Technical Summary: The Meme Is the Message: Generative Memesis and AI Visuals in the 2024 USA Presidential Elections

Problem Statement

Visual content on social media significantly shapes political discourse but historically imposes high production costs that limit civic participation. The emergence of generative AI (GenAI) offers a potential solution by lowering these barriers, yet concerns persist regarding its role in misinformation and the manipulation of information environments. While prior research has examined the impact of social bots and automated distribution, there is a scarcity of literature analyzing the production and curation of GenAI content during critical political events. Specifically, it remains unclear how GenAI interacts with established viral formats like memes, how different ideological groups utilize these tools, and whether synthetic content inherently drives engagement or requires human mediation to be effective.

Methodology

The authors analyzed a dataset of 239,526 Instagram images collected between April 5, 2024, and August 9, 2024, covering the 2024 U.S. presidential election. The data was retrieved via Meta's CrowdTangle API using a snowball sampling strategy of keywords and hashtags related to the election.

The study employed a multimodal workflow combining computer vision, large language models (LLMs), and facial affect analysis:

1. Visual Element Identification: The team utilized GPT-4o-mini with a zero-shot labeling approach to classify images. This included detecting thematic objects (e.g., war, immigration, politicians), identifying meme formats, and estimating the probability of a post being synthetic.
2. Synthetic Detection: The tool Provenance (developed by OpenAI) was used to flag AI-generated images. Manual verification on a random sample of 500 images confirmed the automated labeling accuracy at approximately 85–92%.
3. Affect Analysis: To analyze facial expressions, the authors used Py-Feat, an open-source tool based on the Facial Action Coding System (FACS). This allowed for the quantification of specific facial action units (AUs) to determine emotions (e.g., happiness, anger) in human faces within the images.
4. Opinion Leader Identification: The top 139 content-producing accounts were identified based on a weighted score of post volume and average view counts. These accounts were manually labeled for media bias using external fact-checking resources.
5. Statistical Analysis: The study employed Ordinary Least Squares (OLS) regression to model engagement (Likes and Comments) as a function of content format (meme vs. non-meme), production method (AI vs. non-AI), and their interaction, controlling for follower counts and thematic topics.

Key Contributions

The paper makes three primary contributions to the literature on political communication and AI:

1. Format vs. Production: The study establishes that meme format is a stronger predictor of engagement than the mere presence of AI-generated content. While pure synthetic content does not inherently increase virality, the interaction between AI generation and meme format yields a synergistic increase in engagement.
2. Ideological Divergence: The research characterizes distinct partisan strategies in AI usage. Democrat-leaning users predominantly employ GenAI for in-group support (positive campaigning), whereas Republican-leaning users more frequently utilize it for out-group attacks (negative campaigning).
3. Generative Memesis: The authors define a new socio-technical phenomenon termed "generative memesis." This describes a mode of communication where memes are no longer transmitted person-to-person via editing existing frames, but are mediated by AI through customized visuals. This concept highlights a bifurcation in the production process: generation (automated) and curation (human).

Results

• Engagement Dynamics: Meme formats significantly increased both Likes and Comments. Pure AI-generated content (without meme formatting) showed lower virality compared to the general pool of election-related content. However, posts that were both AI-generated and formatted as memes demonstrated a significant positive interaction effect, particularly for Likes.
• Temporal Correlation: The volume of AI-generated memes correlated with non-AI memes, peaking during major election events (e.g., the assassination attempt on Trump, the Harris-Biden transition announcement). However, a notable gap emerged during Trump's conviction, where non-AI memes dominated (14%) over AI memes (8%), suggesting underlying events moderate the supply of synthetic content.
• Facial Affect: Users generally curated synthetic faces to appear happier and more neutral compared to real photographs, which tended to display more anger. This suggests a psychological bias in human curation of AI content toward positive affect.
• Opinion Leaders: The top 139 opinion leaders utilized memes at a rate of 27.6% (double the average user rate of 13.7%) and AI-generated images at 8.4%. Non-legacy and entertainment-based accounts were the primary drivers of meme and AI content, while legacy media focused more heavily on traditional politician imagery.
• Partisan Asymmetry: Democrat-leaning accounts showed a higher propensity for using AI for positive messaging (16% positive vs. 14% negative), while Republican-leaning accounts skewed toward negative messaging (11% positive vs. 15.5% negative).

Significance and Claims

The paper argues that generative AI is reshaping the political information environment not by displacing human agency, but by reconfiguring it. The authors claim that:

• Shift in Agency: Unlike previous AI interventions (e.g., social bots) that operated in the distribution layer of the information funnel, GenAI operates in the supply layer. This shifts the critical question from "who is sharing?" to "who is speaking?" and challenges source credibility.
• Human Curation is Essential: The "spreadability" of synthetic political content remains contingent on human intervention. AI generates the raw material, but human actors act as cultural intermediaries who select, contextualize, and frame content to align with recognizable political narratives and meme conventions.
• Bifurcation of Production: The study posits that GenAI creates a bifurcation between generation and curation. While AI lowers the cost of production, the engagement value is derived from the human curation that integrates these visuals into culturally salient frames.
• Generative Memesis: The concept of "generative memesis" serves as a new socio-technical layer in visual communication systems, where the transmission of ideas is mediated by customized AI visuals rather than simple replication of existing frames.

The authors conclude that while GenAI empowers civic participation by lowering production barriers, its communicative efficacy depends heavily on the interplay between algorithmic generation and human curation. Future research is needed to understand how different communities develop norms around synthetic curation and how platform affordances shape the spreadability of AI-mediated content.