SSRN Id4749221 Code3223260

Interlopers or Catalysts?
Dissecting the Impact of Incorporating

AI Players on Multiplayer Online Games
Zhechao Yang
Warrington College of Business,
University of Florida, Gainesville, Florida 32611
Email: [email protected]
Qili Wang
Liangfei Qiu
Electronic copy available at: https://ssrn.com/abstract=4749221

Interlopers or Catalysts? Dissecting the Impact of Incorporating AI
Players on Multiplayer Online Games
Abstract
Despite the rapid evolution and extensive application of artificial intelligence (AI), the specific dynamics
of human-AI interaction within the context of multiplayer online games remain underexplored. Our study
fills this research gap by utilizing a popular multiplayer online game’s policy that incorporates AI players
as opponents to human players into matches. We investigate the impact of introducing AI players on human
player engagement, friend team-ups, and the game ecosystem. Our findings indicate a significantly positive
impact of AI players on game engagement in terms of play frequency and duration. Our results suggest that
defeating AI players, who are almost indistinguishable from humans yet less competent, brings enjoyment
to human players, thus fostering self-efficacy. Such enhanced self-efficacy motivates human players to
engage more in the game. Furthermore, we highlight the social impact of incorporating AI players in
encouraging friend team-ups, which may be attributed to the increased confidence in making positive
contributions to the team’s performance. Additionally, our analysis suggests that these benefits of AI
players are moderated by player skill levels and teammate preferences. In particular, the positive effects of
AI players on both player engagement and friend team-ups are stronger among novices. Players who
initially prefer random teammates experience greater social benefits from incorporating AI players, as this
leads to more team-ups with friends for them. Finally, we provide a broader picture of the interplay between
AI players and multiplayer online game ecosystems and highlight the positive roles of AI players as the
base of the player pyramid in facilitating skill development, improving teamwork, and boosting player
retention. Our research offers valuable implications for game developers who seek to enhance player
engagement and foster social interaction in the rapidly evolving landscape of multiplayer online games.
Keywords: human-AI interaction, AI players, social connection, multiplayer online games, game
ecosystem, self-efficacy

1. Introduction
Technological advancements have significantly bolstered the integration of Artificial Intelligence (AI) into
the gaming industry (Uludağlı and Oğuz 2023; Sharma 2024). The Allied Market Research report notes
that the global AI in the game industry witnessed a substantial financial output of $1.08 billion in 2022 and
is expected to reach an impressive $11.42 billion by 2032, with a compound annual growth rate of 26.8%
(PR Newswire 2023). The incorporation of AI in games has marked a significant evolution in how games
are developed and designed, primarily through integrating intelligent non-player characters (NPCs). These
advanced AI-driven NPCs exhibit enhanced realistic behaviors and decision-making abilities, enriching the
gaming environment (Belle et al. 2022). As AI becomes more embedded in gaming, it fundamentally alters
gaming experiences. Therefore, it is essential to understand the impact of integrated AI on player
engagement and the broader game ecosystem.
1.1 Motivations
Multiplayer online games are a prominent category that leverages AI to enrich gaming experiences. In
multiplayer online games, players from diverse locations can interact, collaborate, and compete in real-time
within virtual worlds (Guo et al. 2019; Chan et al. 2023; Mai and Hu 2023). These games, such as League
of Legends, Dota 2, and PlayerUnknown’s Battlegrounds (PUBG), are renowned for their strategic
complexity and the advanced skills required for proficiency (Wang et al. 2023). A critical challenge these
games face is to attract and retain new players (Yamashita et al. 2017; Flaman 2023). Due to the substantial
skill disparity among players, new players often feel overwhelmed by the complexity of the game and the
superior skills of opponents, generating a frustrating experience and consequently leading to early
disengagement.
To address this issue, game developers have implemented several strategies to make the game more
accessible to beginners. For instance, League of Legends and Dota 2 adopt matchmaking systems to pair
players with comparable skill levels based on the evaluation of a player’s performance over time (Chen et
al. 2021). Game developers believe that matchmaking systems can create a more balanced competitive

environment and enhance the enjoyment of players. Despite these efforts, challenges remain in consistently
achieving balanced matches, and beginners still find themselves at a disadvantage when encountering more
experienced opponents. As an innovative move, PUBG introduces AI players to deal with the skill gap issue
(Hossam 2020). These AI players serve as opponents of human players rather than teammates. They can
imitate human behaviors and are hard to distinguish from human players. Nonetheless, their competence is
intentionally kept at a lower level, creating a more manageable and beginner-friendly gaming environment.
Hence, novices might gradually familiarize themselves with the game and develop their skills in a less
competitive environment, thus gaining confidence. Although introducing AI players seeks to improve the
gaming experiences of novices, its impact on the overall player base remains uncertain. Currently, there is
limited empirical evidence available on the outcomes of introducing AI players into multiplayer online
games. Hence, this study aims to fill this research gap.
1.2 Research Questions and Contributions
Prior literature has shown how to implement AI bots as background or supportive characters in games and
highlights the roles of AI bots in establishing adaptive learning environments and facilitating personalized
gaming experiences (Millington 2019; Belle et al. 2022). Despite this extensive research, the specific area
of AI entities acting as opponents to human players in multiplayer online games has not been examined.
This distinctive form of interaction introduces a novel field of study and the impact of introducing such AI
players on human player engagement in multiplayer online games remains underexplored. On the one hand,
introducing AI players into games might potentially deter human player participation. Some players might
feel bored playing against less competent AI opponents (Liu et al. 2013; Hamari and Keronen 2017). Weibel
et al. (2008) find that individuals playing against a computer-controlled opponent experience a reduced
feeling of presence in terms of immersion and involvement, which in turn reduces the overall enjoyment of
the game. Such boredom and decreased enjoyment might result in a reluctance to participate in games.
On the other hand, however, introducing AI players into multiplayer online games can have a positive
impact on player engagement and participation. Since AI players are usually less competent and easier to
defeat, individuals are now able to experience more victories by defeating them and then gain confidence

and self-efficacy, according to the social cognitive theory (Bandura 1989; Compeau et al. 1999; Santhanam
et al. 2016). As a result, the thoughtful incorporation of AI players might potentially make the challenging
game more enjoyable and increase overall engagement among players. Accordingly, we put forward our
first research question: (1) What is the effect of the introduction of AI players on human players’ game
engagement?
Our research investigates this question with data from PUBG, a prominent multiplayer online game
with a vast player base. In May 2020, PUBG launched an innovative policy to incorporate AI players into
matches (Hossam 2020). This policy implementation serves as a natural experiment, allowing us to analyze
how the presence of AI players influences human player engagement within the game. As players began to
join in matches with AI players sequentially, we employ a staggered difference-in-differences (DID) model
to evaluate the impact of this policy. Combined with various robustness checks, our analysis demonstrates
that the introduction of AI players has a significantly positive impact on human player engagement in terms
of playing frequency and duration. One plausible explanation is as follows. Following the policy change,
human players experience increased success in defeating AI opponents who closely resemble humans but
demonstrate less competence. Overcoming these AI opponents can elevate the enjoyment and satisfaction
levels of gaming experiences for human players. Consequently, human players gain greater self-efficacy
and demonstrate increased enthusiasm for continued game participation. These findings underscore the
potential advantages of integrating AI players into multiplayer online games, offering valuable insights and
practical implications for developers aiming to enhance player engagement in such environments.
Our research then extends to the social aspects of gaming, particularly focusing on how introducing
AI players affects players’ tendencies to engage in games with friends. According to Hadji-Vasilev (2024),
78% of U.S. gamers acknowledge that online gaming plays a crucial role in making new friendships and
sustaining existing relationships. Incorporating AI players can potentially alter traditional social
interactions in games in two conflicting directions. On the one hand, the presence of less competent AI
opponents might decrease the excitement and social appeal of the game (Chen et al. 2024). In this sense,
human players might be less motivated to team up with friends. On the other hand, the introduction of AI

players can incentivize individuals to engage in more frequent gaming sessions with friends. Human players
may feel more confident in their ability to make positive contributions to the team’s performance and derive
enjoyment from the social aspect of gaming through collaboration (Zhang et al. 2017). This leads us to the
second research question: (2) Can the introduction of AI players encourage human players to engage in
games with friends more actively?
Our empirical analysis reveals that introducing AI players can foster more engagement in games with
friends. This outcome might be attributed to a notable reduction in the performance pressures that human
players often experience when teaming up with friends (Santhanam et al. 2016). The presence of AI players
mitigates anxiety about underperforming and boosts self-efficacy and confidence in their ability to
contribute meaningfully to the overall team’s performance (Zhang et al. 2017; Zhou et al. 2022). Therefore,
players might favor participating in social activities and teaming up with friends. While little prior literature
has examined the impact of AI on social aspects of gaming, our findings contribute to the literature by
demonstrating the potential positive social impact of AI players. We find that although traditionally treated
as non-social actors, AI entities can indirectly stimulate social behaviors among human players. Our study
also provides crucial insights for game designers who seek to enhance the social connections within games.
By addressing our first and second research questions, we suggest that incorporating AI players
improves the self-efficacy of human players, thus motivating them to participate and play with friends.
However, the initial implementation of AI players did not meet with universal satisfaction, especially
among some experienced players who expressed significant annoyance and frustration with the policy
(Chen et al. 2021). They argued that battles against AI opponents are less exciting and fulfilling than those
against human opponents. Given this negative attitude, we propose that a player’s skill level might moderate
the impact of AI players. Additionally, in multiplayer online games, players exhibit diverse teammate
preferences, with some favoring teaming up with friends and others enjoying random teaming up with
strangers. Incorporating AI players into the game might have a heterogeneous impact on these different
player segments. According to the above two player characteristics, we propose another research question:

(3) Does the introduction of AI players have different impacts across players with different skill levels and
teammate preferences?
Our moderation analysis indicates that these two player characteristics, i.e., skill levels and teammate
preferences, moderate the impact of AI players on human player engagement and friend team-ups. While
the incorporation of AI players positively affects both novices and masters in general, the benefits are
amplified for novice players. With the presence of AI players, novices can gain more self-efficacy and
confidence than masters who are already proficient enough to defeat opponents before the policy. The
moderating role of teammate preference, however, shows a nuanced picture. While it does not significantly
alter the impact on the overall game engagement, teammate preferences can moderate the social benefit of
AI players. Human players who typically prefer random teammates experience a more substantial positive
impact on friend team-ups after the policy, compared with those who already favor playing with friends.
This might be attributed to the reduced performance pressures and enhanced player self-efficacy brought
by AI players (Santhanam et al. 2016; Zhou et al. 2022). The results of our moderation analysis provide
indirect evidence supporting that AI players boost game engagement and friend team-ups by increasing
human players’ self-efficacy and confidence. These findings suggest that the benefits of incorporating AI
players into multiplayer online games are heterogeneous among players, providing insights into managing
players of diverse skill levels and teammate preferences.
Finally, we extend the above analyses and delve deeper into understanding the interplay between AI
players and the multiplayer online game ecosystem, which leads to our fourth research question: (4) How
does the introduction of AI players affect the multiplayer online game ecosystem? We examine this question
through three main aspects: the performance of human players, the interaction among team members, and
the churn rate within the game.
Our results indicate that the introduction of AI players revolutionizes the player pyramid, placing AI
players at the bottom layer, which was previously occupied by novices. Hence, new players can enjoy the
challenging game by eliminating AI players and also practice improving their skills in a less aggressive and
competitive environment. As new players gradually advance, they contribute to a more vibrant and

