The Impact of Generative Artificial Intelligence on Market Equilibrium: Evidence from a Natural Experiment

Generative artificial intelligence (AI) appears to propel our world closer to the realm of Cyberpunk—a futuristic milieu characterized by the coexistence of "high tech and low life": Generative AI demonstrates its ability to generate text, images, and videos that are similar to content created by human experts – but in much less time, at a fraction of the cost, and with amazing creativity [43]. Such advances, however, have raised significant concerns about the potential for job displacement. One notable manifestation is the organized opposition of artists, who have taken to social media to voice their opposition through campaigns such as "NO to AI-generated images" [20].

However, the potential impact of generative AI remains controversial in academia. Distinguished AI scientists such as Turing Award winners Yoshua Bengio, Yann LeCun, and Geoffrey Hinton - recognized with what is often called the "Nobel Prize of computer science" - hold divergent views. While Bengio and Hinton have publicly expressed concern, warning of "the risk of extinction by AI," LeCun dismisses such fears as unfounded, citing his understanding of the limitation of current AI systems. This debate echoes the broader discourse found in the economics literature, which suggests that while information technology may carry the risk of rendering certain jobs obsolete [1, 6, 17, 5, 45], it may also benefit workers by increasing their productivity [7, 29, 28].

To this end, this paper investigates the impact of generative AI on market equilibrium, with empirical data from China’s leading art outsourcing platform. Specifically, we seek to answer the following research questions: RQ1: In what ways does generative AI shape market equilibrium in terms of average price, order volume, and overall revenue? RQ2: What types of demand are eliminated or created by generative AI? RQ3: Who emerges as the primary suppliers in the market—are they incumbents or entrants utilizing generative AI?

The main challenge in assessing the impact of generative AI lies in achieving causal inference. This involves establishing a true cause-and-effect relationship between generative AI and the focal outcomes, rather than mere correlation. One straightforward approach is to contrast the outcome variable (e.g., the number of orders) after and before the emergence of generative AI. However, this method lacks the capacity for causal inference. For example, the demand for anime images could naturally fluctuate due to the seasonal effect of a new semester for students. This intrinsic variability could potentially bias our estimates of the impact of generative AI, leading to incorrect estimates. Another approach is randomized experiments (or A/B testing), which offer the most compelling evidence for causality. However, using experiments to study the impact of generative AI on markets is not feasible. Unlike individuals, markets cannot be randomly assigned to different groups, which limits the ability to implement generative AI effects only on the treated group while leaving the control group unaffected. Macro-level analyses involving countries, cities, and industries face similar limitations.

To overcome the challenge of causal inference, we identify an unanticipated and sudden leak of a highly proficient image-generative AI as a natural experiment, where markets inadvertently encounter experimental and control conditions, resembling a scenario wherein nature orchestrates an experiment.

Among the pioneer platforms in the field, NovelAI stands out as an early entrant, offering text-to-image generative AI services. These services are available for a subscription fee ranging from $10 to $25 per month per user. Powered by Stable Diffusion [34], NovelAI’s AI can quickly generate anime-style images in response to specific keyword-based requests. However, in early October 2022, hackers attacked NovelAI and leaked the AI model and training codes used by NovelAI. As a result, the leaked resources are available for download and local implementation, effectively reducing the cost of image generation to nearly zero.

The inadvertent leak, which originated in the Western world, has inadvertently reverberated throughout the Eastern world. Prior to October 2022, China lacked domestic providers of generative AI, while access to foreign alternatives was also impeded. After October 2022, however, the leaked AI, along with its accompanying tutorials, quickly spread throughout China’s social media landscape. As a result, this leak has expeditiously transitioned China from a state of being unaffected by generative AI to one of being affected by it.

This leak served as an ideal natural experiment. First, China constitutes an environment where painting markets remained unaffected by generative AI prior to the leak due to policy restrictions, language restrictions, and payment restrictions. Second, the leaked AI specializes exclusively in generating anime-style images, thus affecting only the anime painting markets while leaving other painting styles unaffected, allowing for a comparison between treated and control units. Third, the unanticipated and sudden nature of the leak is exogenous. This feature ensures that the focal markets are unlikely to have been affected by unobserved confounding factors as stakeholders did not have time to react to the leak in advance. Thus, alternative explanations are effectively ruled out.

We collected 197,068 records from a leading art outsourcing platform in China from January 2022 to February 2023. This platform allows consumers to place orders with specifications and budgets, and artists as freelancers to fulfill them. The orders are segmented into distinct markets, such as anime, UI, live2d, and wallpaper. We transformed these order records into a market-week panel dataset, focusing on average price, order volume, and overall revenue. Using a difference-in-differences research design, we compare the anime market to the wallpaper market, before and after the unintended leak of NovelAI’s anime generative AI. Our main findings are summarized in Figure 1.

Our analysis reveals a promising effect on market equilibrium following the emergence of generative AI. Specifically, although this emergence leads to a 64% reduction in the average price, the generative AI also leads to a remarkable 121% increase in order volume and a consequent 56% increase in overall revenue. All effects are statistically significant at the 1% level. In addition, the effects are also economically significant, due to their substantial magnitude. The results suggest that generative AI can contribute to a more prosperous market, where each stakeholder benefits: consumers get cheaper services, and human creators and the platform get more orders and revenue.

