The Impact of Generative AI on Social Media: An Experimental Study

Generative Artificial Intelligence (AI) tools are increasingly deployed across social media platforms, yet their implications for user behavior and experience remain understudied, particularly regarding two critical dimensions: (1) how AI tools affect the behaviors of content producers in a social media context, and (2) how content generated with AI assistance is perceived by users. To fill this gap, we conduct a controlled experiment with a representative sample of 680 U.S. participants in a realistic social media environment. The participants are randomly assigned to small discussion groups, each consisting of five individuals in one of five distinct experimental conditions: a control group and four treatment groups, each employing a unique AI intervention—chat assistance, conversation starters, feedback on comment drafts, and reply suggestions. Our findings highlight a complex duality: some AI-tools increase user engagement and volume of generated content, but at the same time decrease the perceived quality and authenticity of discussion, and introduce a negative spill-over effect on conversations. Based on our findings, we propose four design principles and recommendations aimed at social media platforms, policymakers, and stakeholders: ensuring transparent disclosure of AI-generated content, designing tools with user-focused personalization, incorporating context-sensitivity to account for both topic and user intent, and prioritizing intuitive user interfaces. These principles aim to guide an ethical and effective integration of generative AI into social media.

Introduction. The rapid integration of artificial intelligence (AI)-driven text generation tools into social media platforms is reshaping how users create and engage with content, raising new questions about their effects on the quality and dynamics of online interactions [1, 2]. AI writing tools notably reduce barriers to content creation by lowering required effort and expertise. Although these technologies are increasingly adopted across sectors such as journalism [3, 4], education [5], and creative industries [3], their potential impact is particularly pronounced in social media contexts. Social media platforms play a central role in shaping public discourse, influencing democratic engagement, and enabling rapid, large-scale dissemination of information and ideas [6]. Therefore, introducing AI into these platforms may reshape dynamics of interaction, authenticity, and nature of online discussions [7, 8]. Previous studies highlight the promise of AI assistance in advancing human creativity [9], increasing user engagement [10], and facilitating broader inclusivity in online discussions [11, 12]. Yet, this optimism is tempered by concerns about potential drawbacks, such as declining content quality [13], proliferation of misinformation [14], and diminished authenticity of user interactions [15]. Empirical evidence quantifying how AI assistance reshapes online participation dynamics, content quality, and user perceptions remains scarce, particularly in realistic, platform-integrated scenarios [16]. Addressing this critical gap, our study provides empirical insights through a controlled experiment conducted on a realistic social media platform. We approach the debate about AI-assisted content creation from two complementary perspectives: first, how AI assistants in social media affect the experience of content producers, and second, how these interventions shape consumers’ perceptions of content. Our experiment fills a crucial gap by directly assessing how AI assistants transform both producer and consumer experiences on social media. We conduct the experiment using a custom-built platform that closely simulates an online chat room resembling discussions on common social media forums. A representative sample of 680 U.S. participants is partitioned into groups of five people, who are then randomly assigned to one of five conditions: one control (no AI assistance) and four non-overlapping treatment conditions, each featuring distinct AI interventions previously proposed for enhancing online interactions. These interventions include (1) an open-ended chat with an AI assistant [17], (2) AI-generated reply suggestions with varying stances (agreeing, neutral, disagreeing) [7, 18], (3) AI-driven feedback on comment drafts [5, 19], and (4) AI-generated conversation starters [20]; interventions are described in more detail in the Methods section and in “AI Prompts and Settings" in the Supplementary Information (see Fig. 1 for platform screenshots). Together, these interventions represent a comprehensive range of AI-based approaches considered by both researchers and industry practitioners for enhancing online interactions. To assess the heterogeneity of the interventions’ effects across topics, participants sequentially discuss three randomly-ordered topics—ranging from conversational (cats vs. dogs), to scientific (health benefits of oats), to political (universal basic income)—with each topic limited to a 10-minute interaction. We assess various proxies for user engagement and quality of experience through questionnaires before and after participation, and further track participants’ interactions on the platform—including comments, reactions, and AI usage—to comprehensively evaluate how AI interventions impact both content producers and consumers. Overall, our findings indicate that AI assistance substantially influences both usergenerated content quality and consumer perception, although with notable variation across interventions. AI-supported participants demonstrate increased engagement and content production metrics, but these improvements are associated with nuanced, sometimes negative, shifts in consumer perceptions and reactions. Critically, no single AI tool enhances both producer and consumer experiences, highlighting complex trade-offs. AI interventions generally increase participants’ willingness to engage and improve aspects of content creation from the producer perspective. Participants using the Chat and Suggestions features notably report that the AI would increase their willingness to participate in online discussions, compared to control (Fig. 2a). Additionally, all AIsupported tools significantly increase the average length of user comments (Fig. 2c). Participation equality among users, measured by normalized Shannon entropy based on the proportion of comments per user in each round, noticeably improves under the Conversation Starter intervention, indicating more balanced participation (Fig. 2e).

Method. Platform Design responded to. Reactions would be shown to all participants but without identification of who reacted. New comments were highlighted with a brief alert and a “new" tag lasting one minute to support real-time flow. The platform was designed to minimize distractions and reduce cognitive load, allowing participants to focus on the conversation. The setup ensured controlled conditions while preserving key elements of real-world online discussions. Before entering the discussion platform, participants were shown a brief onboarding interface introducing the platform and their task. This included a written description of the discussion structure and a short demonstration video showing how to navigate the platform, post comments, and react to content. For participants assigned to AI-assisted conditions, additional instructions were provided explaining the specific tool available. The demonstration video included a walkthrough of how to access and use the AI feature. During each 10-minute discussion round, we log user activity at the individual level, including comments posted, reactions given, and use of AI-tools.

