Welfare Implications of Democratization in Content Creation: Generative AI and Beyond

Abstract

The rise of content-creation technologies such as generative artificial intelligence (GenAI) has reshaped the way people produce and consume content. These technologies democratize content creation, enabling inexperienced individuals to generate content that rivals the quality produced by professionals. This article adopts a game-theoretic model to study how the democratization of content creation affects welfare. While the content-democratization technology can assist low-quality creators in improving their content quality, it can precipitate a flood of such content in the market and crowd out high-quality content production, leading to welfare losses. An incremental technology development with a small reduction in the quality gap can hurt individual consumers and all creators, leading to a lose-lose market outcome. In contrast, a development with a high reduction in the quality gap benefits individual consumers and all creators, leading to a win-win outcome. Furthermore, when platforms can more effectively screen content and promote high-quality content, welfare rises, but the content-democratization technology is more likely to reduce welfare. The findings help explain the distinct impacts of content democratization in different industries.

Keywords

content democratization content creation technology quality competition consumer search generative AI content screening entry

Our present era is largely shaped by the rise of digital technologies. Among these, generative artificial intelligence (GenAI) is revolutionary. According to the World Economic Forum (Routley 2023), “Generative AI refers to a category of artificial intelligence (AI) algorithms that generate new outputs based on the data they have been trained on. Unlike traditional AI systems that are designed to recognize patterns and make predictions, generative AI creates new content in the form of images, text, audio, and more.” For instance, OpenAI's ChatGPT is backed by a large language model that is adept at generating humanlike text in response to prompts (an input message or a series of input messages given to the model to initiate a response), a feat once believed to be the sole territory of human intelligence (Radford et al. 2019).

GenAI tools have dramatically transformed the landscape of content markets by democratizing content creation, leveling the playing field in sectors previously gated by specialized knowledge or expertise. As indicated by Nvidia's CEO, Jensen Huang, GenAI tools are democratizing creators’ skills (Bajarin 2024). For example, image-creation tools such as DALL-E, Midjourney, and Stable Diffusion can craft intricate artwork, democratizing the creation of digital art (Roose 2022). Google's Magenta and Meta's AudioCraft can produce “complex melodies, harmonies, and soundscapes,” bridging the gap between seasoned composers and amateur enthusiasts. Abundant recent empirical research consistently corroborates that GenAI has democratized content creation in various industries. For example, Noy and Zhang (2023) find in a lab experiment that in writing tasks, performance inequality among workers decreased in the presence of ChatGPT. Zhou and Lee (2024) find that in a digital artwork marketplace, text-to-image GenAI tools (e.g., Midjourney) increased artwork quality more for artists who had been less capable of creating high-quality artwork. Many studies also document that for creativity-related jobs in various industries, GenAI has improved the performance of low-skilled, novice workers much more than high-skilled workers (Brynjolfsson, Li, and Raymond 2025; Chen and Chan 2024; Choi and Schwarcz 2025; Dell’Acqua et al. 2026; Hosanagar and Ahn 2024; Hui, Reshef, and Zhou 2024; Peng et al. 2023).

While GenAI has enabled novices to accomplish complex tasks without grasping the underlying technologies or knowledge, concerns have been raised regarding the impact of the content created by GenAI tools. In May 2023, NewsGuard reported 49 news websites that used GenAI to create hundreds of news articles per day, and the number of such “content farms” skyrocketed to 277 two months later (McMillan 2023). Some GenAI articles were widely circulated on social media such as Facebook, although they often had mediocre quality that could not match authentic news articles on prestigious sites (Sadeghi and Arvanitis 2023). In addition, Burtch, Lee, and Chen (2024) identify declines in the average answer quality provided at Stack Overflow (a popular question-and-answer website for programmers) because of users’ potential reliance on ChatGPT when giving answers.

GenAI is just one example of the various content-democratization technologies. For example, easy-to-use video editing tools such as CapCut significantly lower the skill barrier for individuals to create their own video content (Huang 2023), and social network platforms (e.g., Instagram) also provide tools for their users to edit and enhance the quality of their photos (Instagram 2025). However, the quality of such videos and photos still cannot match those crafted by state-of-the-art video editors and photographers.

Meanwhile, the potential downside of content-democratization technologies can be concerning for policymakers and platform managers. If the introduction or the enhancement of those technologies reduces welfare, policymakers need to be cautious in promoting the technologies, and platform managers should carefully facilitate the adoption of such technologies and be prepared for the potential damage. Our article aims to explore whether and when content-democratization technologies might positively or negatively impact consumer and social welfare in the content market and to unearth the important factors that moderate these effects. Our study can serve as the first step to understand the welfare impact of content-democratization technologies, and we seek to offer insights that can inform the decision-making of policymakers and platform managers regarding technology adoption.

We highlight that we do not intend to make empirical claims on the magnitude of any positive or negative effect brought by the content-democratization technology at present or in the future. Our goal is to reveal a few important underlying forces that govern how the technology affects welfare. In addition, although we use GenAI as a motivating example, it represents only one prominent instance of various content-democratization technologies. Our results apply to content-democratization technologies more broadly and focus on their feature of democratizing content production, rather than making specific predictions about GenAI.

Formally, we propose a game-theoretic model to understand how technologies that can democratize content creation affect the content market and consumer surplus. In the main model, we consider a content market with two types of content creators who differ in their content qualities and make independent decisions on whether to produce content. Consumers, who can consume a limited quantity of content, sequentially search for content to consume. A content-creation technology is available for creators, which can improve a creator's content quality. Moreover, the technology democratizes content creation, in that its quality improvement is more pronounced for low-quality creators than for high-quality ones. Our main analysis focuses on the technology's benefit for the low-quality creators. In an extended analysis, we allow the technology to improve the quality for all creators.

Our first main result identifies conditions under which the democratization of content creation increases or decreases welfare. The condition depends on what content is available in the market. As the technology enables low-skilled creators to create content with better quality, all else being equal, people are more likely to consume the improved (but still low-quality) content instead of continuing to search for high-quality content. This motivates low-quality creators to produce more content but discourages high-quality creators’ production, which reduces consumers’ chance of encountering high-quality content in the search process and in turn further decreases their willingness to search for it. In summary, although the content-democratization technology increases the quality of low-quality content, it crowds out consumers’ consumption of high-quality content. When the content-democratization technology leads to only incremental quality improvement (i.e., the quality gap between low-quality and high-quality content is still large), the substitution of high-quality content by low-quality content yields a significant decline in the average quality of consumed content, which can lead to a loss in consumer surplus. Furthermore, the crowd-out effect may propel many consumers to exit the market, reducing creators’ total payoff and social welfare. By contrast, when the content-democratization technology leads to drastic quality improvement (i.e., the quality gap between low-quality and high-quality content becomes much smaller), consumers and society are better off.

Next, we investigate several important factors that moderate the welfare impact of content democratization. As suggested previously, content democratization can cause an influx of low-quality content to content platforms (e.g., YouTube), which may reduce consumers’ satisfaction when using these platforms. To address this issue, many platforms implement various quality-screening measures, such as deploying AI-based algorithms and hiring human evaluators to better identify and promote high-quality content to consumers. Our model captures content quality screening by allowing consumers to encounter content with higher quality with a higher probability in their search process. Our second main result uncovers that given the availability of content-democratization technology, when the platform can more effectively screen high-quality content, welfare (consumer surplus, creators’ total payoff, and social welfare) indeed increases; however, the content-democratization technology becomes more likely to hurt welfare. The reason is that higher screening efficiency regarding content quality directly increases consumers’ chance of receiving high-quality content in their search process. Meanwhile, with higher screening efficiency, a quality enhancement of the low-quality content brought by the technology more significantly increases the content's probability of passing the screening process and thus yields a larger expansion in its probability of being sampled in consumers’ search process, leading to a more significant crowd-out effect.

Our third main result shows that when the content-democratization technology is more accessible to low-quality creators, the welfare loss due to the crowd-out effect is less likely to occur. This is because, with larger coverage of the technology, fewer high-quality creators are willing to produce content, resulting in less high-quality content that can potentially be crowded out in the market. In this case, the negative impact of the crowd-out effect is less salient compared with the benefits brought by content democratization.

Besides impacting content quality and production efficiency, the content-democratization technology can also affect content variety in the market. When many creators use the same technology to produce content, their content can be highly similar in style and substance (Zhou and Lee 2024). We investigate the effect of the content-democratization technology on content variety by incorporating horizontal varieties of content and consumers’ heterogeneous horizontal preferences for content. Our fourth main result reveals that while content democratization may substantially improve the quality of an individual piece of content, it is more likely to hurt consumers if it reduces the horizontal variation across different content. This is because, with lower content variation, consumers become less likely to encounter highly matched and highly unmatched content and more likely to encounter “mediocre” content. Note that consumers are more likely to choose to consume a piece of content they encounter (instead of scrolling past it) if it has a higher match value. Consequently, the lower content variation shifts consumers’ consumption to “mediocre” content from highly matched content more significantly than from highly unmatched content. Therefore, the reduced consumption probability of highly matched content outweighs the reduced consumption probability of highly unmatched content, resulting in a lower consumer surplus. Additionally, this welfare reduction is further exacerbated by the crowd-out effect of content democratization, which increases the proportion of content with a lower horizontal variation.

In summary, our article provides a game-theoretical framework to systematically understand and predict the welfare impacts of content democratization. The “Managerial and Policy Implications” section illustrates how our framework can explain seemingly conflicting attitudes toward content democratization in different industries observed in reality. In addition, we demonstrate the robustness of our results and derive additional insights in various alternative settings.

Literature Review

Our article is closely related to the literature on content creation. An important research topic of this literature is how content creators determine their content quality and how content platforms can facilitate the production of high-quality content. Yildirim, Gal-Or, and Geylani (2013) and Zhang and Sarvary (2015) explore strategies of competing user-generated content platforms with horizontally differentiated content. Sun and Zhu (2013) empirically show that an advertising revenue-sharing program on a blog platform incentivizes bloggers to produce blogs with better quality and more popular topics. These effects are stronger for participants with moderate popularity. Relatedly, Jain and Qian (2021) theoretically investigate a content platform's optimal ad-revenue-sharing payment schemes. They find that when the platform hosts more creators, it should share more ad revenues to encourage the production of high-quality content. Li, Liao, and Xie (2021) empirically discover that deterrence of article piracy on an online publishing platform increases the quantity of articles produced by writers who receive monetary payments for these articles without compromising their article quality. However, these writers spend less effort on customer care because they face weaker competitive threats of pirated articles. Berman and Katona (2020) and Qian and Jain (2024) suggest that to encourage creators to produce high-quality content, platforms may want to emphasize content quality more than horizontal fit in content recommendations to consumers. Meanwhile, Zou, Wu, and Sarvary (2025) show that emphasizing quality in recommendations may intensify the quality competition between creators and reduce their entry to content platforms, which can reduce content variety and hurt the platforms. In the emerging market of AI-generated content, Chen, Ke, and Shin (2025) show that platforms’ AI-labeling strategies should balance label credibility, creators’ incentives, and consumer engagement, rather than focusing solely on statistical measures such as false positives or false negatives. A common premise of the aforementioned research is that higher content quality always benefits consumers and society. In stark contrast, our research shows that when consumers need to search for content and endogenously decide what content to consume, quality improvement in low-quality content due to content democratization can crowd out consumers’ consumption of high-quality content, which reduces the production incentives for high-quality content creators and may eventually hurt consumers and society. This negative effect can persist even when the quality improvement is equally pronounced for high-quality and low-quality content.

