Research

Generative Artificial Intelligence (AI) technologies have emerged as a transformative force in the content creator economy. This paper examines how the AI stances signaled by specific platform policy decisions shape creator behavior on visual arts platforms. We leverage two independent natural experiments on leading Chinese platforms: Lofter’s launch of an AI image generator, which signaled a pro-AI stance, and Graffiti Kingdom’s prohibition of AI-generated artwork, which signaled an anti-AI stance. Our analysis shows that creators decreased their activity on Lofter following the AI generator launch, while activity increased on Graffiti Kingdom after the AI prohibition.

Multiple lines of evidence show these effects stem from policy-communicated stance signaling: creators respond to what the focal AI policy communicates about AI’s role, the platform’s commitment to human creators, and the future competitive environment for human-created work, rather than merely to specific tool features or enforcement actions. Heterogeneity analysis reveals that higher-popularity, multi-homing, and AI-averse creators show larger activity reductions on Lofter. Through analysis of creator posts, we identify three primary concerns driving resistance: replacement risk, perceived low quality, and copyright infringement.

To our knowledge, this is the first paper to causally identify how the AI stance signaled by a platform policy decision affects creator behavior. Our findings have implications for how platforms communicate AI-related policy decisions, and for policymakers on fostering a constructive relationship between AI and human creators.

Generative AI Search allows content creation platforms to answer users’ questions directly, potentially substituting for the user-generated content (UGC) they help users discover. How can platforms improve search while sustaining demand for creator content? We combine a randomized field experiment involving 100,000 users on a leading Chinese question-and-answer platform with a dynamic structural model of search. Continued access to an established AI search feature increases search frequency and the range of topics explored. Users inspect fewer UGC search results and open fewer UGC pages per session, yet both totals rise as users search more often. The model explains this scale–intensity trade-off: information acquired from AI reduces the benefit of reading additional UGC within a session, while higher expected search value encourages subsequent queries and daily search activity.

Counterfactual simulations show that this trade-off can be mitigated through design. Coordinating AI placement and informativeness, or reallocating AI availability across query intents within the existing budget, can increase both search activity and UGC consumption per query. These improvements need not coincide with greater AI engagement. Optimizing for AI openings favors search expansion at the expense of UGC consumption per query, whereas helping users assess an AI answer’s information content before opening it increases search and total UGC consumption with fewer AI openings. These findings highlight the importance of designing AI search around both the intensity of creator-content consumption and the scale of demand generated by more valuable search.

Many two-sided platforms do not operate a single undifferentiated marketplace. Instead, they organize the same intermediation function through communities that differ in topics, norms, identities, or other non-quality attributes. We study such horizontal differentiation as a platform design choice when users allocate scarce participation resources across communities and the outside option. At first glance, this setting resembles a product-line problem: added or strengthened communities may attract new participation, but they may also cannibalize incumbent communities. We show that this logic is incomplete on a two-sided platform because reallocation across communities can propagate through cross-network feedback. We define the cross-network reallocation effect as the additional equilibrium response generated when an initial movement of users or effort changes participation on the opposite side of affected communities and then feeds back into the side where the movement began. The implication is that horizontal differentiation creates a portfolio-level propagation problem rather than a one-sided trade-off between expansion and cannibalization. This mechanism generates three sign-changing boundaries absent from the conventional one-sided benchmark: whether adding differentiated communities expands aggregate activity or creates net cannibalization, whether unaffected communities merely attenuate a localized shock or reverse its aggregate effect, and whether a locally beneficial policy improves platform-wide outcomes or instead reduces them.

We quantify these mechanisms using Stack Exchange around ChatGPT’s release, focusing on two of its communities: Stack Overflow and Super User. ChatGPT’s release provides a localized disruption to Stack Overflow’s asker demand and audience-feedback environment, while Super User serves as a related, comparatively stable internal alternative through which within-platform reallocation can occur. In this setting, we estimate a dynamic structural model in which askers choose question volume and difficulty composition, while answerers allocate limited weekly capacity across communities and learn about changes in Stack Overflow’s audience-feedback environment from realized feedback.

Counterfactual simulations translate the mechanism into three portfolio-level implications. In stable environments, horizontal differentiation creates more value when added communities are attractive destinations for marginal answerer capacity and when their question pools generate strong answerer-side feedback; when these forces are weak, the differentiated platform can underperform the single-community benchmark. After localized shocks, the same forces make unaffected communities more valuable by helping them retain displaced answerer participation and amplify it enough to move the platform from shock attenuation to aggregate reversal. Community-targeted policies create a related portfolio trade-off: even when a policy improves the treated community, its platform-wide effect can be small or negative when the untreated community already supports expansion or post-shock resilience. In such cases, the policy carries a high opportunity cost because it pulls activity away from a community whose feedback loop is already generating substantial platform-wide value. The results imply that managers should evaluate communities by the reallocation paths they create, including where activity comes from, where it goes, and how gains and losses propagate across sides, rather than by isolated community-level performance.