competitive game ecosystem, keeping it challenging and appealing for experienced players. Additionally,
the introduction of AI players encourages players to collaborate and work together, thus strengthening the
social connections within the game and contributing to a more cohesive community. Overall, incorporating
AI players into the multiplayer online game leads to a more inclusive and engaging gaming ecosystem, with
higher player satisfaction and retention.
To the best of our knowledge, our study offers a pioneering effort to explore the impact of introducing
AI players into the multiplayer online game on human player engagement and social connections.
Theoretically, our research enriches the broad literature on multiplayer online games (e.g., Hamari and
Keronen 2017; Zhao et al. 2022; Mai and Hu 2023) and human-AI interaction (e.g., Luo et al. 2019; Adam
et al. 2023; Wang et al. 2024a), offering new insights into how AI players as opponents reshape human
player behavior in the games. Our findings reveal that incorporating AI players as the base of the player
pyramid motivates human players to participate more in the game and strengthens their tendency to team
up with friends. Practically, our research provides important implications for game developers in improving
gaming experiences and engagement. In particular, our findings suggest that the introduction of AI players
can serve not only as a mechanism to lower difficulty levels and reduce the entry barrier for new players,
but also as a tool to boost social interactions and teamwork within games.
2. Literature Review
Our research is closely related to three streams of literature: (i) multiplayer online games, (ii) human-AI
interactions, and (iii) self-efficacy in the social cognitive theory. This section provides a brief overview of
these domains and highlights our contributions to existing studies.
2.1 Multiplayer Online Games
Existing studies have explored various aspects of multiplayer online games. Some studies examine business
models and profitability enhancement mechanisms of multiplayer online games. Guo et al. (2019)
investigate the effects of selling virtual currency on gamer behavior, the strategies of game designers, and
overall social welfare. Jiao et al. (2021) explore conditions under which game providers can benefit from

implementing transparent selling for virtual items. Meng et al. (2021) compare the outcomes of virtual
selling strategy with those of conventional selling approaches where players can only buy goods with real
currency. Mai and Hu (2023) analyze the free-to-play business model and optimize dynamic pricing and
advertising strategies to increase profitability. Wang et al. (2023) examine the freemium platform-based
business model in the context of multiplayer online games and investigate how specific product
differentiations influence players’ purchasing behavior.
Additional investigations focus on the rationale of players’ participation in these games. For example,
factors such as enjoyment, the appeal of fantasy, curiosity, the excitement of challenges, the pursuit of
achievement, and the attraction of rewards are crucial in spurring player involvement (Hamari and Keronen
2017; Zhao et al. 2022). Additionally, Zhang et al. (2017) uncover that social aspects, including social
connections, social interactions, and social comparison, can also serve as incentives for game participation.
Further research delves into strategies for enhancing player engagement. Huang et al. (2019) and Chen et
al. (2021) explore methods to foster player engagement through implementing matchmaking protocols.
Chen (2023) proposes methods to assess the latent skills of players and show how these insights can enable
developers to design strategies to improve engagement.
Our research contributes to this stream of literature by empirically investigating a novel approach to
improve engagement in multiplayer online games. Unlike previous studies focusing on direct incentives to
encourage player involvement, our research explores the potential outcomes of introducing AI players to
human player engagement. Furthermore, we complement prior research by examining the impact of AI
players on the social aspects of gaming.
2.2 Human-AI interactions
Our work adds to the studies of human-AI interactions. Existing research examines how and why
individuals interact differently with humans and AI entities such as chatbots and service robots (Wang et
al. 2024a). For instance, Adam et al. (2023) reveal that while customers might prefer initial interactions
with human sales agents over automated ones, they are reluctant to share personal information with them.
Mende et al. (2019) discover that humanoid service robots might cause greater consumer discomfort as they

can be perceived as a threat to human identity. The literature specifies possible psychological mechanisms
that explain the variations in these interactions (Castelo et al. 2023; He et al. 2024). For example,
competence and warmth are two major reasons (Luo et al. 2019). Specifically, humans are usually perceived
as more knowledgeable and friendly than bots. Additionally, Adam et al. (2023) show that expectations
also play a significant role. People generally hold higher competence expectations for humans than bots
before and during the interactions. As distinguishing between humans and chatbots becomes increasingly
challenging, recent research investigates the impact of disclosing chatbot identity on consumer purchase
behavior (Luo et al. 2019), service evaluations (Castelo et al. 2023), and the trust from consumers (Gnewuch
et al. 2024). Moreover, Han et al. (2023) explore the potential benefits of integrating AI with the feature of
expressing positive emotions, providing strategies for improving human-AI interactions.
Human-AI collaboration is another emerging research topic that boosts the integration of AI in
organizational settings. Dennis et al. (2023) explore how human team members perceive AI teammates and
whether they are open to collaborating with AI entities. The existing literature also provides various
methods to evaluate and improve the performance of human-AI teams. For instance, Lebovitz et al. (2021)
introduce approaches to assess AI performance in knowledge work. Ge et al. (2021) analyze the impact of
human intervention in adjusting robo-advisor recommendations on investment outcomes. Fügener et al.
(2022) discover that collaborative efforts between humans and AI can lead to superior performance
compared to humans or AI working independently, and they discuss what types of tasks can be allocated to
AI. Wang et al. (2024c) suggest that while experienced workers might better leverage AI, they show more
algorithm aversion, which highlights the complex interplay between humans and AI. Considering the
learning capability of AI, research also examines how human knowledge can augment AI (Fügener et al.
2021) and how AI influences organizational learning frameworks (Sturm et al. 2021).
AI-driven NPCs and their interaction with humans also attract attention from researchers. Uludağlı
and Oğuz (2023) summarize the methods used to improve the decision-making processes of NPCs based
on their current situations in video games. Burgess and Jones (2020) examine players’ emotional attachment
to NPCs. It is important to note that AI players in games like PUBG differ from traditional NPCs.
10

Traditional NPCs are often static and predictable, programmed for fixed roles within the narrative or
mechanics of games. They usually follow predefined paths or exhibit repetitive behaviors, contributing to
the game environment and storyline (Belle et al. 2022). This predictability makes it straightforward for
players to recognize traditional NPCs. In contrast, AI players in games like PUBG are designed to mimic
human-like behaviors, making them hard to distinguish from human players. The advanced intelligence of
AI players ensures that each encounter is potentially unique, offering human players a more dynamic and
unpredictable gaming experience (Hossam 2020).
Our research complements this body of literature by analyzing the impact of integrating AI players,
who actively compete against human players rather than serving as background characters or support roles,
on player engagement and the tendency to team up with friends.
2.3 Self-efficacy in Social Cognitive Theory
Social cognitive theory suggests that learning occurs within a social context, primarily through observation,
imitation, and modeling (Bandura 1989; Compeau et al. 1999; Santhanam et al. 2016). At the heart of this
theory is the concept of self-efficacy, which is the belief in one’s ability to organize and execute the steps
necessary to achieve specific objectives (Bandura 1997, p. 3). According to Bandura (1997), enactive
mastery experiences or successes from actual performances are the most effective means of strengthening
self-efficacy since they offer the most straightforward and authentic evidence of one’s skills. A high level
of self-efficacy affects how an individual faces challenges and overcomes obstacles, playing a crucial role
in determining engagement levels.
The concept of self-efficacy from social cognitive theory has been applied across various fields. In
the context of game-based learning, Santhanam et al. (2016) discover that trainees paired with less skilled
competitors and achieved victory exhibited higher self-efficacy and superior learning outcomes. In the
sphere of online learning, Huang et al. (2021) demonstrate that individuals with higher levels of self-
efficacy are more motivated to overcome procrastination and accelerate task completion. Within the realm
of weight management, Zhou et al. (2022) identify elements that enhance understanding of individuals'
health management efforts, highlighting self-efficacy as a key driver in regulating behaviors.
11

Our study enriches the current literature on self-efficacy in social cognitive theory by exploring its
application in the novel context of multiplayer online games. This investigation provides valuable insights,
showing that the introduction of AI players as the base of the player pyramid can boost player engagement
and the tendency to team up with friends by strengthening players’ self-efficacy.
3. Data and Descriptive Statistics
We obtain data from PUBG, a leading multiplayer online game (Dedmon 2018). In the game, one hundred
players parachute onto an island and search for weapons and equipment to eliminate others while striving
to survive (Kim and Tsvetkova 2022). Players can choose to enter a solo match, competing individually
against others, or join in a squad match in a team of four, leading to a total of 25 teams per match.
Additionally, in PUBG, players can connect with other participants and add them to the friend list, fostering
relationships based on shared in-game experiences. In other words, friends can team up and join in matches
together. For those who do not team up with friends, PUBG automatically assigns them to a group with
three other random players. Note that AI players do not appear as human teammates in these matches. In
the game, to avoid getting damaged every second, players must stay within the safe area, which is not
covered by the blue color as illustrated in Figure 1 (Higham 2017). As the game progresses, the available
safe area on the game map shrinks, compelling the remaining players into a smaller space to force
encounters. Victory is awarded to the last surviving team.
Launched in March 2017, PUBG quickly attracted a vast number of players but experienced a
significant decline during 2018 and 2019. This downturn was largely attributed to the substantial skill gap
among players. Joon H. Choi, the lead project manager of the PUBG console team, noted that while veterans
continued to improve their skills, new players often struggled and were eliminated early due to a lack of
proficiency (Hossam 2020). In response, PUBG introduced AI players into normal mode games (AI
matches) in May 2020, aiming to modify challenge levels and thereby enhance enjoyment and foster
engagement. These AI players mimic human actions, including entering the battlegrounds, looting, shooting,
and moving across the map. Note that these AI players are so intelligent that it is difficult to identify them
12

from human participants (Gnewuch et al. 2024). However, they are typically less challenging opponents,
exhibiting simpler tactical behaviors and decision-making processes than human players (Hossam 2020).
Alongside this change, PUBG kept its non-AI mode that only features human players for a pure player-
versus-player experience. This policy inherently categorizes players into two groups in our analysis: the
treatment group, which consists of players who only played AI matches subsequent to their initial AI match,
and the control group, which includes players who never joined an AI match.
Player 1
Player 2
Player 3
Player 4
Figure 1. The Safe Area and Blue Area in PUBG
We collect data from December 2, 2019, to August 3, 2020, spanning 24 weeks before and 12 weeks
after the focal event date (May 19, 2020). Following prior literature, we utilize a snowball sampling strategy
to ensure a representative sample (e.g., Goodman 1961; Song et al. 2019; Cao et al. 2024). Starting with
five randomly selected active seed players, we tracked all players from their matches within our observation
window to establish the initial pool. Subsequently, we randomly choose five players from the pool and
continue to trace the players from their matches until we obtain a pool with 10,000 players. We then select
active players who join in at least one match both before and after the focal event date. This approach results
in a final dataset with 3,798 players. Furthermore, we collect data on players teaming up with friends. Given
13

the extensive player base, the probability of being paired with the same stranger more than once is nearly
zero. Hence, it is reasonable to classify players who team up multiple times as friends. Thus, we record the
teammates of each player in all the matches from June 27, 2019, to January 18, 2021, covering 24 weeks
before and after the observation window. If player 𝑝 teams up with player 𝑞 (𝑝 ≠ 𝑞) more than once, we
consider players 𝑝 and 𝑞 are friends in the game. This approach allowed us to compile comprehensive data
on friend team-ups. Table 1 shows the definitions and descriptive statistics for the main variables in our
analysis.
Table 1. Definitions and Descriptive Statistics of Variables
Variable Definition Mean SD Min Max