Our article is also related to the literature on quality segmentation and quality competition. For example, Desai (2001) investigates the cannibalization problem that monopoly or duopoly firms face in the presence of two consumer segments who differ in their quality valuation, and the author characterizes when cannibalization affects the firms’ product-line design decisions. Jing (2006) identifies the conditions under which a higher-quality product is more profitable for vertically differentiated firms. Ishibashi and Matsushima (2009) find that in both Cournot and Bertrand models, the entry of low-end firms can benefit incumbent high-end firms because it incentivizes high-end firms to sell their products only to high-end consumers. Amaldoss and Shin (2011) reveal that a larger population of low-valuation consumers may alleviate price competition and improve firm profits. Our research complements this literature and differs from it in the following aspects. First, our model incorporates both high- and low-quality content creators’ endogenous entry decisions. Second, we allow consumers to search for content they would like to consume, resulting in multiple endogenous segments of consumers (depending on their search costs)—some consume high-quality content only, some consume content with any quality, and others do not search. Third, our model focuses on a content market, where creators profit from the consumption level (i.e., number of views or impressions) their content generates; in contrast, firms in the aforementioned literature obtain their payoff from consumer purchases, wherein the firms’ pricing decisions are critical.

Methodologically, our model of consumers’ content search is related to the vast sequential search literature (e.g., McCall 1965; Rothschild 1978; Weitzman 1979). We apply the sequential search framework to a unique setting, where content democratization raises the quality of low-quality content, and we examine how this change affects endogenous outcomes of both the supply side (creators’ entry decisions) and the demand side (consumers’ content search and consumption decisions) in the content market. Furthermore, our model captures a distinctive feature of the content market: The probability of a content item being sampled can depend on its quality, motivated by content quality screening measures widely adopted by content platforms in practice. This contrasts with the typical assumptions in existing literature, which consider the sampling probability of an item to be independent of quality.

Given that GenAI is a prominent example of content-democratization technologies, our article also contributes to the burgeoning theory literature on AI in marketing. This stream of literature has covered the application of AI in forecasting (Miklós-Thal and Tucker 2019), content moderation (Liu, Yildirim, and Zhang 2022), prediction (Choi, Liu, and Shin 2024), product recommendation systems (Cao et al. 2024; Zhou and Zou 2023; Zou and Zhou 2025), health care (Dai and Singh 2025), learning (Li, Li, and Zhang 2023), and decision interpretations (Mohammadi et al. 2024). In contrast to this stream of literature, our focus is content democratization technologies such as GenAI, which can improve content quality and production efficiency, and we investigate the role of such technologies in the content market.

In addition, our research is broadly linked to recent empirical evidence on the potential negative influences of GenAI usage. In the computer science literature, for example, Ji et al. (2023) offer a comprehensive overview of hallucinations in natural language generation. Shumailov et al. (2023) find that as GenAI models are trained by machine-generated data, the resulting models may suffer from irreversible defects. These studies focus on features of content-democratization technologies that are outright detrimental. By contrast, our article reveals how and when the positive feature of content democratization of raising content quality may turn out to be harmful for consumers and society. We provide a theoretical framework to analyze nuanced welfare impacts of content democratization.

Base Model

To establish the key insights, we begin with a base model, which we will extend in several aspects subsequently.

Creators

A unit mass of consumers (indexed by i) consume content produced by different content creators, whose population M is very large (but finite), which is consistent with most real content markets. There are two types of content creators, j ∈ {H, L} (i.e., high quality and low quality), who differ in the content quality q_j they produce: q_H > q_L. We refer to content produced by type-j creators as type-j content, whose quality is q_j. Let θ_j be the fraction of type-j creators, so their population is θ_jM. Each type-j creator, whose mass is negligible compared with θ_jM, produces at most one piece of content. Two subsequent extensions will consider scenarios where creators endogenously decide quality and where there are K ≥ 2 creator types.

Creators of each type produce symmetric content. The revenue of a content creator is assumed to coincide with the number of consumers consuming the content—the revenue per view is normalized to 1. In our setup, creators do not charge consumers for their content or make pricing decisions, and creators’ revenue per view is treated as fixed. This aligns with the practices of most content platforms, where creators primarily profit from platform-shared advertising revenue at predetermined rates (e.g., 55% on YouTube). Let V_j denote the viewership of a type-j content item, defined as the number of consumers who end up consuming the content. The expected payoff for a type-j creator is π_j = 1 × V_j − c, and a type-j creator will produce if and only if π_j ≥ 0. The production cost c follows a uniform distribution on $[0, {\bar{c}}_{j}]$ . We assume that ${\bar{c}}_{j} > 1$ to ensure the interior solution—the type-j creator with the highest cost $c = {\bar{c}}_{j}$ will not produce. We define $β_{j} = \frac{1}{{\bar{c}}_{j}} \in (0, 1)$ as the creation efficiency for type-j creators. Finally, let N_j denote the number of type-j creators who choose to produce content, which is also equal to the amount of type-j content available in the market. For tractability, we follow the literature (e.g., Salop 1979; Wu 2024) and treat the values of M, θ_jM, and N_j as real numbers.¹

Content Consumption

Consumers sequentially search for content to consume. Given the vast amount of content available in the market, consumers can only search and consume a very small subset of content due to their attention and time constraints (e.g., mental fatigue). We assume that consumers consume at most one content item, and their payoff equals the quality of the consumed content q_j. All our results qualitatively apply if consumers consume at most n₀ > 1 content items, as long as the constant n₀ ≪ M.² Additionally, in an extension, we examine the situation where consumers also have horizontal preferences for content and the content-democratization technology can influence the horizontal variety among content.

Consumers can discover a new content piece by making a search attempt. In reality, this reflects that consumers check another section on a video website or scroll to the next screen on a social media platform. A consumer i needs to incur s_i for each search attempt, where s_i ∼ Unif [0, S] across consumers, representing their heterogeneous search costs. We assume the upper bound of search cost, S, is sufficiently large so that the consumer with the highest search cost s_i = S does not make any search attempt. A consumer's total utility when consuming type-j content is u_i = q_j − x_is_i, where x_i ≥ 1 is the number of search attempts the consumer conducts. We allow consumers not to search any content (x_i = 0), in which case their utility is 0. We assume that consumers learn about the quality of a content item after searching it and can evaluate content quality without consuming the entire piece of content. For example, in the case of text content, consumers can evaluate the content quality by examining the first paragraph or article summary. In the case of video content, consumers often decide whether to continue watching a video within the first few seconds (Teleprompter 2023). In addition, in our model, content creators do not derive payoffs from such (brief) examination of the content by consumers.³

In each search attempt, consumers discover a piece of type-j content with probability ρ_j.⁴ The probability ρ_j captures a content platform's matching technology, which displays higher-quality content to consumers with a higher probability: ρ_H ≥ ρ_L. Put differently, higher-quality content is more likely to be recommended by the content platform. We assume that for any two content items whose types are j₁, j₂, $\frac{ρ_{j_{1}}}{ρ_{j_{2}}} = {(\frac{q_{j_{1}}}{q_{j_{2}}})}^{b}$ , where b ≥ 0. We refer to b as the quality screening effectiveness of the matching technology. A higher b indicates that a high-quality content item is more likely to appear in a consumer's search attempt compared to a low-quality one. At an extreme b = 0, as in most traditional sequential search models, consumers encounter all content items with equal probabilities, regardless of quality.

Because ρ_HN_H + ρ_LN_L = 1, the probability for consumers to find a type-j content item in a search attempt can be derived as $ρ_{j} = \frac{q_{j}^{b}}{N_{L} q_{L}^{b} + N_{H} q_{H}^{b}}$ . Content produced by different creators competes for consumers’ attention in their search process: A content item is less likely to be sampled in an attempt when more content is available on the market: ρ_j decreases with N_L and N_H. The notation of the base model is summarized in Table 1.

Table 1.
Summary of Notation.

Symbol Definition

i An index that labels consumers

j ∈ {H, L}: An index that labels creator type

M Population of all creators

θ_jM Population of type-j creators

q_j Content quality of a type-j creator, q_H > q_L

N_j Amount of type-j content available in the market

c $\in [0, {\bar{c}}_{j}]$ , a creator's cost of content creation

${\bar{c}}_{j}$ Upper bound of the content creation cost of type-j creators

β_j = $\frac{1}{{\bar{c}}_{j}}$ , creation efficiency of type-j creators

ρ_j A consumer's probability of encountering a piece of type-j content in a search attempt

b Effectiveness of quality screening

Φ_j = ρ_jN_j, a consumer's probability of encountering some type-j content in a search attempt

P The number of active consumers

s_i ∈ [0, S], a consumer's search cost

S The upper bound of search cost

V_j The viewership of a type-j content item

W Consumer surplus

SW Social welfare

Π_j Total profit for type-j creators

Π Total profit for all creators

□* Superscript that signifies equilibrium outcome

Content-Democratization Technology

Each creator is endowed with a basic capability to produce a unit piece of content. Now a content-democratization technology arrives, and for simplicity, we assume that it can be used by any creator at no cost. The content-democratization technology enables any type-j creator to improve her content quality q_j. For example, content creators can make use of tools such as ChatGPT to obtain higher-quality ideas and content. We allow the improvements in content quality to be different for the two types of creators. In the main analysis, we consider the case where the technology democratizes content creation by helping low-quality creators more, which is consistent with most empirical findings, as discussed previously. For concise illustration, in our base model, the technology helps the low-quality ones only (i.e., q_H does not change). This allows us to focus on the welfare impact of content democratization—how improvements in content quality q_L for type-L creators affect the equilibrium outcomes. In the Web Appendix, we show that our qualitative findings are robust in more general scenarios when the content-democratization technology simultaneously benefits all creators to various degrees (with potentially greater quality improvement for high-quality creators) and when the technology also increases the production efficiency of type-L creators, β_L.

Analysis

Equilibrium Analysis

To begin with, we solve the equilibrium outcome for fixed content quality q_j. Suppose that the number of type-j creators who produce is N_j, which is equal to the number of available content items with quality q_j. In a search attempt, consumers find some content with quality q_j with probability $Φ_{j} = N_{j} ρ_{j} = \frac{N_{j} q_{j}^{b}}{N_{H} q_{H}^{b} + N_{L} q_{L}^{b}}$ .

Consumer search

Now consider a consumer's search decision. Suppose that she has searched at least one content item. If she has found some high-quality content, she will stop searching and consume the content. If all she has found is low-quality content, the expected payoff of making another search attempt is $Φ_{H} (q_{H} - q_{L}) + Φ_{L} \times 0 - s_{i} = \frac{N_{H} q_{H}^{b}}{N_{H} q_{H}^{b} + N_{L} q_{L}^{b}} (q_{H} - q_{L}) - s_{i}$ .

Define
${\bar{s}}_{H} = \frac{N_{H} q_{H}^{b}}{N_{H} q_{H}^{b} + N_{L} q_{L}^{b}} (q_{H} - q_{L}) .$
(1)
Therefore, this consumer will make another search attempt if and only if $s_{i} \leq {\bar{s}}_{H}$ .

We can then derive a consumer's optimal search strategy and the expected utility. For consumers with a relatively low search cost $s_{i} \leq {\bar{s}}_{H}$ , if they make at least one search attempt, they will search until finding a high-quality content item, and in expectation they will make $\frac{1}{Φ_{H}} = \frac{N_{H} q_{H}^{b} + N_{L} q_{L}^{b}}{N_{H} q_{H}^{b}}$ search attempts and incur a total search cost of $\frac{s_{i}}{Φ_{H}}$ . Their payoff is zero if they do not search any content. Thus, their expected utility is $u_{i} = \max {q_{H} - \frac{s_{i}}{Φ_{H}}, 0} = q_{H} - s_{i} \frac{N_{H} q_{H}^{b} + N_{L} q_{L}^{b}}{N_{H} q_{H}^{b}}$ . The last equation is because the first term in the max function is always positive when $s_{i} \leq {\bar{s}}_{H}$ , so these consumers will always search at least one content item.

