P002 → Synthetic methods
Exploring #redpill video shorts with AI
Participants: Sara Hammerschmidt, Johann Pibert, Eva Nieto McAvoy, Yihan Wang, Myrtle Ziwei Zeng, Carolin Gerlitz, Ángeles Briones
This project examines AI in digital research by advancing “synthetic methodologies” for studying short-form video. Situated in Fabricating the People (CRC Transformations of the Popular), it analyzes Turkish #redpill shorts across Instagram, TikTok, and YouTube. We combine ethnographic immersion with machine patterning—zero-shot affect tagging and LLM-driven NER—to trace narratives, archetypes, and affects at feed scale. Treating LLMs as hermeneutic partners rather than truth engines, we iterate between human annotation and automated sorting to surface modal braiding and productive ambiguity. Leaning into and explicating AI’s analytical capacities, we test what AI renders legible and what it excludes, foregrounding knowledge co-produced by humans and machines. That is to say, can we use its recursive logic to our advantage to find different patterns in the data instead of trying to force it to comply with our own interpretations of the space from the get-to. Moving back and forth between human interventions and zero-shot machine sorting of the space, we attempted to structure the data in ways that was neither fully human nor fully machinic.
NER Group collages
1. Given that there are no clear groups of specific individuals connected through affect, we spontaneously decided to run a categorization ZSL on the NERs to extend a grouping logic to these NERs
2. How does this layering of human-machine interpretation find and reveal new patterns in the space? What can and can't we say about this?
3. What affects build or more across the space when one *group* is looked at together? How does the modal braiding -- the dialectics and tensions that build as the interface encourages users to add new modal layers to their messages -- shift these spaces? Which affects (or ideologies) becomes embodied and how?
Muscle Men
Philosopher
Synthetic feed
1. Focus on the level of the feed and how affects (as interpreted by the LLM) are built as they repeat across the cultural space. 2. What did this experience show us we can leverage the LLM to do in terms of identifying affect or ambience? What was different about its vision and our experiences? What did it force us to look at?
Turkish thumbnails
NER networks
1. Is there a clear economy of the images used in this space? Are some clearly associated with specific affects or issues? 2. Is this level of analysis helpful in understanding the way that affect and atmosphere gets built in Turkish redpill feeds?
Monsters
1. We turned our zero-shot classification task of the NERs into an opportunity
2. Building on the idea that our goal is to try to turn the LLM's "weirdness" into opportunities for insight, we asked: When the LLM cannot classify a named entity using the accompanying visual description, what does this suggest about how the LLM interprets bodies? How it see and understand people? Which aspects of the background does it seem to be holding on to and focusing on? Is there an opportunity here to collapse the multimodality of these spaces and capture the "people" and "characters" used here in a one-dimension way that can heighten the aesthetic and affective alliances they rely on to build their community (or spread their ideology?).
Arnold Jesus
Arnold
Jesus