Generative AI Search places an AI-generated answer above conventional search results. This design may improve search effectiveness while changing the attention users devote to the underlying sources. These effects are particularly important for content creation platforms, where user-generated content (UGC) attracts users and user attention sustains creator participation. We examine search activity and UGC consumption in a randomized field experiment involving 100,000 users on a leading Chinese content creation platform. Treatment-group users retained access to an established Generative AI Search feature, while the platform removed access for control-group users.

Access increased how often and how broadly users searched. Most additional same-day queries were classified as follow-up questions or distinct search tasks rather than revisions of earlier queries. The results reveal a scale–intensity trade-off in UGC consumption: consumption was lower per search session but higher in total over the experiment. Access also increased the number of search sessions in which users explored additional creator content after opening an initial UGC page. These findings show why engagement within individual searches can give an incomplete picture of how Generative AI Search affects overall demand for creator content.

Opaque products, which conceal key attributes until after purchase, have been widely adopted across various industries. Yet, empirical evidence on how such opacity influences consumer demand remains scarce. This study introduces a reference price framework to estimate demand for opaque products, focusing on blind boxes in the online retailing context. Using a unique dataset from a leading e-commerce platform, we estimate heterogeneous consumer preferences through an aggregate random coefficient logit model that incorporates gains and losses relative to reference prices.

We uncover a novel behavioral pattern of reversed loss aversion: consumers are more sensitive to gains than to losses when purchasing opaque products. Our findings provide the first empirical validation of this pattern in consumer decision-making using field data and demonstrate that this behavioral tendency makes the demand for opaque products fundamentally different from that of traditional products. The policy simulations show that ignoring reversed loss aversion would lead to misestimated own-price elasticity, resulting in suboptimal pricing strategies and substantial profit losses.

Price gouging regulation (PGR) helps ensure that essential goods, e.g., medical supplies or grocery staples, remain affordable and accessible. During the COVID-19 pandemic, PGR was implemented due to widespread reports of price-gouging on such critical items. Extant research offers few insights for policymakers and marketers into how PGR influences consumer spending and sales during crises.

Using Generalized Synthetic Control, we examine the overall and relative effectiveness of different types of PGR legislation, as well as the moderating effects of consumer mobility and the enactment of strict price caps. We find that PGR is generally highly effective, especially when it is not overly comprehensive and for products that are “essential” (e.g., paper towels) but not obviously related to preventing contagion (e.g., hand sanitizer). A nonparametric mediation analysis suggests that pricing alone does not fully explain PGR’s overall effectiveness, accounting for roughly 27–55%, depending on product class; moreover, it is strongly moderated by both specific price caps and consumer mobility.

We further add a mechanism-focused analysis using Numerator product-category data. To determine which categories are legally essential and theoretically informative, we use eight LLMs to code 1,904 product categories against pandemic price-gouging legal materials, aggregate the ratings with a principal-factor essentiality score, and then isolate three pre-specified classes of mechanism-eligible household essentials. In these categories, hierarchical bootstrap models show modestly lower prices, sharply higher online sales, lower offline sales, and a large shift from offline to online purchasing relative to inessential categories. Our results have direct implications for policymakers deciding whether to enact PGR as well as setting specific design features depending on locality.

“Host Residency Enforcement” policies have been widely adopted by cities in the US, mandating hosts of short-term rentals to reside in those properties to varying extents. For example, New York City’s “Physical Presence” policy requires hosts to remain present with guests during short-term rentals, while Los Angeles’s “Primary Residence” policy mandates hosts’ residence for at least six months annually. Similar regulations have been widely adopted across other cities. Although these policies aim to address housing issues, they may unintentionally amplify racial disparities due to existing racial discrimination on Airbnb. This paper examines whether and how “Host Residency Enforcement” policies exacerbate these racial disparities on the platform.

Our findings indicate that NYC’s “Physical Presence” policy increased the daily revenue gap between minority hosts (i.e., Hispanic and Black hosts) and White hosts by approximately $12. Similarly, LA’s “Primary Residence” policy widened the revenue disparity specifically between Black and White hosts by about $8, while having no significant effect on the gap between Hispanic and White hosts. Additionally, on the supply side, both policies increased the gap in the likelihood of listings remaining active between minority and White hosts; however, in LA, this effect was statistically significant only among Hispanic hosts. Moderating analysis further indicates that neighborhoods with higher income, higher proportions of White residents, and more native-born residents (i.e., neighborhoods with higher socioeconomic status) mitigate these exacerbated racial disparities. Additionally, listings signaling higher quality through their overall Airbnb ratings experienced reduced disparities. This finding suggests that statistical discrimination, primarily driven by information asymmetry, serves as a key underlying mechanism. Lastly, minority female hosts were disproportionately negatively affected compared to minority male hosts, highlighting intersectional discrimination. Although our analysis specifically focuses on NYC and LA, our findings can be generalized to other cities that have implemented the same or similar policies, as NYC and LA represent opposite ends of the regulatory stringency spectrum.

Our study is the first to demonstrate that the widely adopted “Host Residency Enforcement” policies can unintentionally amplify racial disparities on Airbnb, illustrating an important but understudied interaction between public policy and existing discrimination on online platforms. Drawing from our findings, we offer several practical recommendations for policymakers and online platforms.