For those with a relatively high search cost $s_{i} > {\bar{s}}_{H}$ , they will consume the content they find regardless of its quality. Therefore, they will search at most one content item, which has high quality with probability $Φ_{H} = \frac{N_{H} q_{H}^{b}}{N_{H} q_{H}^{b} + N_{L} q_{L}^{b}}$ , and has low quality with probability $Φ_{L} = \frac{N_{L} q_{L}^{b}}{N_{H} q_{H}^{b} + N_{L} q_{L}^{b}}$ . Their expected payoff is thus $u_{i} = \max {\frac{N_{H} q_{H}^{b + 1} + N_{L} q_{L}^{b + 1}}{N_{H} q_{H}^{b} + N_{L} q_{L}^{b}} - s_{i}, 0}$ . Define
${\bar{s}}_{L} = \frac{N_{H} q_{H}^{b + 1} + N_{L} q_{L}^{b + 1}}{N_{H} q_{H}^{b} + N_{L} q_{L}^{b}} .$
(2)

They will make the search attempt if and only if ${\bar{s}}_{H} < s_{i} \leq {\bar{s}}_{L}$ , and will not make any search attempt if $s_{i} > {\bar{s}}_{L}$ . Note that ${\bar{s}}_{L}$ coincides with the expected content quality in the market. Intuitively, consumers will search if and only if the expected benefit from a search attempt is higher than the search cost.

In summary, consumers belong to one of three segments depending on their equilibrium search patterns. (a) “Extensive search” consumers are those with low search costs $s_{i} \leq {\bar{s}}_{H}$ , who search until finding high-quality content. Their population is $\frac{{\bar{s}}_{H}}{S} = \frac{N_{H} q_{H}^{b} (q_{H} - q_{L})}{(N_{H} q_{H}^{b} + N_{L} q_{L}^{b}) S}$ . (b) “Search once” consumers are those with medium search costs ${\bar{s}}_{H} < s_{i} \leq {\bar{s}}_{L}$ , who search only once and consume the content regardless of its quality. Their population is $\frac{{\bar{s}}_{L} - {\bar{s}}_{H}}{S} = \frac{q_{L}}{S}$ . (c) “Inactive” consumers are those with high search costs $s_{i} > {\bar{s}}_{L}$ , who do not participate in the market. The total expected consumer surplus is hence
$W = \int_{0}^{{\bar{s}}_{H}} (q_{H} - s \times \frac{N_{H} q_{H}^{b} + N_{L} q_{L}^{b}}{N_{H} q_{H}^{b}}) \frac{1}{S} ds + \int_{{\bar{s}}_{H}}^{{\bar{s}}_{L}} (\frac{N_{H} q_{H}^{b + 1} + N_{L} q_{L}^{b + 1}}{N_{H} q_{H}^{b} + N_{L} q_{L}^{b}} - s) \frac{1}{S} ds = \frac{N_{H} q_{H}^{b + 2} + N_{L} q_{L}^{b + 2}}{2 S (N_{H} q_{H}^{b} + N_{L} q_{L}^{b})},$
(3)
where the two integrals are the total expected surplus of extensive-search consumers and of search-once consumers, respectively. The number of active consumers, which consists of these two types of consumers, is $P = \frac{{\bar{s}}_{L}}{S} = \frac{N_{H} q_{H}^{b + 1} + N_{L} q_{L}^{b + 1}}{S (N_{H} q_{H}^{b} + N_{L} q_{L}^{b})}$ .

Creator production and equilibrium outcome

We proceed to examine creators’ production decisions. The total viewership of all N_H high-quality content items is $\frac{{\bar{s}}_{H}}{S} + \frac{{\bar{s}}_{L} - {\bar{s}}_{H}}{S} \times \frac{N_{H} q_{H}^{b}}{N_{H} q_{H}^{b} + N_{L} q_{L}^{b}} = \frac{N_{H} q_{H}^{b + 1}}{S (N_{H} q_{H}^{b} + N_{L} q_{L}^{b})}$ , and the total viewership of all N_L low-quality content items is $\frac{{\bar{s}}_{L} - {\bar{s}}_{H}}{S} \times \frac{N_{L} q_{L}^{b}}{N_{H} q_{H}^{b} + N_{L} q_{L}^{b}} = \frac{N_{L} q_{L}^{b + 1}}{S (N_{H} q_{H}^{b} + N_{L} q_{L}^{b})}$ . Therefore, the viewership of a type-j content item is
$V_{j} = \frac{1}{N_{j}} \times \frac{N_{j} q_{j}^{b + 1}}{S (N_{H} q_{H}^{b} + N_{L} q_{L}^{b})} = \frac{q_{j}^{b + 1}}{S (N_{H} q_{H}^{b} + N_{L} q_{L}^{b})} .$
(4)

Type-j creators, whose population is Mθ_j, will produce if and only if c < V_j, which occurs with probability Pr(c < V_j) = β_jV_j. The equilibrium quantity of type-j content satisfies
$N_{j} * = M θ_{j} β_{j} V_{j} * = M θ_{j} β_{j} \times \frac{q_{j}^{b + 1}}{S (N_{H} * q_{H}^{b} + N_{L} * q_{L}^{b})} .$
(5)

The total profit for type-j creators is $Π_{j} * = N_{j} * \times E [V_{j} * - c ∣ c < V_{j} ] = \frac{N_{j} V_{j} }{2} = \frac{{(N_{j} )}^{2}}{2 M θ_{j} β_{j}}$ .

We can solve the equilibrium content quantity $N_{j}^{}$ using Equation 5 for j = H, L. Lemma 1 presents the equilibrium outcome. All technical details and proofs are relegated to the Web Appendix.
Lemma 1.
The equilibrium quantity for type-j content is $N_{j}^{} = \sqrt{\frac{M}{S}} \times \frac{β_{j} θ_{j} q_{j}^{1 + b}}{\sqrt{β_{H} θ_{H} q_{H}^{1 + 2 b} + β_{L} θ_{L} q_{L}^{1 + 2 b}}}$ . The number of active consumers is $P * = \frac{β_{H} θ_{H} q_{H}^{2 + 2 b} + β_{L} θ_{L} q_{L}^{2 + 2 b}}{S (β_{H} θ_{H} q_{H}^{1 + 2 b} + β_{L} θ_{L} q_{L}^{1 + 2 b})}$ . The consumer surplus is $W * = \frac{β_{H} θ_{H} q_{H}^{3 + 2 b} + β_{L} θ_{L} q_{L}^{3 + 2 b}}{2 S (β_{H} θ_{H} q_{H}^{1 + 2 b} + β_{L} θ_{L} q_{L}^{1 + 2 b})}$ . The total payoff of type-j creators is $Π_{j} * = \frac{β_{j} θ_{j} q_{j}^{2 + 2 b}}{2 S (β_{H} θ_{H} q_{H}^{1 + 2 b} + β_{L} θ_{L} q_{L}^{1 + 2 b})}$ . The total social welfare is $SW * = W * + Π_{H} * + Π_{L} * = \frac{β_{H} θ_{H} q_{H}^{2 + 2 b} (1 + q_{H}) + β_{L} θ_{L} q_{L}^{2 + 2 b} (1 + q_{L})}{2 S (β_{H} θ_{H} q_{H}^{1 + 2 b} + β_{L} θ_{L} q_{L}^{1 + 2 b})}$ .

Impact of the Content-Democratization Technology

The content-democratization technology improves the content quality created by type-L creators, q_L. In this section, we examine the technology's marginal equilibrium effects, that is, how market outcomes change when q_L increases by a small amount. In reality, the content-democratization technology may increase q_L discretely, and its equilibrium effects will be the integration of the marginal effects. In the Web Appendix, we show that the qualitative insights remain unchanged when q_H also increases (potentially by more than the increase in q_L).

We start by examining how a higher q_L affects the production decisions for both type-L and type-H creators in Lemma 2.
Lemma 2.
As the quality of the low-quality content, q_L, increases, more type-L creators and fewer type-H creators will produce in equilibrium: $\frac{{dN}_{L}^{}}{{dq}_{L}} > 0$ and $\frac{{dN}_{H}^{}}{{dq}_{L}} < 0$ .

Lemma 2 reveals that the content-democratization technology has opposite effects on different types of creators’ production decisions: More type-L creators and fewer type-H creators will produce in equilibrium. The reason is twofold. First, the technology shrinks the quality gap between high-quality and low-quality content q_H − q_L, so when finding low-quality content in a search attempt, more consumers will consume it instead of being fixated on high-quality content. This can be seen from the observation that as q_L increases, the number of extensive-search consumers, $\frac{{\bar{s}}_{H}}{S}$ , decreases, whereas the number of search-once consumers, $\frac{{\bar{s}}_{L} - {\bar{s}}_{H}}{S}$ , increases. Second, when b > 0, an increase in q_L makes consumers more likely to find low-quality content in a search attempt: $\frac{d ρ_{L}}{{dq}_{L}} > 0$ and $\frac{d ρ_{H}}{{dq}_{L}} < 0$ . Both effects incentivize type-L creators but discourage type-H creators from producing content.
Proposition 1
(welfare effects of the content-democratization technology). As q_L increases, consumer surplus W, the number of active consumers P, the total profit of creators Π, and social welfare SW will first decrease and then increase.⁵ The average quality of all produced content, $\frac{N_{L}^{} q_{L} + N_{H}^{} q_{H}}{N_{L}^{} + N_{H}^{}}$ , will also first decrease and then increase.

One might expect that the content-democratization technology, by improving the quality of low-quality content without reducing the quality of high-quality content, would improve consumer surplus, creators’ total profit, and social welfare. Proposition 1 highlights a key finding of our article: The content-democratization technology will reduce the payoffs of consumers and creators when the quality gap between high-quality and low-quality content is large (i.e., when q_L is much lower than q_H); by contrast, the technology will benefit consumers when the quality gap is small. See Figure 1 for an illustration. The result stems from two effects of the improvement in q_L on consumers’ content consumption. First, it has a quality-improvement effect, which directly increases consumers’ welfare when they consume low-quality content. Second, it has a crowd-out effect, which increases the quantity of low-quality content $N_{L} $ but reduces that of high-quality content $N_{H}^{}$ available in the market (see Lemma 2). Therefore, consumers are more likely to find and thus end up consuming low-quality content instead of high-quality content in equilibrium. Mathematically, Equation 4 indicates that when N_L increases and N_H decreases, the total viewership of type-L content (N_LV_L) increases but the total viewership of type-H content (N_HV_H) decreases. The crowd-out effect tends to hurt consumers by reducing the number of active consumers (P), and by reducing the expected content quality consumed by active consumers.

Figure 1.
Equilibrium Effects of q_L and Screening Efficiency Level b.

Next, we explain how the content-democratization technology affects consumer surplus W* as q_L rises from a low level to a high level. When q_L is much lower than q_H, the crowd-out effect significantly reduces consumer surplus, as any replacement of high-quality content with low-quality content lowers the number of active consumers P* and their average consumed content quality. Furthermore, when q_L is low, little low-quality content will be produced, and thus the benefit of the quality-improvement effect on consumer surplus is weak. Hence, the improvement in q_L will reduce consumer surplus. By contrast, when the content-democratization technology significantly increases q_L such that the quality gap between high- and low-quality content becomes sufficiently small, consumer surplus W* as well as average quality will eventually increase. In the limit case of q_L → q_H, the crowd-out effect does not harm consumer surplus at all. The number of active consumers also increases. In a similar vein, we also find that as q_L increases, the average quality of available content in the market, $\frac{N_{L} * q_{L} + N_{H} * q_{H}}{N_{L} * + N_{H} }$ , will also first decrease and then increase.