To test the impact of different paradigms of AI assistance in social media discussions, we integrated four distinct AI tools into the platform. Each tool was designed to reflect approaches to AI-assisted communication found in academic literature and commercial products. All tools were powered by GPT-4o and with a custom prompt tailored to each intervention (see “AI Prompts and Settings" in Supplementary Information). Participants were not informed which AI model was used in treatments or whether any content they encountered was AI-generated.

Chat. The Chat tool allowed for open-ended interaction with an AI assistant through a sidebar window displayed alongside the conversation thread (Fig. 1f). Participants could engage with the assistant up to eight times per discussion topic. The interface supported informal querying, idea generation, and clarification, giving users flexibility to steer the interaction as desired.

Conversation Starter. The Conversation Starter generated AI-suggested openings for participation, aimed at lowering barriers to entry and stimulating discussion (Fig. 1b). The tool was accessed through a separate button on each comment. The conversation starting suggestions were context dependent but could include follow-up questions, engaging comments, and reflective or contextual statements.

Feedback. The Feedback tool offered real-time guidance on draft comments prior to submission (Fig. 1d). Once users began typing a comment, they could click on a dedicated “AI Feedback" button to receive tailored suggestions on how to improve their comment. The feedback varied based on context and comment draft, but could include ideas to clarifying arguments, add personal anecdotes, or maintain a balanced tone. The feedback appeared inline below the draft comment, and users could receive three rounds of feedback per comment.

Suggestions. The Suggestions feature provided three AI-generated replies—each adopting a distinct stance (agree, neutral, disagree)—in response to any selected comment (Fig. 1e). The tool was accessed through the comment modal. Participants could regenerate a new set of suggestions up to three times per comment, allowing them to explore alternatives before selecting a reply.

Each of these tools was embedded into the platform interface to mirror familiar social media interactions while maintaining clarity and minimalism. The tools were designed to be optional, harmoniously integrated into the platform, and supportive of the natural flow of the discussion.

Discussion. In exploratory regression analyses assessing whether demographics (age, gender, sex, education, and political affiliation) would alter these results, we observe only minor effects that did not reach statistical significance. While preliminary, this likely reflects limited sample sizes across treatment groups (see Supplementary Information for additional details).

Building on the duality—AI interventions boosting participation yet risking polluting conversation with low-quality content—we also uncover nuanced differences in how participants employed each tool. Usage patterns are far from uniform and often reflect the tools intended design and conversational context. These variations merit closer auditing, both to clarify which AI-aid paradigm holds the most promise for meaningful engagement and to guide future employments and refinements. Figs. 3a-d and 4hi detail key usage patterns (see the AI Usage Analysis section and “AI Usage" in Supplementary Information for details on the analyses).

These results entail important implications for the deployment of generative AI on social media platforms. Although participants in the treatment conditions produced more content, they often recognized it as generic, impersonal, and of lower quality. Such reduced perceived authenticity diminishes trust and informational value—risks that align with recent concerns about AI-assisted content flooding digital platforms with low-quality text [22–24]. A central concern is that widespread use of generative AI may saturate platforms with superficial or generic content, diluting the visibility and impact of more original contributions—and, as our findings suggest, lowering the quality of subsequent conversations within threads, even among users not using the AI themselves. Nevertheless, our findings show conditions under which AI tools may enhance public discourse. Participants valued the AI for idea generation, clarity, and initiating engagement—particularly in perceived high-stake or cognitive demanding topics. If tools are developed with greater personalization, contextual awareness, and stylistic nuances, they could support more inclusive and constructive interactions— particularly for individuals typically hesitant to engage in public discussions. To guide ethical and effective deployment, we propose four design principles:

1. Appeal of optional use and transparent disclosure: Our findings suggest that AI tools are most positively received when offered as optional. Across all treatment conditions, usage patterns indicate selective engagement. In the Chat condition, participants submitted an average of 3.36 to 3.55 prompts per round, with only 13- 16% of their final comments showed direct textual overlap. This suggests that users primarily engaged with the tool for ideation, clarification, or rhetorical guidance rather than direct copying. Similarly, for the Conversation Starter, only between 16% and 19% of submitted comments overlapped with AI suggestions—indicating that users often adapted, modified, or disregarded them based on text quality or alignment with their intent. The Feedback tool, although more cognitively demanding and less widely adopted, proved highly effective when used: in over 89% of cases where participants used the tool when composing a comment, they subsequently submitted it to the discussion. These behavioral signals are mirrored by open-ended questionnaire responses. Participants reported appreciating AI support for generating ideas, retrieve information, and overcoming initial writer’s block.

Conclusion. These behavioral and perception perspectives affirm the importance of intuitive and convenient interfaces in encouraging tool engagement. Participants in the Chat and Suggestions conditions find these tools as more supportive overall—reporting that they made participation easier, felt more intuitive to use, and led to higher-quality contributions—compared to those using Feedback and Conversation Starter (see “Supplementary Figures" in Supplementary Information). Altogether, these findings support a key design principle: simplicity and familiarity in user interfaces are vital to enhance adoption and usage. When AI tools are embedded seamlessly into platform workflows—with easy entry points and low interaction costs—they are more regarded as a supportive mechanism rather than disruptive. Future implementations should prioritize clarity and accessibility by embedding AI tools into the natural flow of conversation, minimizing user effort without sacrificing optional depth.