Whether a player has ever joined in an AI match
𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟 0.163 0.369 0 1
until the week.
The number of games that a player joins in
𝐺𝑎𝑚𝑒𝑁𝑢𝑚 0.430 0.647 0 33
during a week.
The time (in minutes) that a player spends on
𝐺𝑎𝑚𝑒𝑇𝑖𝑚𝑒 5.265 9.869 0 590.330
the game during a week.
The number of games where a player teams up
𝐺𝑎𝑚𝑒𝑁𝑢𝑚𝐹𝑟𝑖𝑒𝑛𝑑 0.200 0.475 0 27
with her friends during a week.
The time (in minutes) that a player teams up
𝐺𝑎𝑚𝑒𝑇𝑖𝑚𝑒𝐹𝑟𝑖𝑒𝑛𝑑 2.569 7.322 0 419.841
with her friends during a week.
The proportion of game numbers that a player
𝐺𝑎𝑚𝑒𝑁𝑢𝑚𝐹𝑟𝑖𝑒𝑛𝑑𝑅𝑎𝑡𝑖𝑜 teams up with her friends during a week, which 0.167 0.367 0 1
is 𝐺𝑎𝑚𝑒𝑁𝑢𝑚𝐹𝑟𝑖𝑒𝑛𝑑/𝐺𝑎𝑚𝑒𝑁𝑢𝑚.
The proportion of game time that a player teams
𝐺𝑎𝑚𝑒𝑇𝑖𝑚𝑒𝐹𝑟𝑖𝑒𝑛𝑑𝑅𝑎𝑡𝑖𝑜 up with her friends during a week, which is 0.168 0.369 0 1
𝐺𝑎𝑚𝑒𝑇𝑖𝑚𝑒𝐹𝑟𝑖𝑒𝑛𝑑/𝐺𝑎𝑚𝑒𝑇𝑖𝑚𝑒.
4. Empirical Analysis
This section presents our empirical analysis and the corresponding results. Initially, we explore how the
introduction of AI players influences human players’ game engagement and their tendency to team up with
friends. Next, we conduct a series of tests to examine whether our results are robust. After that, we reveal
that player skill levels and teammate preferences can moderate the impact of introducing AI players. Finally,
we examine the overall impact of AI players on the multiplayer online game ecosystem, which provides a
broader picture of the interplay between AI players and human behavior in multiplayer online games.
14

4.1 Main Analysis
We first use a two-way fixed effects model to analyze the impact of introducing AI players on human player
engagement and their inclination to team up with friends. Since human players began to participate in AI
matches in different weeks after the policy, we employ a staggered difference-in-difference (DID) model
(Liang et al. 2024) and utilize the following regression specification to conduct the analysis:
log(𝑦=> ) = 𝛽B + 𝛽D 𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> + 𝛼= + 𝛿> + 𝜀=> , (1)
where the dependent variable, 𝑦=> , represents either player engagement or friend team-ups. To measure
human players’ game engagement, we use the number of games that player 𝑖 joins in during week 𝑡
(𝐺𝑎𝑚𝑒𝑁𝑢𝑚=> ) and the time (in minutes) that player 𝑖 spends on the game during week 𝑡 (𝐺𝑎𝑚𝑒𝑇𝑖𝑚𝑒=> ).
To measure human players’ friend team-ups, we use the number of games where player 𝑖 teams up with
friends during week 𝑡 ( 𝐺𝑎𝑚𝑒𝑁𝑢𝑚𝐹𝑟𝑖𝑒𝑛𝑑=> ) and the corresponding time spent (in minutes)
(𝐺𝑎𝑚𝑒𝑇𝑖𝑚𝑒𝐹𝑟𝑖𝑒𝑛𝑑=> ). The independent variable in the specification, 𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> , is a dummy variable,
indicating whether player 𝑖 has ever engaged in an AI match until week 𝑡. We incorporate player-level fixed
effects 𝛼= to control for individual characteristics among players and weekly fixed effects 𝛿> to capture
week-to-week fluctuations.
Table 2. Impact of AI Players on Human Player Engagement and Friend Team-ups
(1) (2) (3) (4)

Variables
log(𝐺𝑎𝑚𝑒𝑁𝑢𝑚) log(𝐺𝑎𝑚𝑒𝑇𝑖𝑚𝑒) log(𝐺𝑎𝑚𝑒𝑁𝑢𝑚𝐹𝑟𝑖𝑒𝑛𝑑) log(𝐺𝑎𝑚𝑒𝑇𝑖𝑚𝑒𝐹𝑟𝑖𝑒𝑛𝑑)
0.177*** 0.520*** 0.092*** 0.291***
𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=>
(0.005) (0.016) (0.004) (0.012)
Player FE Yes Yes Yes Yes
Weekly FE Yes Yes Yes Yes
Observations 136,728 136,728 136,728 136,728
Note: ***𝑝 < 0.01, **𝑝 < 0.05, *𝑝 < 0.1; Robust standard errors clustered by each player are in parentheses.
The estimation results are reported in Table 2. Our main interest lies in the coefficients of 𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> .
Across the results presented in the four columns, we find that the coefficients of 𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> are
significantly positive. Specifically, the introduction of AI players contributes to about a 17.7% increase in
the number of games that a player engages in and a 52.0% increase in the amount of time that a player
15

spends on the game during a week. Furthermore, after introducing AI players, human players generally play
9.2% more games and spend 29.1% more time with friends. Our results indicate that after the introduction
of AI players, human players show increased engagement in the game and are more likely to team up with
friends, addressing our first and second research questions.
Note that players might team up with friends more frequently simply because they play games more
often. To rule out this possibility and accurately capture the teaming inclination, we introduce two
additional dependent variables: the proportions of games played (𝐺𝑎𝑚𝑒𝑁𝑢𝑚𝐹𝑟𝑖𝑒𝑛𝑑𝑅𝑎𝑡𝑖𝑜=> ) and time
spent (𝐺𝑎𝑚𝑒𝑇𝑖𝑚𝑒𝐹𝑟𝑖𝑒𝑛𝑑𝑅𝑎𝑡𝑖𝑜=> ) with friends by player 𝑖 during week 𝑡. We then employ the following
specifications:
𝑦=> = 𝛽B + 𝛽D 𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> + 𝛼= + 𝛿> + 𝜀=> , (2)
The estimated results are shown in Table 3. Specifically, the proportions of both games played and
time spent with friends by a player rise 11.8%. All the results together suggest that the introduction of AI
players boosts human player engagement and encourages friend team-ups.
Table 3. Impact of AI Players on Friend Team-ups
(1) (2)
Variables
𝐺𝑎𝑚𝑒𝑁𝑢𝑚𝐹𝑟𝑖𝑒𝑛𝑑𝑅𝑎𝑡𝑖𝑜 𝐺𝑎𝑚𝑒𝑇𝑖𝑚𝑒𝐹𝑟𝑖𝑒𝑛𝑑𝑅𝑎𝑡𝑖𝑜
0.118*** 0.118***
(0.004) (0.004)
Player FE Yes Yes
Weekly FE Yes Yes
Observations 136,728 136,728
Note: *** 𝑝 < 0.01, ** 𝑝 < 0.05, * 𝑝 < 0.1; Robust standard errors
clustered by each player are in parentheses.
One plausible explanation for our findings is rooted in the concept of self-efficacy within the social
cognitive theory (Bandura 1989; Santhanam et al. 2016; Zhou et al. 2022). Following the policy, players
achieve more eliminations by successfully competing against AI players, who are hard to distinguish from
human participants but are generally less competent. This increase in eliminations improves players’
enjoyment and sense of mastery, thus elevating their self-efficacy. According to the social cognitive theory,
an enhanced sense of self-efficacy will motivate players to engage more in the game as they feel more
16

competent in their game abilities. Concurrently, it encourages players to participate in social activities, such
as teaming up with friends, since they believe that they can positively contribute to the team’s performance.
4.2 Robustness Checks
In the previous analysis, we reveal that the introduction of AI players can encourage human players to
participate in games and team up with friends. To validate the robustness of these results, we conduct a
series of empirical tests. A summary of these checks is presented in Table 4. In addition, the robustness
checks for two ratio outcomes, i.e., 𝐺𝑎𝑚𝑒𝑁𝑢𝑚𝐹𝑟𝑖𝑒𝑛𝑑𝑅𝑎𝑡𝑖𝑜 and 𝐺𝑎𝑚𝑒𝑇𝑖𝑚𝑒𝐹𝑟𝑖𝑒𝑛𝑑𝑅𝑎𝑡𝑖𝑜, are reported
in Online Appendix A.
Table 4. Overview of Robustness Checks

Empirical Methods Description Key Insights/Takeaways
Propensity Score We implement the PSM to establish a Our findings remain robust after we
Matching (PSM) more comparable control group and correct the selection bias related to
thereby address the potential selection observable covariates.
bias.
Relative Time Model We conduct the RTM to check whether the The results eliminate the possibility of pre-
(RTM) parallel trends assumption is held. treatment trends as a confounding factor.
Goodman-Bacon We employ the Goodman-Bacon Our findings remain robust after
Decomposition Decomposition to recalibrate our DID eliminating weights associated with
estimator and alleviate potential biases problematic comparisons in our DID
arising from variations in treatment timing. estimation.
Two-Way Fixed We perform two-way fixed effects weights Our results demonstrate robustness when
Effects Weights to examine whether our results are robust accounting for the treatment effect
to the heterogeneity in treatment effects. heterogeneity.
Poisson Quasi- We use Poisson QMLE to estimate the We show that the results remain robust
Maximum Likelihood treatment effects to mitigate potential after using alternative approaches to
Estimation (QMLE) biases from using log-like transformations. address the concern about problematic log
transformations with zero values.
4.2.1 DID Combined with PSM
While the DID method partially addresses the selection issue via the fixed effects setting, selection bias
might still arise from the systematic heterogeneity between the treatment and control groups. For instance,
players who prefer extensive exploration of the map are in constant motion. Such movement increases their
visibility to other players, making them easy targets. In addition, exploration inevitably entails crossing
areas with little cover, so that opponents can easily spot and eliminate them. Thus, these players might show
a greater tendency to participate in AI matches to reduce their elimination risk. To counteract this potential
17

bias, we complement the DID model with the propensity score matching (PSM) approach to examine the
impact of introducing AI players. The combination of DID and PSM facilitates the establishment of a more
comparable control group, thereby effectively mitigating bias from unobserved variables (Bavafa et al.
2018). In line with prior literature (e.g., Khurana et al. 2019; Fan et al. 2023), we first implement PSM to
construct a control group that is similar to the treated group in terms of observable characteristics and then
employ the DID regression in Equation (1).
When conducting matching, we first calculate the propensity score for each player with a logistic
regression model. In the model, we incorporate various characteristics recorded before the introduction of
AI players. These characteristics include the average weekly play time (𝐴𝑣𝑔𝑇𝑖𝑚𝑒), the average weekly
time spent playing with friends (𝐴𝑣𝑔𝑇𝑖𝑚𝑒𝐹𝑟𝑖𝑒𝑛𝑑), the average number of games played during weekends
(𝐴𝑣𝑔𝐺𝑎𝑚𝑒𝑊𝑒𝑒𝑘𝑒𝑛𝑑), and during weekdays (𝐴𝑣𝑔𝐺𝑎𝑚𝑒𝑊𝑒𝑒𝑘𝑑𝑎𝑦). Additionally, we employ the average
distance of walking in each match (𝐴𝑣𝑔𝑊𝑎𝑙𝑘𝐷𝑖𝑠𝑡), the average distance traveled by vehicles in each match
(𝐴𝑣𝑔𝑅𝑖𝑑𝑒𝐷𝑖𝑠𝑡), and the average number of different weapon types used to eliminate opponents in a match
(𝐴𝑣𝑔𝑊𝑒𝑎𝑝𝑜𝑛𝑁𝑢𝑚).
Table 5. Statistics of Variables Before and After PSM

