“NAFO Expansion is Non-Negotiable”: Discursive Language of Political Memes through AI-Assisted Metadata for an Activist Meme Archive. Efremov. 6889441.

Abstract

This thesis examines the discursive language of political memes within the pro-Ukrainian online activist movement NAFO (North Atlantic Fella Organization), with a focus on how AI-assisted methods can support the archiving and classification of such materials. Since NAFO’s emergence in March 2022, the movement has produced thousands of memes that combine humor, satire, and ideological critique to confront Russian disinformation and mobilize international solidarity.

The research evaluates the applicability of Courtois and Frissen’s (2022) three-step meme classification framework to a curated set of NAFO memes. Their original method, which relies on image feature matching and clustering through Python scripts, offers a foundation for large-scale visual analysis but struggles to account for the rhetorical and ideological dimensions central to political memes. To address this, the project applied “tool criticism” (van Es et al. 2018), shifting the workflow to OpenAI’s GPT-4o model. By interacting through natural language prompts, GPT-4o was able to simulate metadata generation, propose semantic clusters, and reflect on interpretive ambiguities. This allowed the analysis to foreground discourse pragmatics, intercultural communication, and activist framing, while leaving a transparent audit trail of the tool-assisted process.

The findings highlight several key contributions. First, political meme function often overrides form: identical templates can be repurposed for divergent ideological messages, while visually unrelated memes can convey nearly identical stances. This functional elasticity demonstrates the limitations of visual-only classification systems and underscores the importance of discourse-aware approaches. Second, the project shows that archiving is never neutral. Metadata design choices—such as fields for tone, ideological stance, or target entity—actively shape how memes are framed and interpreted.Third, the thesis demonstrates the promise of hybrid workflows for activist media. While machine vision can detect clusters and patterns at scale, it lacks sensitivity to irony, intertextuality, and cultural nuance. LLMs like GPT-4o, used reflexively and with human oversight, can bridge this gap by generating discourse-aware annotations and clustering suggestions.

Looking ahead, the study proposes future directions for meme research. Large-scale annotation could benefit from hybrid pipelines that combine automated image clustering with LLM-assisted metadata and human validation. Multilingual support will be increasingly important for activist meme analysis, as political memes frequently operate across languages and cultures. Finally, the NAFO archive itself can evolve into a resource for real-time activist deployment, for example as a searchable “meme arsenal” in response to disinformation campaigns.

In conclusion, this thesis argues that political memes such as those created by NAFO are best understood as intercultural speech acts situated within activist communities. Their classification requires workflows that balance automation with human judgement, visual recognition with discourse awareness, and archival preservation with activist framing. By applying tool criticism to both traditional computer vision scripts and contemporary large language models, the study demonstrates how digital humanities methods can capture the ludic, ideological, and intercultural dimensions of political meme culture in ways that serve scholarship, archiving, and activism alike.