Proposition 1 also examines how the increase in q_L affects creators’ total profit and social welfare. The content-democratization technology, by improving q_L, increases the total profit for type-L creators but decreases the total profit for type-H creators due to the crowd-out effect, as expected. Interestingly, Proposition 1 shows that their total profit Π will decrease if q_L is low and will increase if q_L is high. Intuitively, creators’ revenue depends on the total viewership of all content, or equivalently the number of active consumers P, which decreases with q_L when q_L is low and increases with q_L when q_L is high, as discussed previously. For similar reasons, the content-democratization technology will decrease social welfare SW when q_L is low and increase it when q_L is high.

The previous analysis assumes that all low-type creators adopt the technology, which is for simplicity of presentation and is without loss of generality: If we endogenize the adoption decision, all type-L creators will choose to use the technology if there is no direct adoption cost, because doing so increases the quality of, and thus the payoff for, each individual creator. However, each creator's adoption generates negative externalities (congestion) by crowding out high-quality content production, which can ultimately reduce the number of active consumers and the total payoff for all creators. In the Web Appendix, we show that this qualitative insight holds even when content democratization increases the quality of both low- and high-quality creators by the same magnitude. In this scenario, adopting the technology is a dominant strategy for each individual creator, who will unilaterally profit from adoption, yet the total payoff of all creators can still decrease.

It is important to note that the potential negative welfare impact of the content-democratization technology stems from (a) creators’ endogenous content production and (b) the potential dependency of the sampling probability of a content item (ρ_j) on its quality. In fact, if both factors are absent—content quantities (N_j) are fixed and all content items have the same sampling probability (b = 0)—an increase in q_L will always increase the consumer surplus (see Equation 3). The discussion of Proposition 1 mostly focuses on the effect of creators’ endogenous production (i.e., the crowd-out effect). The “Content Quality Screening” extension in the next section illustrates in detail how the dependence of ρ_j on content quality moderates the impact of content democratization.

In summary, our analysis reveals potential unintended consequences of the content-democratization technology despite its direct benefit of improving content quality. By increasing the quality of the low-quality content, the technology facilitates the production of low-quality content, which crowds out the production of high-quality content in the market. The proliferation of low-quality content will reduce consumers’ chance of finding high-quality content, their consumption of it, and their market participation, which can eventually hurt consumers and society if the technology leads to only incremental quality improvement (i.e., the quality gap between low- and high-quality content is still large after the technology is introduced). In contrast, the technology will benefit consumers and society if it significantly narrows the gap between high- and low-quality content.

Other Factors Affecting the Welfare Impacts of Content Democratization

This section extends the base model to identify how several important market factors can influence the welfare implications of content democratization: First, we examine how content quality screening moderates the welfare effects of the content-democratization technology, and also extend the model to incorporate a content platform and its endogenous decision on its screening technology. Second, we show that our results generalize to the case where there are multiple K > 2 content quality types. Third, we explore the impact of the technology's accessibility for creators. Fourth, we study the technology's impact on horizontal content variety in the market. Fifth, we consider consumers’ aversion to content generated by the technology. Lastly, we show the robustness of our results when creators endogenously choose content quality.

Content Quality Screening

We have illustrated how content democratization can hurt consumers by causing an influx of low-quality content to the market, crowding out production and consumption of high-quality content. This poses a threat to content platforms, whose success relies on not only content quantity but also its quality. To help their customers reliably find high-quality content, platforms implement many measures to effectively screen content based on its quality (Berman and Katona 2020). For example, platforms develop AI-based algorithms to more reliably identify high-quality content and thereby distribute it to more consumers.

This extension studies how content quality screening moderates the welfare effects of the content-democratization technology. More effective quality screening enables consumers to find high-quality content more easily, which is reflected in our model as an increase in b. We first treat b as an exogenous parameter and examine its impact on the market outcome. Then, we incorporate a content platform into the model and examine how content democratization influences the platform's endogenous quality-screening investment and the resulting market outcomes.

Proposition 2 summarizes how the screening effectiveness b affects the market outcomes and moderates the welfare effects of the content-democratization technology.
Proposition 2.

More effective content screening (higher b) increases consumer surplus W, creators’ total profit Π, and social welfare SW.

The negative impacts of q_L on consumer surplus, creators’ total profit, and social welfare are more likely to occur when screening is more effective. Specifically, there exist thresholds ${\bar{b}}_{1}, {\bar{b}}_{2}$ , and ${\bar{b}}_{3}$ , such that dW/dq_L ≤ 0 if and only if $b \geq {\bar{b}}_{1}$ , dΠ/dq_L ≤ 0 if and only if $b \geq {\bar{b}}_{2}$ , and dSW/dq_L ≤ 0 if and only if $b \geq {\bar{b}}_{3}$ .

Part i of Proposition 2 shows that when content screening is more effective in selecting high-quality content (b is higher), consumers are better off because now they are more likely to encounter high-quality content rather than low-quality content in a search attempt. Creators’ total profit and social welfare also increase. Correspondingly, one may expect that the content-democratization technology will become less likely to harm consumers and society, as consumers are less likely to encounter low-quality content. Interestingly, part ii of Proposition 2 suggests that more effective quality screening in fact strengthens the tendency of the technology to harm consumers and society. This is because when quality screening is more effective, an improvement in q_L more significantly raises consumers’ probability of encountering low-quality content in their search process, which substantially strengthens type-L creators’ production incentive but weakens that of type-H creators. Therefore, the crowd-out effect of the content-democratization technology becomes more pronounced when b is larger, so the negative welfare impact is more likely to occur. Technically speaking, if q_L increases to q_L(1 + ɛ), the probability for consumers to find a low-quality content item in a search attempt, ρ_L, will increase to approximately (1 + ɛ)^b times, which is higher with a larger b, other things being equal. Figure 1 illustrates the result: When b is higher, consumer surplus W* and creators’ total profit Π* are higher, but they decrease in q_L for a wider region of q_L.

It is important to note that our finding does not imply that content screening harms society; enhanced screening effectiveness will indeed increase consumer surplus and social welfare for a given q_L, as indicated by part i of Proposition 2. Instead, the implication is that when content platforms can more effectively screen content, the content-democratization technology becomes more likely to harm consumers and society.

The previous analysis treats the screening effectiveness b as an exogenous parameter. We now extend the model to incorporate a content platform and its endogenous investment in screening effectiveness and analyze how the content-democratization technology affects its content-screening investment and the resulting market outcomes. We assume that the platform can endogenously choose its screening effectiveness b at a fixed cost C(b), which strictly increases in b and is twice continuously differentiable. For illustration, in the main text, we assume that the platform's payoff is Ω(b) = W(b) − C(b). Although we do not explicitly model platform monetization, it is reasonable to assume that the platform benefits from higher consumer surplus W, which enables it to charge higher subscription fees or display more ads.⁶

We examine how the content-democratization technology, by increasing q_L, affects the platform's optimal screening effectiveness b(q_L) and its optimal payoff Ω(q_L) ≜ Ω(b(q_L); q_L), holding other parameters constant. For tractability, we assume that for any q_L ∈ (0, q_H), b(q_L) can be uniquely solved by the first-order condition $\frac{\partial W}{\partial b} (b * (q_{L}); q_{L}) - \frac{\partial C}{\partial b} (b * (q_{L})) = 0$ . Proposition 3 summarizes the result.
Proposition 3.

When q_L is sufficiently low, the content-democratization technology increases the platform's optimal screening effectiveness b* and reduces its total payoff $Ω^{} : \frac{{db}^{}}{{dq}_{L}} > 0$ and $\frac{d Ω^{}}{{dq}_{L}} < 0$ .

When q_L is sufficiently close to q_H, the content-democratization technology decreases the platform's optimal screening effectiveness b and increases its total payoff $Ω^{} : \frac{{db}^{}}{{dq}_{L}} < 0$ and $\frac{d Ω^{}}{{dq}_{L}} > 0$ .

One might expect that as q_L increases, the platform would have less need to screen out low-quality content and thus choose a lower screening effectiveness b. Interestingly, Proposition 3 shows that if q_L is initially much lower than q_H, the platform should increase b when q_L increases. This occurs because for a given b, an increase in q_L can significantly reduce consumers’ expected surplus from the platform due to the crowd-out effect. To mitigate this negative impact, the platform increases b to partially restore the likelihood for consumers to find high-quality content. By contrast, if q_L is already close to q_H, content screening becomes less necessary, so the platform should reduce b to save on investment costs when q_L further increases. Finally, the content-democratization technology decreases the platform's total payoff when q_L is much lower than q_H but increases its payoff when q_L and q_H are already close, which is consistent with our main findings about the welfare impact of content democratization.

Our results caution a content platform that even when content-democratization technology improves content quality across all creators, it may also trigger a redistribution of content quality on the platform. Therefore, the platform should increase its screening investment when the technology only incrementally narrows the quality gap between the low- and high-quality creators and reduce the investment when the technology drastically shrinks the quality gap. The platform's optimal screening effectiveness reaches its peak when the quality gap is at an intermediate level.

Multiple K ≥ 2 Content Types

This section generalizes the main model to K ≥ 2 content quality types to show the robustness of our qualitative insights. The equilibrium outcomes of this general setting will also be useful for later extensions. The qualities of the K content types are q₁ < q₂ < ··· < q_K, and the number of type-j creators is Mθ_j, where θ_j > 0 and $Σ_{j = 1}^{K} θ_{j} = 1$ . Other settings are the same as the main model. One can derive that the probability for consumers to find a type-j content item in a search attempt is $ρ_{j} = \frac{q_{j}^{b}}{Σ_{k = 1}^{K} N_{k} q_{k}^{b}}$ , and the probability of finding some type-j content in a search attempt is $Φ_{j} = N_{j} ρ_{j} = \frac{N_{j} q_{j}^{b}}{Σ_{k = 1}^{K} N_{k} q_{k}^{b}}$ .

We present the detailed analysis in the Web Appendix. In equilibrium, depending on their search costs, consumers belong to one of the K + 1 segments differing in their equilibrium search patterns. Segment-0 consumers have the highest search cost and will stay inactive. Segment-k consumers with 1 ≤ k ≤ K will search for content, stop searching when finding content whose quality is at least q_j ≥ q_k, and keep searching if q_j < q_k; consumers in a segment with a higher k have lower search costs.⁷ The following lemma summarizes the equilibrium outcome of the general model. The equilibrium outcome of the main model (Lemma 1) is a special case with K = 2 content types.
Lemma 3.
The equilibrium quantity for type-j content is $N_{j}^{} = \sqrt{\frac{M}{S}} \times \frac{β_{j} θ_{j} q_{j}^{1 + b}}{\sqrt{Σ_{k = 1}^{K} β_{k} θ_{k} q_{k}^{1 + 2 b}}}$ . The number of active consumers is $P * = \frac{Σ_{k = 1}^{K} β_{k} θ_{k} q_{k}^{2 + 2 b}}{S \times Σ_{k = 1}^{K} β_{k} θ_{k} q_{k}^{1 + 2 b}}$ The consumer surplus is $W * = \frac{Σ_{k = 1}^{K} β_{k} θ_{k} q_{k}^{3 + 2 b}}{2 S \times Σ_{k = 1}^{K} β_{k} θ_{k} q_{k}^{1 + 2 b}}$ . The total payoff of type-j creators is $Π_{j}^{} = \frac{β_{j} θ_{j} q_{j}^{2 + 2 b}}{2 S \times Σ_{k = 1}^{K} β_{k} θ_{k} q_{k}^{1 + 2 b}}$ . The total social welfare is ${SW}^{} = \frac{Σ_{k = 1}^{K} β_{k} θ_{k} q_{k}^{2 + 2 b} (q_{k} + 1)}{2 S \times Σ_{k = 1}^{K} β_{k} θ_{k} q_{k}^{1 + 2 b}}$ .