Treatment Group Control Group
Before Matching After Matching
Variables Mean Mean Difference Mean Difference
***
𝐴𝑣𝑔𝑇𝑖𝑚𝑒 3.127 2.340 0.787 2.882 0.245
𝐴𝑣𝑔𝑇𝑖𝑚𝑒𝐹𝑟𝑖𝑒𝑛𝑑 1.289 0.997 0.292*** 1.214 0.075
𝐴𝑣𝑔𝐺𝑎𝑚𝑒𝑊𝑒𝑒𝑘𝑒𝑛𝑑 0.085 0.072 0.013** 0.076 0.009
𝐴𝑣𝑔𝐺𝑎𝑚𝑒𝑊𝑒𝑒𝑘𝑑𝑎𝑦 0.178 0.138 0.04*** 0.172 0.006
𝐴𝑣𝑔𝑊𝑎𝑙𝑘𝐷𝑖𝑠𝑡 324.937 258.441 66.496*** 309.817 15.120
𝐴𝑣𝑔𝑅𝑖𝑑𝑒𝐷𝑖𝑠𝑡 250.315 234.582 15.733 278.434 -28.119
𝐴𝑣𝑔𝑊𝑒𝑎𝑝𝑜𝑛𝑁𝑢𝑚 0.035 0.035 0 0.044 -0.009
Note: ***𝑝 < 0.01, **𝑝 < 0.05, *𝑝 < 0.1.
Subsequently, we apply the nearest-neighbor matching algorithm and pair each treated player with a
control player who has the closest propensity score. There are 2,297 players within the treatment group who
only engaged in AI matches after their initial AI match. We match these treated players with control players
with replacement to generate a new matched data sample. To examine the matching quality, we conduct t-
18

tests to compare the variable means for the treatment and control group both before and after matching. As
shown in Table 5, the differences between the groups are statistically insignificant for all characteristics
after matching. This indicates that the selection bias associated with observable characteristics has been
effectively addressed, and the matched dataset is balanced.
Raw Treated Matched Treated

0.4
0.4
Proportion
Proportion
0.2
0.2
0.0
0.0
0.5 0.6 0.7 0.8 0.5 0.6 0.7 0.8
Propensity Score Propensity Score
Raw Control Matched Control

0.4
0.4
Proportion
Proportion
0.2
0.2
0.0
0.0
0.5 0.6 0.7 0.8 0.5 0.6 0.7 0.8
Propensity Score Propensity Score
Figure 2. Propensity Scores Distributions Before and After Matching
Figure 2 depicts the distribution of propensity scores for both the treatment and control groups pre-
and post-matching. Remarkably, the distributions after PSM exhibit high similarity, indicating that the two
groups are well-matched and comparable.
Table 6. Impact of AI Players on Human Player Engagement and Friend Team-ups after PSM
(1) (2) (3) (4)
Variables
log (𝐺𝑎𝑚𝑒𝑁𝑢𝑚) log(𝐺𝑎𝑚𝑒𝑇𝑖𝑚𝑒) log(𝐺𝑎𝑚𝑒𝑁𝑢𝑚𝐹𝑟𝑖𝑒𝑛𝑑) log(𝐺𝑎𝑚𝑒𝑇𝑖𝑚𝑒𝐹𝑟𝑖𝑒𝑛𝑑)
0.205*** 0.616*** 0.107*** 0.342***
(0.005) (0.018) (0.004) (0.014)
Observations 117,432 117,432 117,432 117,432
19

We re-estimate our main regression model with the constructed matched sample. The estimated
results are presented in Table 6. The coefficients of the 𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟 are significantly positive, which is
consistent with the main results obtained from the full sample. Hence, by establishing a more comparable
control group, our analysis demonstrates the positive impact of introducing AI players on human players’
engagement and their tendency to team up with friends.
4.2.2 Relative Time Model
A fundamental assumption of the DID model is that players in the treatment and control groups show
parallel trends in game engagement and the inclination to team up with friends before introducing AI players.
We perform the relative time model (RTM) approach to verify the parallel trends assumption by estimating
the treatment effect within eight weeks before and after introducing AI players (Autor 2003; Chan et al.
2024; Wang et al. 2024b). Specifically, we modify Equation (1) and develop Equation (3).
log(𝑦=> ) = 𝛽B + U 𝛽V 𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=V + 𝛼= + 𝛿> + 𝜀=> , (3)

V
where 𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=V , for 𝜏 ∈ {−8, −7, … − 2,0,1, … ,7,8}, is a dummy variable denoting whether week 𝑡 is
the 𝜏-th week before (for negative 𝜏), during (for 𝜏 = 0), or after (for positive 𝜏) the initial engagement in
an AI match for player 𝑖. The time period just before the treatment (𝜏 = −1) serves as the benchmark for
comparison.
Table 7 reports the estimates of the lead and lagged coefficients for all outcome measurements. The
coefficients of 𝛽à to 𝛽`b capture the weekly variations before the treatment. We find that those
coefficients do not statistically differ from zero and thus verify the parallel trends assumption. Therefore,
we can eliminate the possibility of the pre-treatment trends as a confounding factor influencing the players’
engagement and their tendency to team up with friends.
Table 7. Impact of AI Players on Human Player Engagement and Friend Team-ups over Time
(1) (2) (3) (4)

Variables
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=à 0.000(0.010) 0.047(0.034) -0.008(0.008) -0.011(0.027)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,`c -0.001(0.010) 0.023(0.033) -0.002(0.008) -0.005(0.027)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,`d 0.005(0.010) 0.046(0.033) -0.006(0.008) -0.014(0.026)
20

𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,è 0.007(0.010) 0.041(0.032) 0.001(0.007) 0.012(0.026)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,`f 0.006(0.009) 0.037(0.032) 0.008(0.008) 0.023(0.026)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,`g -0.009(0.009) -0.005(0.030) -0.001(0.007) 0.010(0.025)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,`b -0.012(0.009) -0.041(0.029) -0.010(0.007) -0.039(0.024)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,B ***
0.581 (0.007) ***
1.866 (0.026) ***
0.282 (0.009) 0.930***(0.033)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,D 0.165***(0.010) 0.521***(0.035) 0.082***(0.008) 0.261***(0.029)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,b 0.129***(0.010) 0.374***(0.035) 0.063***(0.008) 0.188***(0.028)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,g ***
0.144 (0.010) ***
0.435 (0.035) ***
0.067 (0.008) 0.218***(0.029)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,f 0.152***(0.011) 0.442***(0.036) 0.071***(0.008) 0.219***(0.029)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,e 0.158***(0.011) 0.490***(0.037) 0.078***(0.009) 0.250***(0.030)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,d ***
0.149 (0.011) ***
0.474 (0.038) ***
0.080 (0.009) 0.271***(0.031)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,c 0.162***(0.011) 0.517***(0.039) 0.091***(0.009) 0.292***(0.033)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,a 0.142***(0.011) 0.444***(0.039) 0.069***(0.009) 0.240***(0.033)
Observations 136,728 136,728 136,728 136,728
4.2.3 Goodman-Bacon Decomposition
As highlighted in Goodman-Bacon (2021), DID estimations can be biased since players began to engage in
AI matches at different times after the policy (i.e., in a staggered adoption context). Essentially, the general
DID estimator is a weighted average of each possible two-group/two-period DID estimator, where the two
groups serve as the treatment and control groups, and the two periods are the time periods before and after
the treatment. Specifically, the estimator combines three pairs of comparisons: (1) the earlier treated as the
treatment while the later treated as the control (players who began to engage in an AI match earlier versus
those later), (2) the later treated as the treatment while the earlier treated as the control (players who began
to engage in an AI match later versus those earlier), and (3) all the treated as the treatment while the never
treated as the control (players who ever participated in an AI match versus those never). Goodman-Bacon
(2021) argues that the potential bias arises from the second comparison. To alleviate this concern, we
decompose the DID estimators (Alfaro et al. 2022; Eftekhari et al. 2023) and employ an “unbiased DID,”
which is the weighted average of the first and third comparisons.
Table 8 presents the decomposition results with separate estimations and the corresponding weights.
The general DID estimator, labeled “Overall DID,” incorporates all three comparisons, while “Unbiased
21

DID” includes only the first and third comparisons. We find that the results of unbiased DID align closely
in direction and size with the general DID estimator, indicating that the temporal variations in the impact
of introducing AI players do not significantly bias our estimation.
Table 8. Goodman-Bacon Decomposition

Comparison Weight Estimate Estimate Estimate Estimate
Earlier treated
0.170 0.398 1.291 0.193 0.649
vs. Later control
Later treated vs.
0.043 0.503 1.623 0.250 0.827
Earlier control
Treated vs.
0.787 0.111 0.294 0.061 0.184
Never treated
Overall DID 0.177 0.520 0.092 0.291
Unbiased DID 0.162 0.471 0.085 0.267
4.2.4 Two-Way Fixed Effects Weights
To further address the potential bias in the staggered DID estimation and ensure the robustness of our
findings, we employ a heterogeneous treatment estimation introduced by De Chaisemartin and
d’Haultfoeuille (2020). In our research context, players began to engage in AI matches at various times,
leading to temporal treatment heterogeneity. Our analysis focuses on the expected coefficients (𝐸[𝛽jkl ]),
where 𝛽jkl represents the treatment effect in group 𝑔 at time 𝑡 (De Chaisemartin and d’Haultfoeuille 2023).
With the parallel trends assumption, 𝐸[𝛽jkl ] is the weighted average of treatment effects among all (𝑔, 𝑡)
cells:
𝐸[𝛽jkl ] = 𝐸 n U 𝑊o,> ∆o,> v,