We then examine what types of demand are created by generative AI. Orders on the platform are categorized into personal orders and commercial orders, with different licenses. We conduct a heterogeneous analysis, showing that the growth of generative AI is mainly driven by personal orders, rather than commercial orders. The "tachie" market achieved significantly more revenue through more personal orders at low prices, while the price, volume, and revenue of commercial orders are relatively stable. Our results indicate the potential of generative AI for business success by satisfying the "low-end" needs of individual users.

Finally, we examine who emerges as the primary suppliers in the market. Our results show that the incumbents, who registered before the leak, won 97% of the orders after the leak. As a result, they reap most of the benefits of increased order volume and revenue. The results rule out the possibility that new entrants using generative AI have taken over the market share of real human creators.

Our work provides valuable managerial implications. First, the increased order volume and overall revenue indicate more jobs and taxes, providing favorable economic evidence of generative AI for policymakers to make supportive AI policies. Second, generative AI mainly benefits incumbent creators, suggesting that they should welcome and incorporate generative AI into their workflow to further increase their productivity. Third, online platforms can conduct marketing campaigns and improve services specifically for individual users to better exploit the blue ocean driven by generative AI.

In conclusion, this paper makes the following contributions:

• •

  To the best of our knowledge, we pioneer the estimation of the impact of generative AI on market equilibrium, making scholarly contributions to the growing literature on the economics of (generative) AI.

• •

  We provide promising empirical evidence that generative AI can promote market prosperity and explain the reason. We also provide valuable managerial insights for various stakeholders.

• •

  Our identified natural experiment facilitates causal inference for future research on generative AI—its effects on customer satisfaction, on employment, and beyond.

This paper adopts empirical research methods to study generative AI, which is related to the field of generative AI and the economics of AI.

In this section, we compare different empirical methods for estimating the causal effects of generative AI. We apply our analysis under the potential outcome framework [35], the most widely used framework for causal inference.

In this section, we answer the research question RQ1, by examining the impact of generative AI on market equilibrium in terms of average price, order volume, and overall revenue.

We employ the standard multi-periods model specification of difference-in-differences with two-way fixed effects [8]:

In this model, the variable $y_{it}$ represents one of our outcome variables by category $i$ in week $t$. $treated_{i}$ is a dummy variable indicating whether category $i$ belongs to the treated group ($treated_{i}=1$) or the control group ($treated_{i}=0$). $after_{t}$ is also a dummy variable indicating whether week $t$ is after the natural experiment ($after_{t}=1$) or before the natural experiment ($after_{t}=0$). The model includes a fixed effect $\delta_{t}$ for each week to control for observed and unobserved nonparametric time trends. Additionally, the model includes a fixed effect $\alpha_{i}$ for each category to account for time-invariant differences between categories. The parameter of interest is $\beta$, which captures the impact of generative AI on the focal outcome variable.

Table 2 shows the estimates for the change in price, volume, and revenue in columns (1) to (3), respectively. The coefficient estimate of price is -646, indicating generative AI lowered the average price by 646 yuan. Notably, the coefficient estimates of volume and revenue are 140 and 56,674, showing that generative AI increased the number of orders by 140 per week and increased the total revenue by 56,674 yuan per week. In addition, columns (4) to (6) take the logarithm of price, volume, and revenue and show their percentage change. The results show that price decreased by 64%, volume increased by 121%, and revenue increased by 56%. All effects are statistically significant at the 1% level.

Public concern about generative AI is often rooted in negative anticipation of the economic ramifications associated with its adoption. In particular, there is an understandable concern among various segments, including artists, that generative AI could lead to feelings of displacement, given the fear of rapid and unceremonious replacement after years of dedicated study and creative toil.

However, our empirical findings provide a counterintuitive insight, showing that contrary to initial impressions, generative AI can benefit all stakeholders across the platform economy. Consumers can fulfill their orders 64% cheaper than before, indicating an improvement in consumer welfare. Furthermore, artists receive 56% more revenue, as each monetary transaction completed by consumers is considered revenue for the artist. They receive 121% more orders, which underscores the increased opportunities that accelerate the evolution of their artistic endeavors by providing an invaluable platform for rapid prototyping and iterative refinement. Moreover, the platform’s revenue also increases, as their commission is proportional to the total revenue. The substantial increase in order volume could further benefit the platform economy through the network effect [40, 24].

In this section, we answer research question RQ2, which aims to identify the specific demand growth that drives market prosperity²²2”Once you start thinking about growth, it’s hard to think about anything else,” said Robert Lucas, a Nobel Prize-winning economist.. To this end, the heterogeneous effects on order licenses are examined.

The orders on the focal platform are classified into two categories based on licenses: personal orders and commercial orders. The artwork of personal orders is for personal use only, whereas the artwork of commercial orders is permitted for commercial purposes. In particular, commercial orders are granted more extensive property rights than personal orders, such as distribution rights and rental rights.