We illustrate the effect of the content-democratization technology on the market outcomes by focusing on the following case. When the technology is absent, the qualities of the K content types are equidistantly distributed: $q_{j} = \underline{Q} + \frac{j - 1}{K - 1} (\bar{Q} - \underline{Q})$ ; the lowest content quality is $q_{1} = \underline{Q}$ and the highest is $q_{K} = \bar{Q}$ . The quantity $1 - \underline{Q} / \bar{Q}$ measures the (relative) quality dispersion of different content types, analogous to the quality gap q_H − q_L in the main model. The technology improves the quality of type-j content to ${q^{'}}_{j} = (1 - δ) q_{j} + δ q_{K}$ ; the quality improvement for type-j content is $Δ_{j} = {q^{'}}_{j} - q_{j} = δ (q_{K} - q_{j})$ , where δ > 0 measures the magnitude of improvement. This setup reflects that the content-democratization technology more significantly improves the quality of content with lower quality $(\frac{d Δ_{j}}{{dq}_{j}} = - δ)$ . To obtain closed-form solutions, we assume that all creator types have the same population and the same production efficiency: β_j = β and θ_j = θ for all j = 1, ···, K, and that the probability for consumers to find a content item, ρ_j, is proportional to its quality: b = 1.⁸

We still consider the marginal effect of content-democratization technology, that is, the situation when δ is small. We find that the qualitative insight of the main model is robust. The content-democratization technology will hurt consumer surplus and social welfare when the quality dispersion $1 - \underline{Q} / \bar{Q}$ is large, in which case the negative impact of the crowd-out effect is significant. By contrast, the content-democratization technology will benefit consumers, creators, and society when the quality dispersion is small. Proposition 4 summarizes the result.
Proposition 4.
The content-democratization technology will reduce consumer surplus, creators’ total profit, and social welfare when the content quality dispersion $1 - \underline{Q} / \bar{Q}$ is sufficiently large. The technology will increase consumer surplus and creators’ total profit when the content quality dispersion is sufficiently small.

Accessibility of the Content-Democratization Technology

In our base model, we have assumed that all low-quality creators have access to the content-democratization technology. In practice, however, access to the technology is not evenly distributed, generating a disparity between creators.⁹ How does creators’ accessibility of the content-democratization technology affect its welfare impact? Will low accessibility mitigate the potentially negative welfare impact of the technology?

To answer these questions, we consider an extended model in which only a proportion γ ∈ (0, 1] of type-L creators can access the technology, where the parameter γ captures the accessibility of the technology to type-L creators. For convenience, in this section, we label the low-type creators who can utilize the technology as type-M creators and the remaining low-type creators still as type-L creators. As such, the population of type-M creators is θ_LγM, and the technology improves their content quality to q_M. Following the main analysis, we continue to assume q_M < q_H. The population of type-L creators is θ_L(1 − γ)M, and their content quality remains q_L. The numbers of type-M and type-L creators who produce content are denoted by N_M and N_L, respectively.

The equilibrium outcomes for given values of q_L and q_M are given in Lemma 3 with K = 3 creator types. We are interested in how improving q_M (while keeping q_L unchanged) would impact consumer and social welfare, and how the result is moderated by the accessibility of the technology (measured by γ).
Lemma 4.
As the content quality of the low-quality creators who use the technology, q_M, increases, more type-M creators, fewer type-L creators, and fewer type-H creators will produce in equilibrium: $\frac{{dN}_{M}^{}}{{dq}_{M}} > 0, \frac{{dN}_{L}^{}}{{dq}_{M}} < 0$ and $\frac{{dN}_{H}^{}}{{dq}_{M}} < 0$ .

The implication of Lemma 4 is similar to that of Lemma 2: Improving one type's quality draws more consumers to consume this type of content and crowds out other types, which increases the production of the former but decreases the latter. Figure 2 illustrates the result. Proposition 5 formally presents the welfare implications.

Figure 2.
Equilibrium Effects of q_M on Content Quantities.
Proposition 5.

As q_M increases, consumer surplus W, creators’ total profit Π, and social welfare SW first decrease and then increase;

The negative welfare impact of the content-democratization technology is more likely to occur when the technology is less accessible (γ is smaller);

Higher accessibility of the technology may reduce welfare when the quality improvement (q_M − q_L) is sufficiently small.

Part i of Proposition 5 and Figure 3 suggest that the welfare effect of the technology is preserved qualitatively in this extension and showcases that the content-democratization technology can reduce consumer surplus and creators’ profit if the improved quality q_M is much lower than q_H. Similar to the base model, the result is driven by the crowd-out effect of the technology. Thus, one might intuit that when the technology is more accessible to creators (i.e., γ is larger), the crowd-out effect is more pronounced, so the technology is more likely to reduce welfare (i.e., $\frac{{dW}^{}}{{dq}_{M}}$ and $\frac{dSW }{{dq}_{M}}$ are more likely to be negative). Interestingly, part ii of Proposition 5 and Figure 3 suggest the opposite: It is when γ is smaller that the technology tends to hurt consumers and social welfare. The intuition is as follows. Note that this negative impact due to the crowd-out effect (relative to the technology's benefits) is more salient when there is more high-quality content remaining in the market to be crowded out. As a result, our key result delineated in the base model (i.e., the technology hurts welfare) is more likely to occur with a lower γ—in this case, more high-quality creators tend to stay in the market. Put differently, our base model (with γ = 1) has focused on a case where the surprising result is least likely to arise.

Figure 3.
Impact of q_M on Consumer Surplus and Creators’ Profit.

In the previous discussion, we focused on the moderating role of accessibility γ on the welfare impact of the technology. Next, we discuss how a change in γ directly affects consumer welfare and creator profit. Part iii of Proposition 5 illustrates that, when q_M is low, consumer welfare and creator profit may decrease in the accessibility of the technology, γ. In this case, the negative impact of the crowd-out effect is salient because q_H − q_M is large, so the same intuition discussed previously applies: The more low-quality content creators adopt the technology, the more high-quality content will be crowded out, resulting in lower welfare. In this case, while higher accessibility reduces the likelihood of the negative welfare impact of the technology, it directly decreases welfare. By contrast, when q_M is sufficiently high, the negative impact of the crowd-out effect is weak, and thus an increase in technology accessibility can drive up the average content quality in the market, leading to higher welfare. In this case, higher accessibility not only lowers the incidence of a welfare loss brought by the technology but also directly increases welfare.

Content Variety

Besides affecting content quality, the content-democratization technology often influences the variety of content in the market. In practice, content created by the same technology can be highly homogeneous and repetitive in format, style, and substance. For example, most GenAI images tend to share “a kind of soft focus, airless, fantastical quality that render them easily identifiable” (Wilson 2023), and they have significantly lower content and visual novelty than human-created artwork, on average (Zhou and Lee 2024). Members of the film and television industries also worry that the increasing reliance on GenAI in plot and script creation will produce more formulaic films and TV shows (Davenport and Bean 2023). Liu, Wang, and Yang (2025) find that a temporary national ban on ChatGPT in Italy significantly increased the linguistic variety of Instagram marketing posts by restaurants, which in turn improved consumer engagement with these posts. When content-democratization technologies affect not only content quality but also horizontal variation, how do creators change their production decisions and technology adoptions, and what are the welfare consequences?

To examine these questions, we extend the base model by incorporating the horizontal varieties of content and consumers’ heterogeneous horizontal preferences for content. We focus on the case with two quality types without content screening (b = 0) for tractability. Consumer i's valuation for a type-j content item (j ∈ {L, H}) is given by v_ij = q_j + e_ij, where e_ij represents the idiosyncratic horizontal match between consumers and content. The random variable e_ij is independently and identically distributed among consumers and type-j content items, although their distributions may vary with j, reflecting potential differences in content variety across content types. For technical tractability, for content type j, let e_ij be a mean-zero discrete random variable with Z_j possible realizations, $e_{j}^{1} < e_{j}^{2} < \dots < e_{j}^{Z_{j}}$ , and their probability masses are $ω_{j}^{z} ≜ Pr (e_{ij} = e_{j}^{z})$ for 1 ≤ z ≤ Z_j. Thus, the mean of e_ij is $Σ_{z = 1}^{Z_{j}} ω_{j}^{z} e_{j}^{z} = 0$ , and its variance is $σ_{j}^{2} = Σ_{z = 1}^{Z_{j}} ω_{j}^{z} (e_{j}^{z})^{2}$ , which captures the horizontal variation within type-j content. When Z_j is large, our setup can approximate the case where e_ij follows a continuous distribution. We assume $- e_{j}^{1} < q_{j}$ to ensure that content does not yield a negative valuation for consumers. Content democratization increases q_L by δ_q but reduces σ_L by δ_σ. Similar to previous analyses, we consider the marginal effects of content democratization: δ_q and δ_σ are small. Let $ψ ≜ | \frac{δ_{σ}}{δ_{q}} |$ denote the magnitude of the horizontal content variety reduction relative to the quality improvement brought by the democratization technology.

In our setup, a consumer faces Z_H + Z_L possible content values. Z_H of these values are from type-H content, accounting for possible realizations of horizontal match values: $q_{H} + e_{H}^{1}, q_{H} + e_{H}^{2}, \dots, q_{H} + e_{H}^{Z_{H}}$ , and their content quantities are $N_{H} ω_{H}^{1}, \dots, N_{H} ω_{H}^{Z_{H}}$ , respectively. Similarly, Z_L of these values are from type-L content: $q_{L} + e_{L}^{1}, q_{L} + e_{L}^{2}, \dots, q_{L} + e_{L}^{Z_{L}}$ , and their content quantities are $N_{L} ω_{L}^{1}, \dots, N_{L} ω_{L}^{Z_{L}}$ , respectively. Following analogous steps in the analysis of the case with multiple K ≥ 2 content types, we derive the number of consumers who end up consuming type-j content with match $e_{j}^{z}$ as
$\frac{N_{j} ω_{j}^{z} (q_{j} + e_{j}^{z})}{S (N_{H} + N_{L})},$
(6)
which increases with the match value $e_{j}^{z}$ , other things being equal. Intuitively, consumers are more likely to consume a piece of content they encounter (instead of skipping it) when it has a better horizontal match.

The total consumer surplus can be derived as
$W = \frac{Σ_{j = H, L} (Σ_{z = 1}^{Z_{j}} N_{j} ω_{j}^{z} {(q_{j} + e_{j}^{z})}^{2})}{2 S (N_{H} + N_{L})} = \frac{Σ_{j = H, L} N_{j} (q_{j}^{2} + σ_{j}^{2})}{2 S (N_{H} + N_{L})} .$
(7)

The total viewership of all type-j content is $Σ_{z = 1}^{Z_{j}} \frac{N_{j} ω_{j}^{z} (q_{j} + e_{j}^{z})}{S (N_{H} + N_{L})} = \frac{N_{j} q_{j}}{S (N_{H} + N_{L})}$ , and a type-j creator's viewership is $V_{j} = \frac{1}{N_{j}} \frac{N_{j} q_{j}}{S (N_{H} + N_{L})} = \frac{q_{j}}{S (N_{H} + N_{L})}$ , which shares the same expression as in the main model (see Equation 4) and is independent of the level of variation $σ_{j}^{2}$ . This is because a larger variation has two opposing effects on content viewership—it increases viewership from consumers with a good match and decreases viewership from consumers without a good match. In our model, the two effects cancel out, so the level of content variation does not affect content viewership. The variation $σ_{j}^{2}$ affects only whether the viewership comes evenly from all consumers or mostly from a small set of consumers with high content match. Because creators’ revenue depends only on their total viewership rather than its composition, creators’ production decisions, their equilibrium payoffs $(π_{j} )$ , and the equilibrium quantities of different content types $(N_{j} )$ also remain identical to the main model, as given in Lemma 1.¹⁰

Substituting $N_{j}^{}$ into the expression of W, the equilibrium consumer surplus is given as
$W^{} = \frac{β_{H} θ_{H} q_{H} (q_{H}^{2} + σ_{H}^{2}) + β_{L} θ_{L} q_{L} (q_{L}^{2} + σ_{L}^{2})}{2 S (β_{H} θ_{H} q_{H} + β_{L} θ_{L} q_{L})} .$
(8)

Interestingly, greater content variation $σ_{j}^{2}$ , while not affecting creators’ content production, increases consumer surplus W. The reason is as follows. With greater content variation, consumers become more likely to encounter highly matched and highly unmatched content, and less likely to encounter “mediocre” content. As discussed after Equation 6, consumers are more likely to consume a highly matched piece of content they encounter than a highly unmatched piece of content. Therefore, the greater content variation shifts consumers’ consumption from “mediocre” content to highly matched content more significantly than to highly unmatched content. Thus, the higher consumption of highly matched content outweighs the higher consumption of highly unmatched content, leading to an increase in consumer surplus.