(o,>):rs,t uD
where ∆o,> denotes the average treatment effect for player 𝑔 in week 𝑡 . 𝐷o,> serves as the treatment
indicator, which is a dummy variable indicating if player 𝑔 began to engage in AI matches in week 𝑡, and
𝑊o,> specifies the weight for each (𝑔, 𝑡) cell, with the summation of one.
De Chaisemartin and d’Haultfoeuille (2023) highlight that 𝑊o,> can be negative, which is a problem
when the treatment effects vary across groups or time periods. In particular, 𝐸[𝛽jkl ] can be positive even if
22

the treatment effects for each player in every time period are consistently negative, which is known as the
“no-sign reversal property.” To mitigate this issue, we utilize the software package developed by De
Chaisemartin and d’Haultfoeuille (2020) and reveal a zero proportion of negative weights. Therefore, we
effectively eliminate the possibility that the variations in treatment effects among players might distort our
coefficient estimation.
4.2.5 Poisson Quasi-Maximum Likelihood Estimation
As indicated by Chen and Roth (2023), log transformations, commonly used in research, encounter
challenges when the outcome of interest (𝑌) can be zero. Since log(𝑌) is undefined at zero, researchers
often employ alternatives such as log(𝑌 + 1) . A significant concern related to these log-like
transformations is that the estimated average treatment effect (ATE) hinges on the units of 𝑌, making
interpretations as percentage effects unreliable, especially when the treatment shifts the outcome from zero
to a non-zero value. Consistent with Chen and Roth (2023), we apply the Poisson quasi-maximum
likelihood estimation (QMLE) to address this issue. Specifically, the number of games that a player joins
in during a week (𝐺𝑎𝑚𝑒𝑁𝑢𝑚=> ) and the number of games where a player teams up with her friends during
a week (𝐺𝑎𝑚𝑒𝑁𝑢𝑚𝐹𝑟𝑖𝑒𝑛𝑑=> ) are nonnegative integers, and thus can serve as dependable variables in the
Poisson QMLE specified in Equation (4).
𝑦=> = exp(𝛽B + 𝛽D 𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> + 𝛼= + 𝛿> + 𝜀=> ), (4)
The results of Poisson QMLE are reported in Table 9. To derive the estimated proportional treatment
effect, we calculate exp(𝛽D ) − 1, which are 0.828 and 0.868, respectively. This indicates that introducing
AI players increases the frequency of participating in the game and teaming up with friends by 82.8% and
86.8%, respectively. Hence, the findings from the Poisson QMLE align with our main results, underscoring
the significant impact of AI players in enhancing player engagement and friend team-ups.
Table 9. Impact of AI Players Using Poisson QMLE
(1) (2)
Variables
𝐺𝑎𝑚𝑒𝑁𝑢𝑚 𝐺𝑎𝑚𝑒𝑁𝑢𝑚𝐹𝑟𝑖𝑒𝑛𝑑
0.603*** 0.625***
(0.020) (0.030)
23

Player FE Yes Yes
Weekly FE Yes Yes
Note: 𝑝 < 0.01, 𝑝 < 0.05, 𝑝 < 0.1; Robust standard
*** ** *
errors clustered by each player are in parentheses.
In addition to the Poisson QMLE, we use 𝑌 as dependent variables (instead of the logarithmic value
of 𝑌) and estimate the specification in Equation (5) to alleviate potential biases from using log-like
transformations. The estimations are shown in Table 10. The significantly positive coefficients of
𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> indicate that our results remain robust and consistent after addressing concerns about
problematic log transformations with zero values.
𝑦=> = 𝛽B + 𝛽D 𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> + 𝛼= + 𝛿> + 𝜀=> , (5)
Table 10. Impact of AI Players without Log Transformations
(1) (2) (3) (4)

Variables
𝐺𝑎𝑚𝑒𝑁𝑢𝑚 𝐺𝑎𝑚𝑒𝑇𝑖𝑚𝑒 𝐺𝑎𝑚𝑒𝑁𝑢𝑚𝐹𝑟𝑖𝑒𝑛𝑑 𝐺𝑎𝑚𝑒𝑇𝑖𝑚𝑒𝐹𝑟𝑖𝑒𝑛𝑑
0.270*** 2.752*** 0.137*** 1.548***
(0.008) (0.131) (0.006) (0.095)
Observations 136,728 136,728 136,728 136,728
4.3 Moderation Analysis
In this section, we delve into the moderating factors that influence the impact of introducing AI players on
human player engagement and friend team-ups. Specifically, we consider the moderating effects of player
skill levels and teammate preferences on the impact of AI players. In addition, the results of the moderation
analysis using two ratio outcomes, i.e., 𝐺𝑎𝑚𝑒𝑁𝑢𝑚𝐹𝑟𝑖𝑒𝑛𝑑𝑅𝑎𝑡𝑖𝑜 and 𝐺𝑎𝑚𝑒𝑇𝑖𝑚𝑒𝐹𝑟𝑖𝑒𝑛𝑑𝑅𝑎𝑡𝑖𝑜, are shown
in Online Appendix A.
4.3.1 Player Skill Levels
In our main analysis, we show that introducing AI players can encourage human players to engage in the
game and team up with friends. We then investigate how these effects of AI players differ based on player
skill levels. To classify players according to skill level, we calculate the average number of opponents a
24

player eliminates per match before the policy. Players who eliminate opponents fewer than the median
among all players are labeled as novices, while those who eliminate more are categorized as masters. These
classified skill levels are then used as a moderating variable in our regression model estimations:
log(𝑦=> ) = 𝛽B + 𝛽D 𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> + 𝛽b 𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> × 𝑁𝑜𝑣𝑖𝑐𝑒= + 𝛼= + 𝛿> + 𝜀=> , (6)
where 𝑁𝑜𝑣𝑖𝑐𝑒= is a dummy variable. 𝑁𝑜𝑣𝑖𝑐𝑒= equals 1 if player 𝑖 is a novice and 0 otherwise. Table 11
presents the results of our analysis. The positive coefficients of 𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> indicates that introducing AI
players improves human player engagement and friend team-ups when 𝑁𝑜𝑣𝑖𝑐𝑒= = 0. In addition, the
coefficients of the interaction term between variables 𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> and 𝑁𝑜𝑣𝑖𝑐𝑒= are significantly positive,
which underscores the amplified beneficial effects of AI players on novice players. A possible explanation
is as follows. Prior to the introduction of AI players, novices, unlike masters, often struggled with opponent
elimination and exited early from the game. The presence of AI players, not identifiable from humans yet
less competent, mitigates this challenge and enhances the probability of eliminating an opponent.
Successfully competing against an AI player brings more enjoyment and satisfaction to novices than
masters. Thus, novices gain more self-efficacy, leading to more game engagement and friend team-ups
(Zhang et al. 2017; Zhou et al. 2022).
Table 11. Impact of AI Players: Moderating Role of Skill Levels
(1) (2) (3) (4)

Variables
0.160*** 0.444*** 0.082*** 0.249***
(0.006) (0.020) (0.005) (0.016)
𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> 0.032*** 0.144*** 0.019*** 0.079***
× 𝑁𝑜𝑣𝑖𝑐𝑒= (0.007) (0.023) (0.006) (0.019)
Observations 136,728 136,728 136,728 136,728
4.3.2 Player Teammate Preferences
We then explore the moderating role of player teammate preferences, specifically whether a player opts to
team up with random strangers or friends. To categorize players based on their teammate preferences, we
25

first obtain the proportion of matches a player participates in with random teammates before implementing
the policy. If a player’s ratio of matches with random teammates exceeds the median of all players, this
player is marked as preferring random teammates. A player with a lower ratio is considered to favor playing
with friends. We then proceed to assess the moderating effect of teammate preference using the following
regression model:
log(𝑦=> ) = 𝛽B + 𝛽D 𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> + 𝛽b 𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> × 𝑃𝑟𝑒𝑓𝑒𝑟𝑅𝑎𝑛𝑑𝑇𝑒𝑎𝑚= + 𝛼= + 𝛿> + 𝜀=> , (7)
where 𝑃𝑟𝑒𝑓𝑒𝑟𝑅𝑎𝑛𝑑𝑇𝑒𝑎𝑚= is a dummy variable indicating whether player 𝑖 prefers random teammates
before introducing AI players. Table 12 presents the results. The significantly positive coefficients of
𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> suggest that, introducing AI players boosts player engagement and fosters friend team-ups
when𝑃𝑟𝑒𝑓𝑒𝑟𝑅𝑎𝑛𝑑𝑇𝑒𝑎𝑚= = 0. For the outcomes of friend team-ups, i.e., log(𝐺𝑎𝑚𝑒𝑁𝑢𝑚𝐹𝑟𝑖𝑒𝑛𝑑) and
log(𝐺𝑎𝑚𝑒𝑇𝑖𝑚𝑒𝐹𝑟𝑖𝑒𝑛𝑑), the coefficients of the interaction term are significantly positive, which suggests
that players initially inclined towards random teammates will engage in the game with friends more
frequently and for longer durations, compared to those who initially favor playing with friends. One
possible explanation is that the introduction of AI players reduces the performance pressures associated
with playing with friends. Players gain more self-efficacy and feel more capable in the game by defeating
AI players. As a result, they believe that they can positively contribute to the team’s performance and might
pursue more socially fulfilling experiences, which are often found in teaming up with friends, according to
the social cognitive theory (Santhanam et al. 2016; Zhou et al. 2022). This effect is more pronounced among
players who are hesitant to team up with friends due to performance anxiety and prefer random teammates
before policy. Our finding highlights a shift in player behavior towards more social engagement with friends
in the gaming environment after introducing AI players.
However, for the outcomes of individual engagement, i.e., log(𝐺𝑎𝑚𝑒𝑁𝑢𝑚) and log(𝐺𝑎𝑚𝑒𝑇𝑖𝑚𝑒),
the coefficients for the interaction term between 𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> and 𝑃𝑟𝑒𝑓𝑒𝑟𝑅𝑎𝑛𝑑𝑇𝑒𝑎𝑚= are statistically
insignificant, indicating that a player’s initial teammate preference will not significantly moderate the
impact of AI players on total game engagement. One possible explanation is that teammate preferences
26

might not affect the overall efforts that a player puts into the game and do not directly influence total
engagement.
The above moderation analyses tackle our third research question. The corresponding results provide
further evidence supporting the mechanism that AI players boost game engagement and friend team-ups by
increasing human players’ self-efficacy and confidence.
Table 12. Impact of AI Players: Moderating Role of Teammate Preferences
(1) (2) (3) (4)

Variables
0.174*** 0.505*** 0.076*** 0.244***
(0.006) (0.020) (0.005) (0.016)
𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> 0.006 0.033 0.035*** 0.100***
× 𝑃𝑟𝑒𝑓𝑒𝑟𝑅𝑎𝑛𝑑𝑇𝑒𝑎𝑚= (0.007) (0.023) (0.005) (0.019)
Observations 136,728 136,728 136,728 136,728
4.4 AI Players and Game Ecosystem
Extending previous analysis, we further explore the interplay between AI players and the multiplayer online
game ecosystem by investigating the additional impact of AI players, in terms of the performance of human
players, the interactions among team members, and the churn rate within the game.
4.4.1 Player Game Performance
First, our analysis examines the impact of introducing AI players on individual game performance. We
employ the metric of the average distance for eliminating an opponent to assess player proficiency for
several reasons. It is necessary to have acute situational awareness and familiarity with the map to
accurately detect long-range opponents (Biswas 2020). Effectively targeting and deciding whether to
confront distant opponents typically requires players to have a more refined sense of tactical positioning,
compared to that in close combat. Furthermore, successfully eliminating an opponent far away requires
advanced skills, such as the ability to aim precisely and a comprehensive understanding of the bullet drop
and travel time (Knight 2017). Importantly, we avoid using more direct measurements, such as the average
27

number of opponents eliminated per match, to rule out potential bias in evaluating player skill levels after
introducing AI players. On the one hand, the number of total opponents eliminated in a match, if including
AI opponents, cannot precisely measure the actual performance of a player since defeating AI players does
not require advanced skills. On the other hand, using the number of total human opponents eliminated in a
match can be biased. The presence of AI players reduces the number of human players in a match, thereby
decreasing the maximum number of human opponents that can be eliminated and potentially distorting the
true measure of performance. Nevertheless, we have explored these direct measurements as alternative
dependent variables and obtained substantially similar results.
The regression in Equation (1) is employed to estimate the results and 𝑦=> denotes the average
distance at which player 𝑖 eliminates a human opponent in week 𝑡 (𝐴𝑣𝑔𝐸𝑙𝑖𝑚𝑖𝑛𝑎𝑡𝑒𝐷𝑖𝑠𝑡𝑎𝑛𝑐𝑒=> ). The higher
this measurement, the greater the performance. The results in Column 1 of Table 13 uncover that the
introduction of AI players significantly boosts the performance of human players. Specifically, it leads to
an average increase of 28.1% in the distance from which a player can successfully eliminate an opponent.
As noted by social cognitive theory (Compeau et al. 1999), the introduction of AI players boosts the self-
efficacy of human players, subsequently encouraging them to participate more frequently and practice more
intensively. Hence, human players exhibit higher performance.
Table 13. Impact of AI Players on the Multiplayer Online Game Ecosystem
(1) (2) (3)