Figure 3 shows the market equilibrium outcomes of personal orders and commercial orders in the ”tachie” market. Prior to and following the advent of generative AI, the price, volume, and revenue of commercial orders remain relatively stable. In contrast, personal orders exhibit a substantial increase in volume and revenue with a corresponding decrease in price, which aligns with the overall "tachie" market. This suggests that the growth in the "tachie" market is primarily driven by personal orders.

Formally, we replace the term, $treated$, in Equation 1 with $personal\_treated$ and $commercial\_treated$ to capture heterogeneous treatment effects for different order licenses:

Table 3 presents the heterogeneous impact of Generative AI with different licenses. The results indicate that both personal and commercial orders exhibit a similar trend with regard to the primary average effects, namely a reduction in price and an increase in volume and revenue. Furthermore, the results demonstrate that the magnitude of the estimators for personal orders is notably higher than that of commercial orders, suggesting that generative AI benefits the "tachie" market by increasing the demand for personal orders with lower prices.

Our findings indicate that, as a blue ocean strategy [21], generative AI can achieve its success by fulfilling the "low-end" needs of individual users, which were previously unmet due to the high cost. In contrast, the requisite quality for commercial orders exceeds the capabilities of current generative AI. For example, there is a notable disparity between the demo images produced by NovelAI (Figure 4 in Appendix) and successful commercial anime images (such as [18]) that feature considerably more intricate and detailed designs. Conversely, individual users may not require a very high level of quality for their orders. Generative AI can fulfill their demand directly or with a slight modification by human creators. Furthermore, generative AI may present potential copyright issues that limit its commercial applications. This is because the training of generative AI often involves the use of a substantial number of images collected from the Internet without copyright permission.

Our previous findings indicate that generative AI has the potential to benefit artists by increasing their order volume and revenue. However, it remains unclear whether incumbents or new entrants utilizing generative AI will reap most of the benefits. In this section, we answer research question RQ3, by examining the incumbents in the market.

We focus on the completed orders in order to observe which artists won the orders. We define incumbents as the group of artists who registered prior to the leak, i.e., before October 2022. We will examine the market equilibrium before and after the leak with the orders completed by incumbents.

We employ the standard difference-in-differences in Equation (1). Table 4 presents the estimation results. Panel A’s dataset comprises all completed orders, whereas Panel B’s dataset comprises only orders completed by incumbents. The results in Panel B are consistent with our main effects, indicating that incumbents benefit from generative AI by gaining more orders (116%) and more revenue (65%). Furthermore, the estimators in Panel B is close to those in Panel A, indicating that the majority of the orders were completed by incumbents, resulting in a similar sample for Panel A and B. According to our data, the incumbents won 97.41% of the orders and 97.96% of the revenue after the leak.

One potential explanation for this result is that the current generative AI is unable to fulfill orders without human artists. This is because the generation of content is an iterative process between consumers and artists, with artists required to make ongoing modifications in response to consumer feedback. The current generative AI, which outputs images, audio, and videos, lacks the capacity to make slight modifications independently due to the limitations of its underlying algorithms.

Our results rule out the possibility that new entrants using generative AI have taken over the market share of genuine human creators. In contrast, generative AI serves to expand the market itself, with the incumbent creators deriving benefits from such expansion. The findings provide empirical evidence of a favorable economic consequence of generative AI for artists and suggest that they should embrace generative AI rather than resist it. The incorporation of generative AI into one’s workflow has the potential to increase productivity and leverage the novelty and enhanced capabilities provided by generative AI to attract new customers who previously were not engaging with the art market, thereby "growing the pie."

This paper examines the impact of the advent of generative AI on market equilibrium, in terms of price, volume, and revenue. We identify an unanticipated and sudden leak of an advanced image-generative AI as a natural experiment, which facilitates difference-in-differences comparisons for causal inference. The results indicate that generative AI substantially increases order volume and overall revenue with a corresponding decrease in average price. The increase in volume and revenue is primarily driven by the rising demand for personal orders with lower prices, rather than expensive commercial orders. Moreover, the incumbent creators maintained the majority of the market share, reaping most of the benefits from generative AI.

This paper makes a significant scholarly contribution to the growing literature on the economics of AI, by providing one of the first empirical evidence on how generative AI affects the platform economy through diverse perspectives. Moreover, the findings also offer valuable practical insights. Creators should embrace generative AI rather than resist it to gain greater benefits from enhanced productivity. Online platforms can conduct marketing campaigns and improve services specifically for "low-end" personal orders, the emerging demand driven by generative AI. Policymakers may also implement supportive policies for generative AI to create more jobs and tax revenue.

It is our contention that generative AI is having and will continue to have a profound impact on our world. In the context of the marketplace, future research can explore how generative AI affects product quality, consumer satisfaction, and beyond through our identified natural experiment.

Acknowledgments. We would like to thank Reviewer HvRd, QbGb, and Area Chair Y8tc for their constructive feedback. This work was supported by the National Natural Science Foundation of China (Grant No.92370204), National Key R&D Program of China (Grant No.2023YFF0725001), Guangzhou-HKUST(GZ) Joint Funding Program (Grant No.2023A03J0008), Education Bureau of Guangzhou Municipality.