AI tells are the stylistic habits that readers associate with text written by language models: favorite words such as “delve,” “tapestry” or “underscore,” constructions such as “it’s not just X – it’s Y,” lists of exactly three, relentless upbeat summaries and, most famously, the em dash. Spotting them has become a popular pastime and a source of suspicion – and a poor way to prove who wrote something.
Where tells come from
Language models are trained on vast amounts of text and then tuned on examples of “good” answers, often written or rated by contractors. That fine-tuning is widely thought to reward a particular register: polished, balanced, explanatory and slightly formal, in which certain words and structures become overused relative to ordinary human writing. Researchers have measured the effect. A study by Dmitry Kobak and colleagues, published in Science Advances in 2025, tracked “excess vocabulary” in biomedical abstracts and found a sudden rise in words like “delves,” “showcasing” and “underscores” after ChatGPT’s release, concluding that at least 13.5 percent of 2024 abstracts had been processed with language models. Wikipedia editors maintain a public guide to signs of AI writing to help volunteers clean up machine-generated articles.
The em dash affair
No tell caused more argument than the em dash. In 2025, as users noticed how often chatbots used it, writers who had used dashes for decades complained that their work was being dismissed as machine-made. In November 2025, OpenAI said ChatGPT would finally obey custom instructions to avoid em dashes, though it still uses them by default. The episode illustrated a broader problem: most tells are ordinary features of careful writing that models overuse, not proof of machine authorship. Humans also pick up model habits; researchers led by Iyad Rahwan found that words ChatGPT favors, such as “delve,” became more common in unscripted podcast conversations after its launch.
Detection and its limits
Because tells are statistical tendencies, they say little about any single text. A human can write like a model, a model can be told to write like a human, and editing blurs everything. Studies have found that AI detectors frequently misclassify human writing, especially by non-native English writers, as machine-made. Experienced readers can be fairly accurate in aggregate, but no single tell proves authorship. Provenance tools such as watermarking and content credentials can help, mainly for images and video; text watermarks remain rare and fragile. Meanwhile, AI tells have become a cultural shorthand for low-effort content, feeding the backlash against AI slop and workslop.
Why it matters
AI tells sit at the center of a new kind of social suspicion: “did a machine write this?” The question matters for trust, education and authorship, but tells are weak evidence and can unfairly target people. A healthier norm is disclosure – saying when AI helped – combined with judging writing by its substance. As models diversify their styles, today’s tells may fade and new ones will likely replace them, while the honest question remains whether the work is good and whether the author stands behind it (see AI-assisted authorship).
Written with Claude, made by Anthropic; this entry may contain some of the tells it describes.
Related
AI Slop · Workslop · Watermarking · AI-Assisted Authorship · Glazing · AI Shame
Sources: Dmitry Kobak et al.: Delving into LLM-assisted writing in biomedical publications through excess vocabulary (Science Advances) (July 2, 2025) · Wikipedia: Signs of AI writing · Hiromu Yakura et al.: Empirical evidence of Large Language Model’s influence on human spoken communication (arXiv) (September 2024; revised July 2026) · Weixin Liang et al.: GPT detectors are biased against non-native English writers (Patterns) (July 2023) · Jenna Russell, Marzena Karpinska, Mohit Iyyer: People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text (arXiv) (January 2025) · Sarah Perez: OpenAI says it’s fixed ChatGPT’s em dash problem (TechCrunch) (November 14, 2025)
Leave a Reply