Variables
log(𝐴𝑣𝑔𝐸𝑙𝑖𝑚𝑖𝑛𝑎𝑡𝑒𝐷𝑖𝑠𝑡𝑎𝑛𝑐𝑒) log(𝐴𝑣𝑔𝑇𝑒𝑎𝑚𝑚𝑎𝑡𝑒𝐴𝑠𝑠𝑖𝑠𝑡) 𝐶ℎ𝑢𝑟𝑛𝑃𝑟𝑜𝑏
0.281*** 0.111*** -0.232***
(0.031) (0.012) (0.006)
Player FE Yes Yes Yes
Weekly FE Yes Yes Yes
Observations 26,538 42,323 136,728
Note: ***𝑝 < 0.01, **𝑝 < 0.05, *𝑝 < 0.1; Robust standard errors clustered by each player are
in parentheses.
4.4.2 Team Member Interactions
Second, we investigate how the introduction of AI players affects the interaction among team members. In
PUBG, players who are knocked down by opponents are not immediately eliminated from the game. Instead,
28

they are given a chance to be revived by teammates before succumbing to their injuries (Higham 2017).
For example, when a player is seriously injured due to enemy fire, they enter a “knocked down” state where
they cannot use weapons but can crawl slowly for cover or to facilitate rescue by a teammate. Teammates
are notified when a player is knocked down and can locate the downed player on their map. Reviving the
downed player requires a teammate to physically reach the player, stay nearby, and perform the revival
process for a few seconds without interruption. If the revival process is successfully completed before the
downed player’s health bar depletes fully, the downed player is restored to a minimal health state, allowing
them to heal and return to the game battle.
This aspect of the game provides us with a unique metric for assessing team interactions: the
frequency at which a player is knocked down but not fully eliminated. This metric reflects the frequency
with which a player receives help from her teammates. We estimate the model in Equation (1), where 𝑦=>
represents the average number of times that player 𝑖 is knocked down without being completely eliminated
during a specific week 𝑡 (𝐴𝑣𝑔𝑇𝑒𝑎𝑚𝑚𝑎𝑡𝑒𝐴𝑠𝑠𝑖𝑠𝑡=> ). A higher value of this measurement indicates more
team interactions and support. We report the results in Column 2 of Table 13. The significantly positive
coefficient of 𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> reveals that introducing AI players boosts team interactions. More concretely, we
show an 11.1% increase in the frequency of a player receiving a revival from teammates after the
introduction of AI players. This is because when players gain more self-efficacy and engage more in the
game with friends, their focus might shift from individual performance to collective strategy and teamwork,
according to the social cognitive theory (Compeau et al. 1999).
4.4.3 Player Retention
Third, we examine the potential of AI players to prevent players from leaving the game and retain them.
We introduce a dummy variable, 𝐶ℎ𝑢𝑟𝑛𝑃𝑟𝑜𝑏=> , to represent whether player 𝑖 discontinues her engagement
with the game in week 𝑡. 𝐶ℎ𝑢𝑟𝑛𝑃𝑟𝑜𝑏=> = 1 indicates no game activity for player 𝑖 in that week and
𝐶ℎ𝑢𝑟𝑛𝑃𝑟𝑜𝑏=> = 0 indicates active participation. To evaluate the impact, we apply the regression model in
Equation (2) and 𝑦=> corresponds to 𝐶ℎ𝑢𝑟𝑛𝑃𝑟𝑜𝑏=> . Our analysis, detailed in Column 3 of Table 13,
29

uncovers a significantly negative coefficient for 𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> , demonstrating that the introduction of AI
players leads to a 23.2% reduction in the likelihood of players churning. In other words, the presence of AI
players can effectively retain players. Consequently, introducing AI players not only boosts the engagement
of human players but also substantially increases their retention.
The above regression results respond to our fourth research question, implying that introducing AI
players into the multiplayer online game ecosystem can bring significant and substantial benefits.
Specifically, it can effectively address the obstacles encountered by novices while simultaneously
improving the engagement of advanced players. In the evolving world of multiplayer online games, a
prevalent and significant challenge is the widening skill gap between experienced and new players. This
situation often results in new players abandoning the game. Indeed, the player population of the multiplayer
online game ecosystem resembles a pyramid shown in Figure 3, where experienced players dominate the
apex while novices populate the base. Before the advent of AI players, these novices occupy the lowest
level and consistently suffer more defeats without obtaining satisfactory gaming experiences. Hence, they
eventually quit the game. Once novices exit, amateurs take their place at the bottom, encountering similar
unenjoyable experiences and also leaving the game. This cycle continues, progressively shrinking the player
base and damaging the game ecosystem.
However, the introduction of AI players revolutionizes this pyramid structure, positioning AI players
at the foundational layer and establishing a more enriching ecosystem. This strategic change enables new
players to derive fulfillment via victories against AI opponents and build self-efficacy, thus creating a
situation where they can improve skills through continuous practice with these AI opponents. New players
are able to learn at their own pace without the pressure of competing against advanced and experienced
human opponents. Furthermore, the introduction of AI players also benefits experienced players. The
enhanced skills and competence of novices improve the interest and engagement of more experienced
players since they now encounter an expanded pool of capable opponents, which ensures that the game is
competitive and challenging.
30

Master
Master
Advanced
Introduction of
Advanced
AI Players
Intermediate
Intermediate
Amateur
Amateur
Novice
Novice
AI Player
Less Enjoyment
Easy to Churn
Figure 3. Impact of AI Players on the Multiplayer Online Game Ecosystem
Additionally, introducing AI players transforms the social dynamics of the multiplayer online game
ecosystem. The presence of AI players improves social interaction and teamwork within the game,
contributing to a more collective and engaging gaming experience. Overall, the presence of AI players
significantly improves player retention and maintains a vibrant player base by enriching the gaming
experience for players across all skill levels and fostering social connections.
5. Discussion and Conclusions
The integration of AI into the gaming industry represents a rapidly expanding field with both practical and
theoretical importance. With the evolution of digital games, incorporating AI not only reshapes the gaming
experiences through dynamic and responsive interactions but also offers a fertile context to study human-
AI interplay within virtual settings. While prior research provides valuable implications for the human-AI
interaction (Ge et al. 2021; Fügener et al. 2022; Abdel-Karim et al. 2023), the impact of AI players, who
act as opponents instead of supportive entities, is not fully understood. Our research fills this gap and
concentrates on how introducing AI players in multiplayer online games influences player engagement and
social connections. Our results demonstrate the crucial impact of AI players and show that they can
encourage game engagement and friend team-ups. These benefits of AI players are moderated by player
skill levels and teammate preferences. In addition, we analyze the impact of AI players as the base of the
player pyramid on the overall game ecosystem. Consequently, our study contributes to the literature on
31

multiplayer online games and human-AI interactions and provides game designers with tangible guidance
for developing more engaging and immersive gaming experiences.
5.1 Theoretical and Managerial Implications
Firstly, we investigate the impact of introducing AI players into multiplayer online games on human player
engagement. Different from prior research that focuses on supportive service bots (Mende et al. 2019;
Castelo et al. 2023) and human-AI collaboration (Fügener et al. 2022; Dennis et al. 2023), our research
examines another specific kind of human-AI interaction, where AI entities serve as the opponents of human
players. The positive impact of AI players on game engagement broadens our understanding of human-AI
interaction. Specifically, we show that AI can play a role in boosting enjoyment and self-efficacy in
competitive settings, thus leading to more engagement, as noted in the social cognitive theory (Compeau et
al. 1999; Santhanam et al. 2016). From the practical perspective, game developers can encourage human
players to participate by thoughtfully designing AI players that are challenging but still manageable.
Secondly, we find that incorporating AI players can encourage friend team-ups. To the best of our
knowledge, this study is the first to empirically analyze the social impact of AI players and underscores the
crucial potential of AI to strengthen social connections. Although traditionally treated as non-social actors,
AI entities can indirectly stimulate social behaviors among human players. This finding offers game
developers and managers feasible approaches to enriching the social experience within games (Ai-Admin
2023). For instance, managers can design AI opponents that inherently require cooperative effort for victory
to encourage friend team-ups. This strategy can improve the game community by fostering social
connections and collaborative work.
Thirdly, our results uncover the heterogeneous impact of introducing AI players across players with
different skill levels and teammate preferences. In particular, the positive effects of AI players on both
player engagement and friend team-ups are stronger among novices. Moreover, the social benefit of more
friend team-ups is stronger for players who initially favor random teammates. These findings shed light on
the personalized nature of gaming experience and emphasize the intricate dynamics of human-AI
interaction (Belle et al. 2022). Therefore, game developers need to account for individual heterogeneity to
32

fully harness the benefits of AI players (Uludağlı and Oğuz 2023). For example, in matches involving a
substantial number of novices, integrating more AI players can reduce the difficulty levels and ensure that
these players find the game engaging and manageable, thus enhancing their overall experience.
Finally, our research provides the broader implications of incorporating AI players in multiplayer
online game ecosystems, contributing to the literature on multiplayer online games (e.g., Mai and Hu 2023;
Wang et al. 2023). We highlight the positive roles of AI players in altering the competitive environment,
facilitating skill development, improving teamwork, and boosting player retention. By positioning AI
players at the base of the player pyramid, this challenging and competitive game becomes more accessible
and welcoming to novices. As novices develop skills, master players can also enjoy the game since they
face more sophisticated competitors. Consequently, the presence of AI players can enrich the gaming
experiences across diverse player segments. These findings guide game developers on how to strategically
integrate AI players to maintain a vivid and vibrant game ecosystem (Sharma 2024). For example,
developers can create AI players as stepping stones that help new players adapt to the game mechanics,
progressively preparing them for more challenging human competition. Additionally, developers should
continuously monitor and refine their AI integration strategies based on player performance. As new players
become advanced, developers should adjust the difficulty levels of AI players to maintain game enjoyment
(Michail and Alepis 2023).
5.2 Limitations and Future Research Directions
We acknowledge several limitations of this research. One limitation is that we rely on data from a single
game, PUBG, which might not represent the various multiplayer online game landscape. PUBG, a battle
royale game with specific characteristics like survival elements and a shrinking play area, provides
gameplay distinct from other multiplayer online games related to role-playing or sports simulations. Thus,
the impact of AI players in PUBG might not be directly generalized to other game contexts with different
mechanics and player experiences. Our research is also limited by data constraints. Our dataset mainly
focuses on quantifiable aspects of player behavior, such as play frequency, duration, and achievement,
potentially neglecting the qualitative dimensions of gaming experiences. Crucial factors such as player
33