Consequently, if content democratization significantly reduces the variation of type-L content, consumer surplus and social welfare will decrease. By contrast, if the content-democratization technology only slightly decreases (or even improves) content variation of type-L content, consumers and society can become better off if the quality gap between low- and high-quality content becomes sufficiently small. Proposition 6 formally summarizes the welfare impact of content democratization when it also affects content variation.
Proposition 6.
The presence of the content-democratization technology reduces consumer surplus and social welfare when the technology significantly reduces the variation of type-L content (ψ is sufficiently large).

Consumer Aversion to Technology Usage

In practice, consumers may be able to recognize whether a piece of content is generated with the assistance of the technology. Additionally, platforms such as TikTok and YouTube require creators to disclose when their content is AI-generated and employ automated algorithms to detect and label such content for consumers (TikTok 2024; YouTube 2024). In such cases, consumers may value content less just because* it is labeled as AI generated (vs. human generated), all else being equal (Béchard and Kreiman 2025; Carney, Riveros, and Tully 2024; Millet et al. 2023). This section examines the implications of consumer aversion to content created by the technology.

Formally, suppose that the type-L creators can use the technology to improve the content quality to q_L + δ. Consumers discount the value of content by a factor µ ∈ [0, 1] if it is produced by the content-democratization technology—that is, consumers’ valuation is only ${\tilde{q}}_{L} = μ (q_{L} + δ)$ , which can be considered the content's effective quality due to the use of the technology. This formulation allows the content-democratization technology to increase the quality and consumers’ technology aversion simultaneously. To reflect the detection and labeling practices adopted by many content platforms, which inform consumers about and demote technology-generated content, we assume that if a type-L creator adopts the technology, the visibility of her content (i.e., its sampling probability ρ_j) and consumers’ search decisions are both determined by its effective quality ${\tilde{q}}_{L}$ , which factors in consumer aversion to the technology.

If consumers are too averse to using the technology in content creation (i.e., µ is too small), their valuation of the content can be below what it would have been in the absence of the technology, in which case ${\tilde{q}}_{L} = μ (q_{L} + δ) < q_{L}$ . In this case, the content-democratization technology can reduce the visibility of the AI-generated content, and as a result, the type-L creators would rather not use it. Only when the net effect is still positive in the sense that the perceived content quality is improved would a creator choose to employ the technology. Next, we summarize how consumers’ aversion to technology usage influences consumer surplus and social welfare.
Proposition 7.
When consumers can recognize and are averse to the usage of technology, consumer surplus and social welfare
remain unchanged when $μ \leq \frac{q_{L}}{q_{L} + δ}$ ;

improve when the technology is available (vs. not available) when µ or q_L is sufficiently large;¹¹

decrease when the technology is available (vs. not available) when q_L is sufficiently small and µ is intermediate and larger than $\frac{q_{L}}{q_{L} + δ}$ .

Part i confirms that when consumers are strongly averse to technology usage, type-L creators will not use the technology, and therefore all content quality and quantity stay the same with or without the technology, and consumer surplus and social welfare remain unchanged. Parts ii and iii speak to the scenario where creators choose to use the technology $(μ > \frac{q_{L}}{q_{L} + δ})$ . In this case, Proposition 7 shows that the availability of the content-democratization technology tends to decrease welfare when consumers are more averse to technology usage (i.e., when µ is smaller). This is because the negative welfare impact of the crowd-out effect is more prominent. Note that when µ → 1⁻, this extended model reduces to our base model. This implies that our base model represents a case where it is least likely to observe the negative effect of technology on welfare.

Endogenous Content Quality

This extension studies creators’ endogenous decisions on their content quality. The content-democratization technology narrows the efficiency gap in quality provision between low- and high-skilled creators. Moreover, the technology also affects quality decisions of high-skilled creators, even if they may not adopt the technology. This extension examines whether the negative impact of the content-democratization technology will be attenuated if high-quality creators can increase their content quality to avoid being crowded out.

Specifically, we assume that each creator's content quality is a random variable with binary outcomes, with a high realization q = q_H with probability ν, and a low realization q = q_L with probability 1 − ν, where ν ∈ [0, 1]. The expected content quality is q = νq_H + (1 − ν)q_L. In this extension, we fix the values of q_H and q_L and capture variations in expected content quality through changes in ν. In this setting, creator types may differ from realized qualities, so to avoid confusion, we use {h, l} to denote creator types and use {H, L} to denote realized content quality levels. Each creator endogenously chooses ν ∈ [0, 1], which is referred to as the quality investment level. Creators’ cost function is specified as $κ_{j} (ν) = \frac{ν^{2}}{ζ_{j}} + ε$ , where ζ_h > ζ_l indicates that high-quality creators can more efficiently improve content quality, and ɛ ∼ Unif[0, 1] captures the quality-independent component of cost, which are heterogeneous among creators. Creators endogenously decide whether to produce content and, if so, their choice of ν. The content-democratization technology increases the quality provision efficiency of low-quality creators, ζ_l.

We solve for the equilibrium in which all type-j creators who decide to produce content will choose the same $ν_{j}^{}$ . Let $m_{j}^{}$ denote the equilibrium number of type-j creators who choose to produce content, where j ∈ {h, l}. In equilibrium, the quantities of high- and low-quality content are $N_{H}^{} = Σ_{j = h, l} m_{j}^{} ν_{j}^{}$ and $N_{L}^{} = Σ_{j = h, l} m_{j} * (1 - ν_{j}^{})$ , respectively. Consumers have rational expectations about $ν_{j} $ and $m_{j}^{}$ , and we assume that they do not update such beliefs based on the realized qualities of content they have searched. We assume that ζ _l and ζ_h are sufficiently small so that all creators will choose interior $ν_{j}^{} < 1$ in equilibrium. To obtain closed-form solutions, our analysis focuses on the situation without content filtering (b = 0), although one can numerically verify that the qualitative outcomes remain unchanged for b > 0. The detailed analysis is presented in the Web Appendix.

Proposition 8 summarizes the welfare impact of the content-democratization technology.
Proposition 8.
When creators endogenously decide quality, the equilibrium effects of content-democratization technology, captured as an increase in ζ _l , are summarized as follows:
When ζ _l increases, more type-l creators will produce content, and they will choose a higher quality investment level $ν_{l} * : \frac{{dm}_{l}^{}}{d ζ_{l}} > 0 and \frac{{d ν}_{l}^{}}{d ζ_{l}} > 0$ ; in contrast, fewer type-h creators will produce, and they will choose a lower quality investment level $ν_{h}^{} : \frac{{dm}_{h}^{}}{d ζ_{l}} < 0 and \frac{{d ν}_{h}^{}}{d ζ_{l}} < 0$ . The total amount of producing creators will increase: $\frac{d (m_{h}^{} + m_{l}^{})}{d ζ_{l}} > 0$ .

Moreover, suppose the quality gap between low- and high-quality content is sufficiently large, that is, q_L/q_H is sufficiently low. When ζ _l increases, both consumer surplus, W, and creators’ total profit, Π, will first decrease and then increase.

Proposition 8 demonstrates that our qualitative insights are robust when creators endogenously invest in content quality, provided that the quality gap between low- and high-quality content is significant and thus the crowd-out effect of content democratization is pronounced. More interestingly, we show that the high-skilled creators, instead of raising quality investment to mitigate the crowd-out effect of content democratization, will not only produce less content but also reduce quality investment. This is because more low-quality content dilutes high-skilled creators’ expected viewership, decreasing their marginal benefit of improving quality. The decreased quality investment aggravates the negative consequence of content democratization on consumer surplus and social welfare.

In the next section, we discuss how our results are related to various industry realities and empirical evidence, and how our framework can provide important managerial insights.

Managerial and Policy Implications

Content democratization empowers creators to produce content with better quality and at lower costs, narrowing the skill gap between novice and experienced creators. This article provides a game-theoretic framework to understand the welfare impacts of the content-democratization technology in different contexts. We show that the technology tends to benefit consumers and society (1) when there is a small quality gap between content types or when the technology can significantly narrow the quality gap, (2) when content platforms cannot easily screen content quality and distribute high-quality content, (3) when the accessibility of the technology to content creators is high enough, (4) when the content generated by the technology does not significantly reduce content variations, and (5) when consumers are not strongly averse to AI-generated content. Otherwise, the technology may hurt consumers and society. Next, we discuss how our framework is connected to several industry realities, using GenAI as a primary example.

While content democratization has been warmly embraced in some industries, other industries have been worried about its potential negative impacts. Recent empirical research has also produced mixed evidence on whether GenAI is beneficial or harmful across industries. Our framework helps explain these seemingly conflicting attitudes and evidence regarding content democratization. As mentioned in the introduction, the news industry is increasingly concerned that many “content farms” empowered by GenAI can produce numerous news articles with mediocre news values, which dilute consumers’ attention to original, high-quality news. This is because the quality of news crucially depends on whether news media can get firsthand, exclusive materials and synthesize and analyze them in depth. This job can be competently done by professional journalists (e.g., Wall Street Journal) but not by GenAI. Because of the large quality gap, content democratization in this case tends to be harmful to the news industry. However, in some areas, such as companies’ earnings reports and sports news, AI technologies can efficiently recap the earnings data or event outcomes and successfully generate high-quality news articles, and so news media have been using AI to write such news for years (Etherington 2014).

Similarly, in online communities such as Reddit or Stack Overflow, the impact of GenAI also depends largely on how the quality of AI-generated content compares to that without the assistance of AI. For example, Shorakaei et al. (2025) find that GenAI reduced posting activities on Reddit most significantly on topics where most content had been posted by high-skilled experts before GenAI; by contrast, GenAI increased posting activities in more objectively based topics, for which GenAI tended to generate answers of decent quality. Additionally, Shankar and Sim (2024) find that, in the absence of expert oversight over answers on Stack Overflow, GenAI only moderately improved the answer quality of novice contributors but did enable them to answer more questions, which ultimately lowered the overall answer quality on the platform.

The film industry also exhibits mixed attitudes toward the adoption of GenAI, depending on its application. The success of movies often depends on high-quality stories and productions. Many small-movie producers possess high-quality stories but are hindered by the production costs, which are often as high as millions of dollars (Shaw 2023). Recent AI technologies have enabled small producers to significantly reduce the cost and time required for many production stages, such as visual and sound effects, thereby allowing them to produce high-quality movies at a fraction of the traditional cost. For example, a small yet highly acclaimed production, Everything Everywhere All at Once, greatly benefited from using the cutting-edge AI visual-effect tool Runway, which shortened “days of work into minutes” (Tangcay 2023). The savings in production cost allow more high-quality movies to be produced, benefiting consumers and the industry. Meanwhile, as mentioned previously, the industry remains concerned about using GenAI to create stories, scripts, and dialogues. In this case, the adoption of GenAI can result in formulaic and repetitive movies (i.e., lower content variety), which leads to lower welfare, even if each movie maintains a decent quality level.

