TL;DR
This paper argues that identifying AI-generated writing is less about spotting specific "tells" and more about detecting pragmatic infelicity—a mismatch between a text and its social context. By analyzing Wikipedia editors and students, the researchers show that "AI-likeness" is a moving target used to police sociolinguistic norms, meaning that what one community calls "AI" another might call "human."
What Problem Does This Paper Address?
As LLMs become ubiquitous, the "human provenance" of a text can no longer be assumed. While technical research focuses on building better automated detectors, this paper looks at the folk theories and social practices of AI identification.
The authors identify a critical gap: most detection tools treat "AI-likeness" as an inherent, context-free property of a string of text. However, in the real world, humans identify AI based on social expectations. If a detection tool is trained on one context (e.g., academic essays) and applied to another (e.g., Wikipedia or personal texts), it doesn't just detect AI; it enforces the social norms of its training data at scale. The paper seeks to understand how people actually decide something "feels like ChatGPT" and how these judgments evolve within communities.
Methods
The researchers employed a mixed-methods approach combining linguistic anthropology with computational linguistics:
- Qualitative Coding: They developed a two-level schema to categorize the reasons people give for flagging AI, covering Textual Features, Content, Author traits, "Vibes," and External Resources.
- Thematic Analysis: Iterative analysis of open-ended survey responses to understand the social motivations behind AI accusations.
- Computational Lexical Analysis: They applied Heaps’ Law (or Herdan’s Law) to measure how the vocabulary of "AI markers" grows and compresses over time. They also tracked hapax legomena (words used only once) to see if the community was settling on a fixed set of features or constantly inventing new ones.
Datasets
The study draws on two distinct real-world contexts:
- Wikipedia Metadata: 11,882 instances of English Wikipedia editors using the
{{AI-generated}}cleanup template between October 2024 and September 2026. This included 6,046 tags with specific "reason" strings provided by the editors. - Student Surveys: 74 surveys (yielding 220 free-text responses) from high school and undergraduate students in Vermont, reflecting on their perceptions of AI writing and its social impact.
Key Insights and Ideas
- The "Goldilocks" Problem: AI identification is often based on contradictions. One person flags a text for being "too flowery," while another flags it for being "too bland." As shown in the study's findings, the same features can be used as evidence for opposite conclusions.
- Contextual Inversion: What counts as "AI-like" depends entirely on the community's norms. For students, neutrality and a lack of emotion signal AI. For Wikipedia editors—who value a neutral encyclopedic tone—enthusiasm and "promotional" language signal AI.
- Enregisterment: The community is actively "enregistering" AI—creating a recognizable "AI voice" or "accent." This isn't a static list of features but a dynamic process where new markers (like the much-discussed em-dash) are constantly added to the "AI register."
- Detection as a Social Weapon: Exploratory results suggest that calling something "very ChatGPT" is often a rhetorical strategy to shut down an argument or dismiss an interlocutor as "fake" or "inhuman," regardless of whether AI was actually used.
Key Results
- Vocabulary Growth: The Heaps exponent ($\beta \approx 0.59$) for Wikipedia reasons indicates an "open" vocabulary. The list of features editors cite is still growing, suggesting the "AI register" hasn't fully settled yet.
- Shift to Authority: Over time, Wikipedia editors stopped describing textual features and started deferring to community authority. Citations of textual features dropped from ~80% to under 20%, while citations of the "AI noticeboard" rose to over 80%.
- Pragmatic Infelicity: The researchers found that identifications are rarely about a feature existing, but about it being "too much" (e.g., "overly formal," "too detailed"). This confirms that AI detection is a judgment of appropriateness rather than provenance.
Limitations and Future Work
The authors acknowledge that their survey sample is small and geographically specific (Vermont). Furthermore, the Wikipedia data represents a relatively short 10-month snapshot of a rapidly evolving phenomenon.
Future work plans to move beyond static text to look at interactional components: How do people react when accused of using AI? How do these accusations shape identity and community formation on platforms like Reddit and TikTok? The authors also suggest investigating how "Author" and "External" signals (like editing speed or reputation) are becoming more important than the text itself as LLMs become better at mimicking human styles.
Summary updated