satisfaction and emotional reactions to AI players might not be adequately captured. Future research can
delve deeper into this topic as more data becomes available.
Our research can be extended in various directions. First, further studies can explore how the
introduction of AI players affects the effectiveness of current matchmaking systems and how to adapt the
matchmaking strategy accordingly. Matchmaking systems are designed to create balanced and enjoyable
matches by pairing players based on skill levels and experience. The presence of AI players introduces new
challenges and opportunities for optimizing these systems. Moreover, investigating the economic impact
of AI integration on the business models of the game industry, particularly concerning monetization
strategies and virtual item sales, represents a vital future research avenue. This exploration can provide
meaningful insights into how AI transforms traditional revenue models.
References
Abdel-Karim, B. M., Pfeuffer, N., Carl, K. V., & Hinz, O. (2023). How AI-based systems can induce
reflections: The case of AI-augmented diagnostic work. MIS Quarterly, 47(4), 1395-1424.
Adam, M., Roethke, K., & Benlian, A. (2023). Human vs. automated sales agents: How and why customer
responses shift across sales stages. Information Systems Research, 34(3), 1148-1168.
Ai-Admin. (2023). The future of artificial intelligence NPC - How AI is revolutionizing non-player
characters in gaming. https://aiforsocialgood.ca/blog/the-future-of-artificial-intelligence-npc-how-
ai-is-revolutionizing-non-player-characters-in-gaming, last accessed on Feb 10, 2024.
Alfaro, L., Faia, E., Lamersdorf, N., & Saidi, F. (2022). Health externalities and policy: The role of social
preferences. Management Science, 68(9), 6751-6761.
Autor, D. H. (2003). Outsourcing at will: The contribution of unjust dismissal doctrine to the growth of
employment outsourcing. Journal of Labor Economics, 21(1), 1-42.
Bandura, A. (1989). Human agency in social cognitive theory. American Psychologist, 44(9), 1175–1184.
Bandura, A. (1997). Self-Efficacy: The Exercise of Control W. H. Freemanand Company, New York.
Bavafa, H., Hitt, L. M., & Terwiesch, C. (2018). The impact of e-visits on visit frequencies and patient
health: Evidence from primary care. Management Science, 64(12), 5461-5480.
Belle, S., Gittens, C., & Graham, T. N. (2022). A framework for creating non-player characters that make
psychologically-driven decisions. In 2022 IEEE International Conference on Consumer Electronics
(ICCE) (pp. 1-7).
Biswas, S. D. (2020). GamingBytes: Five PUBG tips for better enemy spotting.
https://www.newsbytesapp.com/news/sports/pubg-mobile-tips-to-improve-enemy-spotting/story,
last accessed on Feb 10, 2024.
Burgess, J., & Jones, C. (2020). “I harbour strong feelings for Tali despite her being a fictional character:”
Investigating videogame players’ emotional attachments to non-player characters. Game
Studies, 20(1).
Cao, Z., Zhu, Y., Li, G., & Qiu, L. (2024). Consequences of information feed integration on user
engagement and contribution: A natural experiment in an online knowledge-sharing
community. Information Systems Research. Forthcoming.
34

Castelo, N., Boegershausen, J., Hildebrand, C., & Henkel, A. P. (2023). Understanding and improving
consumer reactions to service bots. Journal of Consumer Research, 50(4), 848–863.
Chan, T., Chen, Y., & Wu, C. (2023). Collaborate to compete: An empirical matching game under
incomplete information in rank-order tournaments. Marketing Science, 42(5), 1004-1026.
Chan, J., He, S., Qiao, D., & Whinston, A. (2024). Shedding light on the dark: The impact of legal
enforcement on darknet transactions. Information Systems Research. Forthcoming.
Chen, J. (2023). Learning and skill set formation: A structural examination of version upgrades, user
visibility, and AI strategies. Production and Operations Management, 32(12), 3856-3872.
Chen, J., He, S., & Yang, X. (2024). Platform loophole exploitation, recovery measures, and user
engagement: A quasi-natural experiment in online gaming. Information Systems Research.
Forthcoming.
Chen, M., Elmachtoub, A. N., & Lei, X. (2021). Matchmaking strategies for maximizing player engagement
in video games. Available at SSRN 3928966.
Chen, J., & Roth, J. (2023). Logs with zeros? Some problems and solutions. The Quarterly Journal of
Economics, qjad054.
Compeau, D., Higgins, C. A., & Huff, S. (1999). Social cognitive theory and individual reactions to
computing technology: A longitudinal study. MIS Quarterly, 23(2), 145-158.
De Chaisemartin, C., & d’Haultfoeuille, X. (2020). Two-way fixed effects estimators with heterogeneous
treatment effects. American Economic Review, 110(9), 2964-2996.
De Chaisemartin, C., & d’Haultfoeuille, X. (2023). Two-way fixed effects and differences-in-differences
with heterogeneous treatment effects: A survey. The Econometrics Journal, 26(3), C1-C30.
Dedmon, T. (2018). PUBG’s developer API is good news for both devs and players.
https://comicbook.com/gaming/news/pubg-developer-api-good-news-for-devs-and-players/, last
accessed on Feb 10, 2024.
Dennis, A. R., Lakhiwal, A., & Sachdeva, A. (2023). AI agents as team members: Effects on satisfaction,
conflict, trustworthiness, and willingness to work with. Journal of Management Information
Systems, 40(2), 307-337.
Eftekhari, S., Yaraghi, N., Gopal, R. D., & Ramesh, R. (2023). Impact of health information exchange
adoption on referral patterns. Management Science, 69(3), 1615-1638.
Fan, W., Zhou, Q., Qiu, L., & Kumar, S. (2023). Should doctors open online consultation services? An
empirical investigation of their impact on offline appointments. Information Systems Research, 34(2),
629-651.
Flaman, A. (2023). Examining the impact of skill-based matchmaking and the removal of skill gap in call
of duty: Modern Warfare 2 (2022). https://medium.com/@8flamz/examining-the-impact-of-skill-
based-matchmaking-and-the-removal-of-skill-gap-in-call-of-duty-88b1428d3d4b, last accessed on
Feb 10, 2024.
Fügener, A., Grahl, J., Gupta, A., & Ketter, W. (2021). Will humans-in-the-loop become borgs? Merits and
pitfalls of working with AI. MIS Quarterly, 45(3),1527-1556.
Fügener, A., Grahl, J., Gupta, A., & Ketter, W. (2022). Cognitive challenges in human–artificial
intelligence collaboration: Investigating the path toward productive delegation. Information Systems
Research, 33(2), 678-696.
Ge, R., Zheng, Z., Tian, X., & Liao, L. (2021). Human–robot interaction: When investors adjust the usage
of robo-advisors in peer-to-peer lending. Information Systems Research, 32(3), 774-785.
Gnewuch, U., Morana, S., Hinz, O., Kellner, R., & Maedche, A. (2024). More than a bot? The impact of
disclosing human involvement on customer interactions with hybrid service agents. Information
Systems Research. Forthcoming.
Goodman, L. A. (1961). Snowball sampling. The Annals of Mathematical Statistics, 32(1), 148-170.
Goodman-Bacon, A. (2021). Difference-in-differences with variation in treatment timing. Journal of
Econometrics, 225(2), 254-277.
Guo, H., Hao, L., Mukhopadhyay, T., & Sun, D. (2019). Selling virtual currency in digital games:
Implications for gameplay and social welfare. Information Systems Research, 30(2), 430-446.
35

Hadji-Vasilev, A. (2024). 23 video game and online gaming statistics, facts & trends for 2024.
https://www.cloudwards.net/online-gaming-statistics/, last accessed on Feb 10, 2024.
Hamari, J., & Keronen, L. (2017). Why do people play games? A meta-analysis. International Journal of
Information Management, 37(3), 125-141.
Han, E., Yin, D., & Zhang, H. (2023). Bots with feelings: Should AI agents express positive emotion in
customer service?. Information Systems Research, 34(3), 1296-1311.
He, Y., Xu, X., Huang, N., Hong, Y., & Liu, D. (2024). Enhancing user privacy through ephemeral sharing
design: Experimental evidence from the online dating context. Information Systems Research.
Forthcoming.
Higham, M. (2017). PlayerUnknown’s Battlegrounds review.
https://www.gamespot.com/reviews/playerunknowns-battlegrounds-review/1900-6416830/, last
accessed on Feb 10, 2024.
Hossam, M. (2020). PUBG Update 7.1 adds bots to public games to balance skill gap.
https://techraptor.net/gaming/news/pubg-update-71-adds-bots-to-lower-skill-gap-between-veterans-
and-newbies, last accessed on Feb 10, 2024.
Huang, N., Zhang, J., Burtch, G., Li, X., & Chen, P. (2021). Combating procrastination on massive online
open courses via optimal calls to action. Information Systems Research, 32(2), 301-317.
Huang, Y., Jasin, S., & Manchanda, P. (2019). “Level up”: Leveraging skill and engagement to maximize
player game-play in online video games. Information Systems Research, 30(3), 927-947.
Jiao, Y., Tang, C. S., & Wang, J. (2021). Selling virtual items in free-to-play games: Transparent selling
vs. opaque selling. Service Science, 13(2), 53-76.
Khurana, S., Qiu, L., & Kumar, S. (2019). When a doctor knows, it shows: An empirical analysis of doctors’
responses in a Q&A forum of an online healthcare portal. Information Systems Research, 30(3), 872-
891.
Knight, W. (2017). Zeroing distance guide and aiming tips. https://samurai-gamers.com/playerunknowns-
battlegrounds-pubg/zeroing-distance-aiming-tips/, last accessed on Feb 10, 2024.
Kim, J. E., & Tsvetkova, M. (2022). Cheating in online gaming spreads through observation and
victimization. Network Science, 9(4), 425-442.
Lebovitz, S., Levina, N., & Lifshitz-Assaf, H. (2021). Is AI ground truth really true? The dangers of training
and evaluating AI tools based on experts’ know-what. MIS Quarterly, 45(3), 1501–1526.
Liang, C., Wu, J., & Li, X. (2024). Do “likes” in a brand community always make you buy
more?. Information Systems Research. Forthcoming.
Liu, D., Li, X., & Santhanam, R. (2013). Digital games and beyond: What happens when players
compete?. MIS Quarterly, 37(1), 111-124.
Luo, X., Tong, S., Fang, Z., & Qu, Z. (2019). Frontiers: Machines vs. humans: The impact of artificial
intelligence chatbot disclosure on customer purchases. Marketing Science, 38(6), 937-947.
Mai, Y., & Hu, B. (2023). Optimizing free-to-play multiplayer games with premium
subscription. Management Science, 69(6), 3437-3456.
Mende, M., Scott, M. L., van Doorn, J., Grewal, D., & Shanks, I. (2019). Service robots rising: How
humanoid robots influence service experiences and elicit compensatory consumer responses. Journal
of Marketing Research, 56(4), 535-556.
Meng, Z., Hao, L., & Tan, Y. (2021). Freemium pricing in digital games with virtual currency. Information
Systems Research, 32(2), 481-496.
Michail, T., & Alepis, E. (2023). Design of real-time multiplayer word game for the Android platform using
firebase and fuzzy logic. In 2023 14th International Conference on Information, Intelligence,
Systems & Applications (IISA) (pp. 1-8).
Millington, I. (2019). AI for Games. CRC Press.
PR Newswire. (2023). AI in video games market to reach $11.4 billion, globally, by 2032 at 26.8%% CAGR:
Allied Market Research. https://finance.yahoo.com/news/ai-video-games-market-reach-
071000146.html, last accessed on Feb 10, 2024.
36