The welfare implications of the content-democratization technology in an industry depend on consumers’ inherent preferences for technology-generated content. For example, in different industries, consumers may have distinct attitudes toward AI-generated images even though they are generated with similar GenAI technologies. The art industry is more averse to AI-generated images because consumers often display a subjective bias against them even when their objective quality is higher than those created by human artists (Grassini and Koivisto 2024). Evidence also shows that image-generation AI has triggered negative consumer reactions, creator exodus, and declining platform performance across online art communities (e.g., Lin 2025; Matatov et al. 2024; Wei and Tyson 2024). By contrast, AI-generated images are better received in functional applications such as advertising. According to a survey, 60% of consumers support advertisers to use GenAI to create non-informational images (e.g., background images) in their ads (Yahoo and Publicis Media 2024). In such applications, Goldberg and Lam (2025) find that image-generation AI increased overall image quality and sales on a stock image platform, although it crowded out the production of non-AI-generated images. Our results reconcile these seemingly contradictory empirical findings on the welfare impacts of AI-generated images in different industries.

Given the mixed attitudes across industries toward GenAI, interventions regarding the disclosure of its use have emerged, thereby allowing the public to exercise discretion in consuming AI-assisted content. For instance, YouTube mandates creators to disclose certain AI-generated content such as “content that is meaningfully altered or synthetically generated when it seems realistic” (Google 2024), and social media platforms such as TikTok employ detection algorithms to identify and label AI-generated content (TikTok 2024; YouTube 2024). Our analysis in the “Consumer Aversion to Technology Usage” section implies that platform interventions that reduce traffic to AI-assisted content can alleviate the negative welfare impact of the technology.

Concluding Remarks

This article examines the influences of content-democratization technologies in content markets, epitomized by GenAI content production tools such as ChatGPT, Nano Banana, and DALL-E. These technologies have radically altered the online content landscape, fostering an environment with both innovation and disruption.

We reveal the nuanced welfare implications of content democratization. The technology, despite its improvement in content quality by low-quality creators, may harm consumers and society when the content quality gap between high- and low-skilled creators is large, when the quality improvement brought by the technology is small, when content platforms can effectively screen content quality, when the technology has limited accessibility to creators, when the technology significantly reduces content variations, and when consumers are inherently averse to technology-generated content. When content democratization results in a welfare loss, the negative consequences arise mainly because the proliferation of low-quality content crowds out the production of high-quality content. The decreased share of high-quality content in the market reduces consumers’ chance of finding and consuming high-quality content in their search process and also discourages them from market participation, which in turn hurts creators’ profit and social welfare. Our results can explain divergent attitudes toward content democratization by various industries.

In the Web Appendix, we also discuss other factors affecting the welfare impact of content-democratization technologies and confirm the robustness of our main insights in alternative model settings. First, we demonstrate that the content-democratization technology can reduce both consumer and social welfare even when it equally* enhances the quality of low-quality and high-quality content. Second, we examine a general setting where the technology can also increase the production efficiency of low-quality creators. We find that the crowd-out effect is stronger, making it more likely to reduce consumer surplus and the total profit of all creators. Third, when creators’ revenue per view increases with content quality, the technology is more likely to reduce welfare, because the technology increases not only the viewership but also the revenue per view of low-quality creators, further strengthening their production incentives and magnifying the crowd-out effect. Fourth, we show that our results are robust under an alternative demand formulation based on attention competition.¹²

It is worth noting that it is not our intent to suggest that policymakers should limit the development and usage of content-democratization technology. Instead, the welfare impacts discussed in our article reveal what policymakers could do to minimize the potential negative effects of content democratization. Additionally, while we use GenAI as the primary example, the results and the insights garnered from our model can be extrapolated to a broader context. As alluded to in the introduction, technologies that democratize content creation are not confined to GenAI alone. For example, free and easy-to-use video editing tools such as CapCut enable individuals who previously did not possess video editing skills to express their creativity and contribute to the rich tapestry of content. Our results suggest that the prevalent use of such tools may initially reduce welfare because improved mediocre content can crowd out high-quality content. Once the effectiveness (in terms of quality improvement) of such tools surpasses a threshold, further enhancement of the technologies can raise welfare.

Although our research does not account for content pricing, which aligns with the practices of most content markets, we conjecture that some of our insights may still apply to markets where consumers purchase content, such as online education platforms like Udemy. Consider a simple setting where the “market-level” price of content with quality q_j is kq_j, where k ∈ (0, 1). In this setting, a consumer's net utility from the content is (1 − k)q_j, and creators’ revenue per view is kq_j. Then, this setup can be transformed to a special case in the “Creators’ Revenue-per-View Depends on Content Quality” section of the Web Appendix, up to some rescaling of parameters. We acknowledge that while it is reasonable to expect content prices to increase with quality, this simplified setting does not capture creators’ endogenous pricing decisions. We leave the exploration of creators’ pricing strategies for future research.

We conclude by suggesting several directions for future research. First, our article concisely illustrates the nuanced welfare consequences of content democratization with a static model. Future research can extend our analysis to a dynamic setting in which consumers adapt to content democratization. For example, in response to the proliferation of low-quality content, consumers may learn to better discern content quality or concentrate their attention more on reputable content outlets. Second, creators may endogenously invest in improving their aptitude for using the content-democratization technology. Future research can examine how such investments depend on creator characteristics and the equilibrium implications of such investments. Lastly, our study focuses on the impact of content-democratization technology as it helps creators improve content quality and production efficiency. Future research can investigate other impacts that can be brought by content-democratization technologies or GenAI, such as how it connects to intellectual property infringement and the associated impact on market outcomes.

Supplemental Material

sj-pdf-1-mrj-10.1177_00222437261423540 - Supplemental material for Welfare Implications of Democratization in Content Creation: Generative AI and Beyond

Supplemental material, sj-pdf-1-mrj-10.1177_00222437261423540 for Welfare Implications of Democratization in Content Creation: Generative AI and Beyond by Tianxin Zou, Zijun (June) Shi and Yue Wu in Journal of Marketing Research

Symbol	Definition
i	An index that labels consumers
j	∈ {H, L}: An index that labels creator type
M	Population of all creators
θ_jM	Population of type-j creators
q_j	Content quality of a type-j creator, q_H > q_L
N_j	Amount of type-j content available in the market
c	$\in [0, {\bar{c}}_{j}]$ , a creator's cost of content creation
${\bar{c}}_{j}$	Upper bound of the content creation cost of type-j creators
β_j	= $\frac{1}{{\bar{c}}_{j}}$ , creation efficiency of type-j creators
ρ_j	A consumer's probability of encountering a piece of type-j content in a search attempt
b	Effectiveness of quality screening
Φ_j	= ρ_jN_j, a consumer's probability of encountering some type-j content in a search attempt
P	The number of active consumers
s_i	∈ [0, S], a consumer's search cost
S	The upper bound of search cost
V_j	The viewership of a type-j content item
W	Consumer surplus
SW	Social welfare
Π_j	Total profit for type-j creators
Π	Total profit for all creators
□*	Superscript that signifies equilibrium outcome

Footnotes

Acknowledgments

The authors thank Kaifu Zhang; seminar participants at Columbia University, Purdue University, Washington University in St. Louis, University of Connecticut, and University of Pittsburgh; and attendees of the 18th Annual UTD FORMS Conference, University of Texas at Dallas and 2024 AI in Management (AIM) Conference, University of Southern California. The authors used AI tools during manuscript preparation to search for information and references and generate draft wording for limited portions of the text, which were subsequently reviewed, edited, and verified by the authors. The tools were not used to generate data, analysis, or results, and the authors take full responsibility for the accuracy and integrity of the manuscript.

Coeditor

Brett Gordon

Associate Editor

Tony Ke

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

ORCID iDs

Tianxin Zou

Zijun (June) Shi

Yue Wu

Data Availability

No data were created or analyzed for this article.

Notes

References

Amaldoss

Wilfred

Shin

Woochoel

(2011), “Competing for Low-End Markets,” Marketing Science, 30 (5), 776–88.

Anderson

Simon P.

de Palma

André

(2009), “Information Congestion,” RAND Journal of Economics, 40 (4), 688–709.

Bajarin

Tim

(2024), “Nvidia’s CEO on the Democratization of Coding,” Forbes (March 25), https://www.forbes.com/sites/timbajarin/2024/03/20/nvidias-ceo-on-the-democratization-of-coding/.

Béchard

Deni Ellis

Kreiman

Gabriel

(2025), “People Want AI to Help Artists, Not Be the Artist,” Scientific American (September 7), https://www.scientificamerican.com/article/survey-results-show-people-prefer-more-human-involvement-in-ai-driven-art/, accessed on September 18, 2025.

Berman

Ron

Katona

Zsolt

(2020), “Curation Algorithms and Filter Bubbles in Social Networks,” Marketing Science, 39 (2), 296–316.

Brynjolfsson

Erik

Danielle

Raymond

Lindsey

(2025), “Generative AI at Work,” Quarterly Journal of Economics, 140 (2), 889–942.

Burtch

Gordon

Lee

Dokyun

Chen

Zhichen

(2024), “The Consequences of Generative AI for Online Knowledge Communities,” Scientific Reports, 14 (1), 10413.

Cao

H. Henry

Liye

Ning

Z. Eddie

Sun

Baohong

(2024), “How Does Competition Affect Exploration vs. Exploitation? A Tale of two Recommendation Algorithms,” Management Science, 70 (2), 1029–51.

Carney

Stephan

Riveros

Ignacio

Tully

Stephanie

(2024), “Made with AI: Consumer Engagement with Media Containing AI Disclosures,” SSRN, https://doi.org/10.2139/ssrn.4988760.

10.

Chen

Jieteng

T. Tony

Shin

Jiwoong

(2025), “Designing Detection Algorithms for AIGenerated Content: Consumer Inference, Creator Incentives, and Platform Strategy,” SSRN, https://doi.org/10.2139/ssrn.5271493.

11.

Chen

Zenan

Chan

Jason

(2024), “Large Language Model in Creative Work: The Role of Collaboration Modality and User Expertise,” Management Science, 70 (12), 9101–17.

12.

Choi

Jonathan H.

Schwarcz

Daniel

(2025), “AI Assistance in Legal Analysis: An Empirical Study,” Journal of Legal Education, 73 (2), 384–420.

13.

Choi

W. Jason

Liu

Qihong

Shin

Jiwoong

(2024), “Predictive Analytics and Ship-Then-Shop Subscription,” Management Science, 70 (2), 1012–28.

14.

Dai

Tinglong

Singh

Shubhranshu

(2025), “Artificial Intelligence on Call: The Physician’s Decision of Whether to Use AI in Clinical Practice,” Journal of Marketing Research, 62 (5), 854–75.

15.

Davenport

Thomas H.

Bean

Randy

(2023), “The Impact of Generative AI on Hollywood and Entertainment,” MIT Sloan Management Review (June 19), https://sloanreview.mit.edu/article/the-impact-of-generative-ai-on-hollywood-and-entertainment/, accessed on August 25, 2025.

16.

Dell’Acqua

Fabrizio

McFowland

Edward

Mollick

Ethan R.

Lifshitz-Assaf

Hila

Kellogg

Katherine

Rajendran

Saran

Krayer

Lisa

Candelon

François

Lakhani

Karim R.

(2026), “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality,” Organization Science, 37 (2), 403–23

17.

Desai

Preyas S.

(2001), “Quality Segmentation in Spatial Markets: When Does Cannibalization Affect Product Line Design?” Marketing Science, 20 (3), 265–83.

18.

