our main insights regarding when the decentralized dispute system should be preferred by the platform and obtain additional insights regarding how the platform’s preference can be affected by other factors. Recall that we have identified two conditions from our main model regarding how a platform can successfully implement the decentralized dispute system: a condition to achieve fair voting (which requires the voters’ degree of subjectivity to be sufficiently high) and a condition for the decentralized dispute system to generate more profit than the centralized dispute system (which requires the freelancer skill level to be sufficiently high). We focus on checking the robustness of the condition for the decentralized dispute system to be more profitable (i.e., Theorem 3) because the condition to achieve fair voting is unaffected in the extensions. In this section, we present a summary of the five model extensions we analyze. The detailed model formulation, results, and proofs are provided in E-Companions EC.3–EC.7.

7.1. Third-Party Arbitration

In E-Companion EC.3, we extend the model of the centralized dispute system to allow the freelancer and the client to appeal the platform’s decision, in which case a third-party arbitrator will be involved to re-evaluate the dispute case and make a binding decision. This adds two more decision-making stages to the sequence of events, namely an “appealing” stage and a “third-party arbitration” stage. The third-party arbitrator chooses its own threshold to evaluate the freelancer’s work, taking into consideration both the industry standard and the decision of the platform. We find that regardless of the involvement of a third-party arbitrator, the decentralized dispute system dominates the centralized dispute system when the freelancer’s skill level is sufficiently high. Moreover, a third-party arbitrator whose incentive is more aligned with the platform can hurt the platform under the centralized dispute system and make the decentralized dispute system the preferred system in more cases. Thus, if the platform cannot convince its users that the third-party arbitrator is impartial, fully decentralizing dispute resolution to individual platform users can remove the need for the platform to commit to a fairer dispute resolution because the decision-making process is entirely entrusted to independent parties with no conflicts of interest.

7.2. Price-Dependent Industry Standard

In E-Companion EC.4, we allow the industry standard to be contingent on the contracting terms; hence, the contracting parties are able to influence the industry standard. Specifically, we assume that the industry standard is a linear increasing function of the contract price. We confirm that our main insight about the value of decentralizing dispute resolution is not driven by the platform users’ lack of power in influencing the industry standard, as our previous finding that the decentralized dispute system dominates the centralized dispute system when the freelancer’s skill level is sufficiently high continues to hold. Moreover, the decentralized dispute system is more likely to be the preferred system when the industry standard is less sensitive to the contract price.

7.3. Double-Sided Dispute Fees

In E-Companion EC.5, we model the scenario where the platform requires the dispute fee to be paid by both parties in order for the dispute case to be handled. In this case, after the freelancer initiates the dispute, the client decides whether to participate in the dispute by paying the dispute fee. If the client also pays the dispute fee, the dispute case will be evaluated. However, if the client decides not to pay the dispute fee, the freelancer will be automatically awarded a “win,” and the client will then have to pay the freelancer. We find that the decentralized dispute system still dominates the centralized dispute system when the freelancer’s skill level is sufficiently high. Thus, our main insight remains robust when the platform is allowed to choose which dispute fee structure to use under each dispute system. Moreover, we find that when the freelancer’s skill level is only moderately high, the platform’s optimal strategy is to combine the decentralized dispute system with the double-sided dispute fee structure. This indicates that platforms that intend to adopt the decentralized dispute system can use the double-sided dispute fee structure as a transitional step when the freelancers’ skill levels are not sufficiently high and switch to the single-sided dispute fee structure when the freelancers become more proficient.

7.4. Differential Subjectivity Between Platform and Voters

In E-Companion EC.6, we consider a model extension that allows the centralized dispute system to be associated with a lower degree of subjectivity in dispute judgment compared with the decentralized dispute system. This captures the scenario where the platform’s dispute team members may be less subjective in quality evaluation because they are professionally trained. We find that as long as the degree of subjectivity is not too different between the two systems, our previous results continue to hold exactly, and the platform should prefer the decentralized dispute system when the freelancer’s skill level is sufficiently low. Moreover, the decentralized dispute system can be the preferred system even for lower-skilled freelancers when it substantially increases the degree of subjectivity in dispute judgment from that under the centralized dispute system.