Santhanam, R., Liu, D., & Shen, W. C. M. (2016). Research note—Gamification of technology-mediated
training: Not all competitions are the same. Information Systems Research, 27(2), 453-465.
Sharma, N. (2024). This is how AI is empowering the gaming industry!
https://www.analyticsvidhya.com/blog/2023/12/ai-in-gaming-industry/, last accessed on Feb 10,
2024.
Song, T., Tang, Q., & Huang, J. (2019). Triadic closure, homophily, and reciprocation: An empirical
investigation of social ties between content providers. Information Systems Research, 30(3), 912-926.
Sturm T, Gerlach JP, Pumplun L, Mesbah N, Peters F, Tauchert C, Nan N & Buxmann, P. (2021).
Coordinating human and machine learning for effective organizational learning. MIS
Quarterly, 45(3), 1581–1602.
Uludağlı, M. Ç., & Oğuz, K. (2023). Non-player character decision-making in computer games. Artificial
Intelligence Review, 56, 14159–14191.
Wang, L., He, Y., Huang, N., Liu, D., Guo, X., & Chen, G. (2024a). The role of AI assistants in livestream
selling: Evidence from a randomized field experiment. Available at SSRN 4365103.
Wang, L., Lowry, P. B., Luo, X., & Li, H. (2023). Moving consumers from free to fee in platform-based
markets: An empirical study of multiplayer online battle arena games. Information Systems
Research, 34(1), 275-296.
Wang, Q., Qiu, L., & Xu, W. (2024b). Informal payments and doctor engagement in an online health
community: An empirical investigation using generalized synthetic control. Information Systems
Research. Forthcoming.
Wang, W., Gao, G., & Agarwal, R. (2024c). Friend or foe? Teaming between artificial intelligence and
workers with variation in experience. Management Science. Forthcoming.
Weibel, D., Wissmath, B., Habegger, S., Steiner, Y., & Groner, R. (2008). Playing online games against
computer-vs. human-controlled opponents: Effects on presence, flow, and enjoyment. Computers in
Human Behavior, 24(5), 2274-2291.
Yamashita, S., Matsuda, A., Hamanishi, N., Suwa, S., & Rekimoto, J. (2017). Demulti display: A
multiplayer gaming environment for mitigating the skills gap. In Proceedings of the Eleventh
International Conference on Tangible, Embedded, and Embodied Interaction (pp. 457-463).
Zhang, C., Phang, C. W., Wu, Q., & Luo, X. (2017). Nonlinear effects of social connections and interactions
on individual goal attainment and spending: Evidences from online gaming markets. Journal of
Marketing, 81(6), 132-155.
Zhao, Y., Yang, S., Shum, M., & Dutta, S. (2022). A dynamic model of player level-progression decisions
in online gaming. Management Science, 68(11), 8062-8082.
Zhou, T., Yan, L., Wang, Y., & Tan, Y. (2022). Turn your online weight management from zero to hero:
A multidimensional, continuous-time evaluation. Management Science, 68(5), 3507-3527.
37

Online Appendix A
In this section, we present the robustness checks and moderation analysis for two ratio outcomes, i.e.,
𝐺𝑎𝑚𝑒𝑁𝑢𝑚𝐹𝑟𝑖𝑒𝑛𝑑𝑅𝑎𝑡𝑖𝑜 and 𝐺𝑎𝑚𝑒𝑇𝑖𝑚𝑒𝐹𝑟𝑖𝑒𝑛𝑑𝑅𝑎𝑡𝑖𝑜.
Table A.1 displays the DID results after propensity score matching (PSM), which is consistent with
our main findings.
Table A.1. Impact of AI Players on Friend Team-ups after PSM

(1) (2)
Variables
0.136*** 0.136***
(0.005) (0.005)
Player FE Yes Yes
Weekly FE Yes Yes
Note: *** 𝑝 < 0.01 , ** 𝑝 < 0.05 , * 𝑝 < 0.1 ; Robust standard errors
clustered by each player are in parentheses.
Table A.2 shows the results from the relative time model, indicating that the parallel trends
assumption is satisfied.
Table A.2. Impact of AI Players on Friend Team-ups Over Time
(1) (2)
Variables
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=à -0.010(0.010) -0.010(0.010)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,`c -0.005(0.010) -0.005(0.010)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,`d -0.006(0.010) -0.007(0.010)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,è -0.002(0.009) -0.001(0.009)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,`f 0.009(0.010) 0.009(0.010)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,`g -0.002(0.009) -0.001(0.009)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,`b -0.014(0.009) -0.014(0.009)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,B ***
0.348 (0.012) 0.347***(0.012)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,D 0.100***(0.010) 0.100***(0.010)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,b 0.081***(0.010) 0.080***(0.010)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,g ***
0.088 (0.010) 0.089***(0.011)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,f 0.086***(0.010) 0.089***(0.011)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,e 0.099***(0.011) 0.100***(0.011)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,d ***
0.106 (0.011) 0.105***(0.011)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,c 0.116***(0.012) 0.114***(0.012)
𝑅𝑒𝑙𝑎𝑊𝑒𝑒𝑘=,a 0.094***(0.012) 0.096***(0.012)
38

Player FE Yes Yes
Weekly FE Yes Yes
Note: 𝑝 < 0.01, 𝑝 < 0.05, 𝑝 < 0.1; Robust standard errors clustered by each player are in parentheses.
*** ** *
Table A.3 reports the results from Goodman-Bacon Decomposition, which reveal that our findings
are robust after ruling out weights related to problematic comparisons in our DID estimation.
Table A.3. Goodman-Bacon Decomposition

Comparison Weight Estimate Estimate
Earlier treated
0.170 0.244 0.244
vs. Later control
Later treated vs.
0.043 0.307 0.307
Earlier control
Treated vs.
0.787 0.081 0.081
Never treated
Overall DID 0.118 0.118
Unbiased DID 0.110 0.110
We use the following regression specifications to estimate the results of moderating analysis in terms
of skill levels and teammate preferences.
𝑦=> = 𝛽B + 𝛽D 𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> + 𝛽b 𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> × 𝑁𝑜𝑣𝑖𝑐𝑒= + 𝛼= + 𝛿> + 𝜀=> ,
𝑦=> = 𝛽B + 𝛽D 𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> + 𝛽b 𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> × 𝑃𝑟𝑒𝑓𝑒𝑟𝑅𝑎𝑛𝑑𝑇𝑒𝑎𝑚= + 𝛼= + 𝛿> + 𝜀=> .
Table A.4 and Table A.5 detail the results related to two moderating factors. The findings are
consistent with those in the main paper. That is, the positive effect of introducing AI players on friend team-
ups is stronger among novice players and those who initially favor random teammates.
Table A.4. Impact of AI Players: Moderating Role of Skill Levels
(1) (2)
Variables
0.107*** 0.107***
(0.006) (0.006)
0.021*** 0.021***
𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> × 𝑁𝑜𝑣𝑖𝑐𝑒=
(0.007) (0.007)
Player FE Yes Yes
Weekly FE Yes Yes
39

Note: ***𝑝 < 0.01, **𝑝 < 0.05, *𝑝 < 0.1; Robust standard errors clustered by
each player are in parentheses.
Table A.5. Impact of AI Players: Moderating Role of Teammate Preferences
(1) (2)
Variables
0.099*** 0.099***
(0.006) (0.006)
𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> 0.042*** 0.042***
× 𝑃𝑟𝑒𝑓𝑒𝑟𝑅𝑎𝑛𝑑𝑇𝑒𝑎𝑚= (0.007) (0.007)
Player FE Yes Yes
Weekly FE Yes Yes
Note: ***𝑝 < 0.01, **𝑝 < 0.05, *𝑝 < 0.1; Robust standard errors clustered by each
player are in parentheses.
40

SSRN Id4749221 Code3223260

Uploaded by

Copyright:

Available Formats

SSRN Id4749221 Code3223260

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

SSRN Id4749221 Code3223260

Uploaded by

Copyright:

Available Formats

Interlopers or Catalysts?

Dissecting the Impact of Incorporating

Electronic copy available at: https://ssrn.com/abstract=4749221

Electronic copy available at: https://ssrn.com/abstract=4749221

gaming experiences. Therefore, it is essential to understand the impact of integrated AI on player

engagement and the broader game ecosystem.

Electronic copy available at: https://ssrn.com/abstract=4749221

games. Hence, this study aims to fill this research gap.

1.2 Research Questions and Contributions

Electronic copy available at: https://ssrn.com/abstract=4749221

Electronic copy available at: https://ssrn.com/abstract=4749221

games with friends more actively?

Electronic copy available at: https://ssrn.com/abstract=4749221

players of diverse skill levels and teammate preferences.

the churn rate within the game.

Electronic copy available at: https://ssrn.com/abstract=4749221

higher player satisfaction and retention.

these domains and highlights our contributions to existing studies.

2.1 Multiplayer Online Games

Electronic copy available at: https://ssrn.com/abstract=4749221

differentiations influence players’ purchasing behavior.

developers to design strategies to improve engagement.

players on the social aspects of gaming.

2.2 Human-AI interactions

Electronic copy available at: https://ssrn.com/abstract=4749221

expressing positive emotions, providing strategies for improving human-AI interactions.

human intervention in adjusting robo-advisor recommendations on investment outcomes. Fügener et al.

Electronic copy available at: https://ssrn.com/abstract=4749221

unpredictable gaming experience (Hossam 2020).

on player engagement and the tendency to team up with friends.

2.3 Self-efficacy in Social Cognitive Theory

in determining engagement levels.

health management efforts, highlighting self-efficacy as a key driver in regulating behaviors.

Electronic copy available at: https://ssrn.com/abstract=4749221

and the tendency to team up with friends by strengthening players’ self-efficacy.

3. Data and Descriptive Statistics

encounters. Victory is awarded to the last surviving team.

Electronic copy available at: https://ssrn.com/abstract=4749221

Figure 1. The Safe Area and Blue Area in PUBG

Electronic copy available at: https://ssrn.com/abstract=4749221

Table 1. Definitions and Descriptive Statistics of Variables

Variable Definition Mean SD Min Max

Electronic copy available at: https://ssrn.com/abstract=4749221

log(𝑦=> ) = 𝛽B + 𝛽D 𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> + 𝛼= + 𝛿> + 𝜀=> , (1)

(𝐺𝑎𝑚𝑒𝑇𝑖𝑚𝑒𝐹𝑟𝑖𝑒𝑛𝑑=> ). The independent variable in the specification, 𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> , is a dummy variable,

Table 2. Impact of AI Players on Human Player Engagement and Friend Team-ups

(1) (2) (3) (4)

Electronic copy available at: https://ssrn.com/abstract=4749221

friends, addressing our first and second research questions.

𝑦=> = 𝛽B + 𝛽D 𝐴𝐼𝑃𝑙𝑎𝑦𝑒𝑟=> + 𝛼= + 𝛿> + 𝜀=> , (2)

players boosts human player engagement and encourages friend team-ups.

Table 3. Impact of AI Players on Friend Team-ups

Electronic copy available at: https://ssrn.com/abstract=4749221

4.2 Robustness Checks

Table 4. Overview of Robustness Checks

4.2.1 DID Combined with PSM

Electronic copy available at: https://ssrn.com/abstract=4749221

employ the DID regression in Equation (1).

(𝐴𝑣𝑔𝐺𝑎𝑚𝑒𝑊𝑒𝑒𝑘𝑒𝑛𝑑), and during weekdays (𝐴𝑣𝑔𝐺𝑎𝑚𝑒𝑊𝑒𝑒𝑘𝑑𝑎𝑦). Additionally, we employ the average