Cultural AI Lab

Duke University

Judgment, taste, and AI slop

Generative AI has made content abundant and judgment scarce. We study how people judge and discern in this new environment, with “slop”, the word that emerged almost overnight for unwanted AI content, as our central case. Slop is less a property of content than a social judgment: what one community dismisses, another embraces, and the label travels, hardens, and gets fought over. We measure these judgments at scale on platforms like Reddit, run experiments on how making AI content changes what people value, and ask how new evaluative capacities, human and computational, might be built to match.

  1. AI ‘Slop’ Is a Social Judgement Richard Jean So et al. Nature Human Behaviour, 2026. Read paper
  2. The Social and Cultural Life of Artificial Intelligence (AI) Slop: A Research Framework Richard Jean So and Hoyt Long. Big Data & Society, 2026. Read paper
  3. Why Slop Matters Cody Kommers et al. ACM AI Letters, 2026. Read paper
  4. The Rise of AI Slop In progress.
  5. What’s to Like about AI Slop? Creating It Changes How You Feel about It In progress.

Creativity and interpretive communities

All stories are social, and so is creativity: what counts as original or moving is decided by communities of readers who share conventions and argue over them. Current systems optimize for a single generic reader, an aggregate that quietly reproduces what is already dominant. We ask whether reader difference can be modeled instead. Can populations of AI agents, grounded in real readers, simulate an interpretive community, with its shared tastes, its disputes, and its sense of what is new? What does such a simulation capture about collective judgment, and what does it miss? We use these simulated communities to evaluate AI writing and to surface their responses back to writers, so that AI storytelling answers to actual audiences rather than to no one.

  1. The Social AI Author: Modeling Creativity and Distinction in Simulated Cultural Fields Edwin Roland, Richard Jean So, and Hoyt Long. AI & Society, 2026. Read paper
  2. Evaluating AI Creative Writing through Interpretive Communities In progress.
  3. Readers in the Loop: Building Interpretive Communities for AI Storytelling In progress.

Downstream effects of LLMs on digital platforms

We study what happens when LLMs are deployed at scale on live platforms like Reddit, Substack, and Spotify, and what their increasingly obvious problems — homogenization, model collapse — do to communication and creativity in the wild. We track how AI content spreads through online communities, identify which models are doing the writing, and audit whether streaming recommendations treat AI-made music differently from human artists. Model homogenization is well documented, but whether it produces cultural homogenization is an open question, and one that needs evidence from how people actually make and consume culture. We also study detection. Does it work, or does it mostly punish the wrong people? And how does the suspicion that anything might be machine-made change what artists create and whom audiences trust?

  1. Playing for the Detector: How AI Suspicion Changes What Artists Make In progress.
  2. Model Fingerprints: Which LLMs Are Writing the Internet? In progress.