7.5. Heterogeneous Freelancers

In E-Companion EC.7, we extend our model to allow for a heterogeneous freelancer pool consisting of two types of freelancers with different skill levels. Although the client can choose the contract price based on the type of freelancer she is contracting with, the platform needs to set the same dispute fee up front. This extension serves as a robustness check for our main insight when the platform is no longer able to use a single dispute fee to extract surplus from all platform users. Through numerical analysis, we observe that the decentralized dispute system dominates the centralized dispute system if the proportion of high-type freelancers is sufficiently high, which indicates a higher average skill level of the freelancer pool and is consistent with our main insight. In addition, we observe that the degree of heterogeneity within the freelancer pool also impacts the platform’s optimal dispute system. Specifically, the centralized dispute system tends to perform better when the freelancers are more diverse in their skill levels.

8. Conclusion and Discussion

Disputes are an inevitable part of projects that involve multiple parties, and any intermediary platform must take into account dispute management when designing its policies. In this paper, we analyze and compare two types of dispute resolution systems for online labor platforms: the centralized dispute system, where the platform serves as the arbitrator, and the decentralized dispute system, where a tribunal consisting of individual platform users votes on the dispute case. Our findings shed light on the value and implementation of the decentralized dispute system. We demonstrate that the critical value of decentralizing dispute resolution lies in the collective input of individual users, which can result in a fairer resolution and eliminate any decision-making bias of the platform. However, realizing this value requires the platform to ensure the heterogeneity of voter assessments and the skill level of the freelancer pool.

Although our evaluation of the decentralized dispute system focuses on its fairness perspective, there are other factors that platforms should consider when adopting this system, which are beyond the scope of our model. For example, platforms need to consider the implementation costs of the tribunal system, including the costs of setting up the architecture and incentivizing users to participate as voters. However, such investments may pay off in the long run as the decentralized dispute system can reduce the expenditure of resources that platforms would otherwise incur in compensating in-house adjudicators. Moreover, because the decentralized dispute system crowdsources dispute resolution from the public, it has the potential to handle an increasing number of disputes as the number of platform users grows, ensuring its scalability. In addition, the rapid expansion of gig platforms and their substantial autonomy to dictate users’ contractual terms have recently prompted regulations aimed at limiting the platforms’ sweeping authority (Herman 2017). By requiring platforms to relinquish more decision-making power to their users, the decentralized dispute system can help reduce negative connotations and enable platforms to retain their status as part of the “sharing economy,” where certain commercial and labor laws may not apply.

It is worth noting that many emerging platforms are decentralized autonomous organizations deployed on the blockchain. Although blockchain is often viewed as a convenience way to implement the decentralized dispute system because of its ability to automate payment transactions and ensure credibility of payments, it is not the only option available. Platforms can use any trust-based payment intermediary to implement the incentive scheme of the tribunal system. For example, Ortolani (2015) proposes a preauthorization model using credit cards. Under this approach, a voter will be subject to a credit card preauthorization to participate in the tribunal. Once the voting is complete, the voters will be credited with the appropriate amount through the intermediary: in this case, the credit card system. This adjudication process resembles that used by the emerging platforms but does not require the use of blockchain. Moreover, some emerging platforms, such as Kleros and Jur, are third-party service providers that specialize in providing decentralized dispute resolution services. As a result, a centralized platform can outsource dispute resolution to such platforms and integrate their dispute system with its own to create a decentralized dispute system. Thus, a centralized platform does not need to fully decentralize its operations to adopt the decentralized dispute system.

Finally, we hope our work could trigger more future research to study the operational aspect of dispute management for gig platforms. For example, we have focused on the case where the skill level of the freelancer is known. One future research direction could be incorporating asymmetric information with regard to the freelancer’s skill level. Moreover, we have focused on the type of dispute that is initiated by the freelancer when the client refuses to pay. There are other types of disputes that gig platforms may need to handle, such as user conduct on the platform or intellectual property right of online gig work. It would be interesting to examine whether a decentralized arbitration mechanism can work effectively for other types of disputes. In addition, whereas this paper considers a setting where the client sets the contract price, future research could investigate dispute management when the contract price is specified by the freelancer. Finally, it would be interesting to study the long-term reputational effects of decentralizing dispute resolution for gig platforms and empirically test the theoretical predictions of these effects.