Etherington

Darrell

(2014), “The AP Is Using Robots to Write Earnings Reports,” TechCrunch (July 1), https://techcrunch.com/2014/07/01/the-ap-is-using-robots-to-write-earnings-reports/.

19.

Goldberg

Samuel

Lam

H. Tai

(2025), “Generative AI in Equilibrium: Evidence from a Creative Goods Marketplace,” SSRN, https://doi.org/10.2139/ssrn.5152649.

20.

Google (2024), “Disclosing Use of Altered or Synthetic Content,” (accessed August 26, 2025), https://support.google.com/youtube/answer/14328491.

21.

Grassini

Simone

Koivisto

Mika

(2024), “Understanding how Personality Traits, Experiences, and Attitudes Shape Negative Bias Toward AI-Generated Artworks,” Scientific Reports, 14 (1), 4113.

22.

Hosanagar

Kartik

Ahn

Daehwan

(2024), “Designing Human and Generative AI Collaboration,” arXiv, https://doi.org/10.48550/arXiv.2412.14199.

23.

Huang

Raffaele

(2023), “TikTok’s Chinese Parent Has Another Wildly Popular App in the U.S.,” Wall Street Journal (March 19), https://www.wsj.com/articles/tiktoks-chinese-parent-has-another-wildly-popular-app-in-the-u-s-e14c41fc?.

24.

Hui

Xiang

Reshef

Oren

Zhou

Luofeng

(2024), “The Short-Term Effects of Generative Artificial Intelligence on Employment: Evidence from an Online Labor Market,” Organization Science, 35 (6), 1977–89.

25.

Instagram (2025), “Add Effects and Filters,” (accessed August 25, 2025), https://help.instagram.com/608433622656862.

26.

Ishibashi

Ikuo

Matsushima

Noriaki

(2009), “The Existence of Low-End Firms May Help High-End Firms,” Marketing Science, 28 (1), 136–47.

27.

Iyer

Ganesh

Katona

Zsolt

(2016), “Competing for Attention in Social Communication Markets,” Management Science, 62 (8), 2304–20.

28.

Jain

Sanjay

Qian

Kun

(2021), “Compensating Online Content Producers: A Theoretical Analysis,” Management Science, 67 (11), 7075–90.

29.

Ziwei

Lee

Nayeon

Frieske

Rita

Tiezheng

Dan

Yan

Ishii

Etsuko

Bang

Ye Jin

Madotto

Andrea

Fung

Pascale

(2023), “Survey of Hallucination in Natural Language Generation,” ACM Computing Surveys, 55 (12), https://doi.org/10.1145/3571730.

30.

Jing

Bing

(2006), “On the Profitability of Firms in a Differentiated Industry,” Marketing Science, 25 (3), 248–59.

31.

Krista J.

Changying

Zhang

Jianhu

(2023), “Artificial Intelligence: Information Collection and Behavior-Based Pricing Under Privacy Concerns,” working paper.

32.

Xiaolin

Liao

Chenxi

Xie

Ying

(2021), “Digital Piracy, Creative Productivity, and Customer Care Effort: Evidence from the Digital Publishing Industry,” Marketing Science, 40 (4), 685–707.

33.

Lin

Sijie

(2025), “Hiding from Generative AI,” tSE Digital Conference Paper, https://www.tse-fr.eu/sites/default/files/TSE/documents/conf/2025/digital/lin.pdf.

34.

Liu

Chaoran

Wang

Tong

Yang

S. Alex

(2025), “Generative AI and Content Homogenization: The Case of Digital Marketing,” SSRN, https://doi.org/10.2139/ssrn.5367123.

35.

Liu

Yildirim

Pinar

Zhang

John Z.

(2022), “Implications of Revenue Models and Technology for Content Moderation Strategies,” Marketing Science, 41 (4), 831–47.

36.

Matatov

Hana

Le Quéré

Marianne Aubin

Amir

Ofra

Naaman

Mor

(2024), “Examining the Prevalence and Dynamics of AI-Generated Media in Art Subreddits,” arXiv, https://doi.org/10.48550/arXiv.2410.07302.

37.

McCall

John J.

(1965), “The Economics of Information and Optimal Stopping Rules,” Journal of Business, 38 (3), 300–317.

38.

McMillan

Robert

(2023), “AI Junk Is Starting to Pollute the Internet,” Wall Street Journal (July 12), https://www.wsj.com/tech/chatgpt-already-floods-some-corners-of-the-internet-with-spam-its-just-the-beginning-9c86ea25? .

39.

Miklós-Thal

Jeanine

Tucker

Catherine

(2019), “Collusion by Algorithm: Does Better Demand Prediction Facilitate Coordination Between Sellers?” Management Science, 65 (4), 1552–61.

40.

Millet

Kobe

Buehler

Florian

Guanzhong

Kokkoris

Michail D.

(2023), “Defending Humankind: Anthropocentric Bias in the Appreciation of AI Art,” Computers in Human Behavior, 143, 107707.

41.

Mohammadi

Behnam

Malik

Nikhil

Derdenger

Timothy

Srinivasan

Kannan

(2024), “Regulating eXplainable Artificial Intelligence (XAI) May Harm Consumers,” Marketing Science, 44 (3), 711–24

42.

Noy

Shakked

Zhang

Whitney

(2023), “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence,” Science, 381 (6654), 187–92.

43.

Peng

Sida

Kalliamvakou

Eirini

Cihon

Peter

Demirer

Mert

(2023), “The Impact of AI on Developer Productivity: Evidence from GitHub Copilot,” arXiv, https://doi.org/10.48550/arXiv.2302.06590.

44.

Qian

Kun

Jain

Sanjay

(2024), “Digital Content Creation: An Analysis of the Impact of Recommendation Systems,” Management Science 70 (12), 8668–84.

45.

Radford

Alec

Jeffrey

Child

Rewon

Luan

David

Amodei

Dario

Sutskever

Ilya

(2019), “Language Models are Unsupervised Multitask Learners,” OpenAI Blog, 1 (8), https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf.

46.

Roose

Kevin

(2022), “An AI-Generated Picture Won an Art Prize. Artists Aren’t Happy,” New York Times (September 2), https://www.nytimes.com/2022/09/02/technology/ai-artificial-intelligence-artists.html.

47.

Rothschild

Michael

(1978), “Searching for the Lowest Price When the Distribution of Prices Is Unknown,” in Uncertainty in Economics. Academic Press, 425–54.

48.

Routley

Nick

(2023), “What Is Generative AI? An AI Explains,” World Economic Forum (February 6), https://www.weforum.org/agenda/2023/02/generative-ai-explain-algorithms-work/.

49.

Sadeghi

McKenzie

Arvanitis

Lorenzo

(2023), “Rise of the Newsbots: AI-Generated News Websites Proliferating Online,” NewsGuard (May 1), https: //www.newsguardtech.com/special-reports/newsbots-ai-generated-news-websites- proliferating/.

50.

Salop

Steven C.

(1979), “Monopolistic Competition with Outside Goods,” Bell Journal of Economics, 10 (1), 141–56.

51.

Shankar

Ramesh

Sim

Jaeung

(2024), “Generative AI on the Loose: Impact of Improved AI and Expert Oversight on Knowledge Sharing,” SSRN, https://doi.org/10.2139/ssrn.4867815.

52.

Shaw

Lucas

(2023), “Indie Films Are the Latest Casualty of the Streaming Wars,” Bloomberg (June 4), https://www.bloomberg.com/news/newsletters/2023-06-04/indie-films-are-the-latest-casualty-of-the-streaming-wars.

53.

Shorakaei

Hasan

Zhang

Mengxia

Bayley

Tiffany

Begen

Mehmet A.

(2025), “Empowering or Eroding Contributions? How Generative AI Impacts User-Generated Content Across Diverse Communities,” SSRN, https://doi.org/10.2139/ssrn.5191809.

54.

Shumailov

Ilia

Shumaylov

Zakhar

Zhao

Yiren

Gal

Yarin

Papernot

Nicolas

Anderson

Ross

(2023), “The Curse of Recursion: Training on Generated Data Makes Models Forget,” arXiv, https://doi.org/10.48550/arXiv.2305.17493.

55.

Sun

Monic

Zhu

Feng

(2013), “Ad Revenue and Content Commercialization: Evidence from Blogs,” Management Science, 59 (10), 2314–31.

56.

Tangcay

Jazz

(2023), “‘Hollywood 2.0’: How the Rise of AI Tools Like Runway Are Changing Filmmaking,” Variety (February 22), https://variety.com/2023/artisans/news/artificial-intelligence-runway-everything-everywhere-all-at-once-1235532322/.

57.

Teleprompter (2023), “TikTok 3 Second Rule: Mastering Rapid Engagement,” (March 21), https:// www.teleprompter.com/blog/tiktok-3-second-rule.

58.

TikTok (2024), “AI-Generated Content,” (accessed August 26, 2025), https://www.tiktok.com/tns-inapp/pages/ai-generated-content.

59.

Tullock

(1980), “Efficient Rent Seeking,” in Towards a Theory of the Rent-Seeking Society, Tollison

Buchanan

Tullock

, eds. Texas A&M University Press, 97–112.

60.

Van Zandt

Timothy

(2004), “Information Overload in a Network of Targeted Communication,” RAND Journal of Economics, 35 (3), 542–60.

61.

Wei

Yiluo

Tyson

Gareth

(2024), “Understanding the Impact of AI-Generated Content on Social Media: The Pixiv Case,” in Proceedings of the 32nd ACM International Conference on Multimedia. ACM, 6813–22.

62.

Weitzman

Martin L.

(1979), “Optimal Search for the Best Alternative,” Econometrica, 47 (3), 641–54.

63.

Wilson

Jared

(2023), “AI’s Real Problem Is That It’s Boring,” The Hill (August 10), https://thehill.com/opinion/technology/4146423-ais-real-problem-is-that-its-boring/.

64.

Yue

(2024), “Creation, Consumption, and Control of Sensitive Content,” Marketing Science, 43 (4), 885–902.

65.

Yahoo and Publicis Media (2024), “Future of AI and Advertising: Trust Through Transparency,” https://assets-global.website-files.com/64232e2dde875cd0589eaa88/65de3d9a600aff9ea15f9ce7_%5BResearch%20Study%5D%20Future%20of%20AI%20%26%20Advertising_Yahoo%20%26%20Publicis%20OFFICIAL.pdf.

66.

Yildirim

Pinar

Gal-Or

Esther

Geylani

Tansev

(2013), “User-Generated Content and Bias in News Media,” Management Science, 59 (12), 2655–66.

67.

YouTube (2024), “New Disclosures and Labels for Generative AI Content on YouTube,” YouTube Help Community (March 18), https://support.google.com/youtube/thread/264550152/newdisclosures-and-labels-for-generative-ai-content-on-youtube.

68.

Zhang

Kaifu

Sarvary

Miklos

(2015), “Differentiation with User-Generated Content,” Management Science, 61 (4), 898–914.

69.

Zhou

Zou

Tianxin

(2023), “Competing for Recommendations: The Strategic Impact of Personalized Product Recommendations in Online Marketplaces,” Marketing Science, 42 (2), 360–76.

70.

Zhou

Eric

Lee

Dokyun

(2024), “Generative Artificial Intelligence, Human Creativity, and Art,” PNAS Nexus, 3 (3), 10.1093/pnasnexus/pgae052.

71.

Zou

Tianxin

Yue

Sarvary

Miklos

(2025), “Quality-Variety Tradeoffs in Recommendation Systems on Content Platforms,” Management Science (published online November 19), https://doi.org/10.1287/mnsc.2025.00300.

72.

Zou

Tianxin

Zhou

(2025), “Self-Preferencing and Search Neutrality in Online Retail Platforms,” Management Science, 71 (5), 4087–4107.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.46 MB

0.00 MB