AI tells are the recognizable signals that reveal a piece of content was generated or edited by AI, from stock phrasing and em dashes to visual artifacts.
Last verified · 2026-10-14 · by Moe Ameen
AI tells are the recognizable signals in a piece of content that give away that a model wrote, drew, or edited it. The word borrows from poker, where a "tell" is an involuntary habit that leaks what a player is holding. In content, a tell is a pattern a generative model produces far more often than a human would: a stock opener, a particular sentence shape, a punctuation habit, a too-even rhythm, or a visual glitch. No single tell is proof, but a cluster of them is what makes a reader think "this reads like AI" before they can say why.
The best-documented tells are textual. According to an analysis by the AI-detection firm Pangram, generative models use em dashes at roughly 17 per 10,000 words against a human baseline of about 2, with OpenAI models closer to 45; they produce bullet lists at about nine times the human rate, lean on "negative parallelism" (the "it's not just X, it's Y" construction) about three times as often, and reach for tidy three-item lists far more than people do. A small vocabulary became notorious for the same reason: "delve," "testament," "tapestry," "underscore," and "navigate" spike in model output. Pew Research Center's August 20, 2026 study of the web found the same fingerprints rising in the wild, reporting that since ChatGPT's release negative parallelism had nearly tripled and AI-favored vocabulary had more than doubled across the pages it sampled.
The single most important thing to understand is that a tell is a probability, not a verdict. Humans use em dashes, write in threes, and love a good bullet list; plenty of AI text has none of the stock markers. Pew itself cautions that its detector can misclassify individual documents in both directions and that its patterns only become meaningful when looking at very large collections of text together, and detection experts stress that no one signal proves a machine wrote something. Tells are also a moving target: the moment one becomes famous, model makers train it down, so the signal that worked last year is weaker this year. That arms race, not any fixed checklist, is the real nature of AI tells.
The discourse around AI tells grew up alongside ChatGPT's mass adoption in 2023 and 2024. "Delve" became the first celebrity tell when researchers and writers noticed how unnaturally often the word appeared in model output, a quirk widely traced to the preferences of the human raters who tuned the models. The em dash followed and overtook it: by 2025 a dense scatter of em dashes was the layperson's shorthand for "a bot wrote this," to the point that human writers who had always used the punctuation found themselves accused of being AI.
That backlash made the em dash the clearest case study in how tells decay. On November 14, 2025, after sustained public ribbing, OpenAI's Sam Altman posted that ChatGPT would finally honor a custom instruction to stop using em dashes, calling it a "small-but-happy win." Models trained against the giveaway: Pangram's later data showed Google's Gemini 3 Pro using em dashes below the human rate, a tell effectively engineered out. The academic grounding arrived the same cycle. Pew Research Center's Data Labs team published "How Much of the Internet Is Written With AI?" on August 20, 2026, analyzing English-language pages from the Common Crawl archive with Pangram's open-weight classifier; it found about 10% of all sampled pages, and over one-third of pages published after ChatGPT's release, carried signs of AI authorship, with the rate far higher on .com domains than on .edu or .gov. By late 2026 the consensus had settled on a nuance the early checklists missed: tells are real and measurable in aggregate, useless as single-document proof, and constantly shifting.
| Platform | Behavior |
|---|---|
| Text | The richest and best-studied tells. Stock openers ("Certainly! Here's…"), hedges ("It's worth noting that"), em-dash density, bullet-list overuse, negative parallelism ("not just X, it's Y"), rule-of-three triads, uniform sentence length, and a signature vocabulary (delve, testament, tapestry) all appear at multiples of the human rate. Also the fastest-decaying tells, since model makers actively train the famous ones down. |
| Images | Tells live in the details a generator struggles to keep coherent: garbled text on signs and labels, extra or merged fingers, warped jewelry and teeth, impossible reflections, backgrounds that dissolve on inspection, and the plasticky, over-smooth "AI sheen." Quality has improved fast, so the reliable checks moved from gross anatomy errors to small inconsistencies at the edges. |
| Video | Temporal tells a single frame hides: objects that morph between frames, hands and fingers that flicker in count, text that shimmers, physics that drifts (liquids, hair, crowds), and unnatural blink or lip-sync timing on avatars. The giveaway is usually motion over time, not any one still. |
| Audio / voice | Synthetic speech tells include too-even pacing, absent breaths, flat or mistimed emotional emphasis, and clipped, uniform pauses. As cloning improves these narrow, and the durable signal shifts to provenance (a watermark or disclosure) rather than the ear alone. |
| Detectors & watermarks | A different class of signal entirely. Statistical detectors like Pangram estimate the probability of AI authorship from the patterns above; watermark and provenance checks (SynthID, C2PA Content Credentials) read a signal the model deliberately planted. The latter is close to decisive when present; a style-based detector only ever reports a likelihood. |
The trap I watch people fall into is treating AI tells as a formatting problem you solve with find-and-replace. Strip the em dashes, ban "delve," cut the triads, and you get content that is marginally harder to pin as AI and exactly as forgettable as before. The tells readers actually punish are not the punctuation; they are the absence of a specific point of view, a real example, a number that came from somewhere. Generic is the tell. The fix is to make the output genuinely yours before it ships, not to sand off the fingerprints after. In practice that means governing voice and substance at the source: Kompozy runs every generation through a [Persona Brief](/glossary/persona-brief) that owns your phrasing and a banned-word filter that kills the signature AI vocabulary by default, and through [quality gates](/glossary/quality-gates) that reject invented statistics and off-voice drafts into review rather than onto your feed. That is a different goal from beating a detector. Beating a detector is an arms race you lose on the next model update; sounding like a person who knows the subject is durable, because the one tell no model can fake for you is having something specific to say. Disclose where a platform or the law asks, then compete on quality, and the tells take care of themselves.
AI tells are the recognizable signals in a piece of content that reveal a generative model produced or heavily edited it — stock openers, em-dash density, bullet-list overuse, "not just X, it's Y" phrasing, rule-of-three triads, a signature vocabulary (delve, testament, tapestry) in text, and glitches like garbled signage or extra fingers in images. A model produces these patterns far more often than a human would, so a cluster of them makes content read as AI.
Less than it was. The em dash became the most famous AI tell, and model makers responded: in November 2025 OpenAI said ChatGPT would honor an instruction to stop using them, and by 2026 Google's Gemini 3 Pro used them below the human rate. The statistical gap persists in aggregate — Pangram still measures AI output using them at several times the human rate — but a single em dash proves nothing, and human writers use them constantly.
No. A tell is a probability, not a verdict. Humans use every one of the stock markers, and plenty of AI text lacks them. Pew Research Center, which measured these patterns across the web, warns that its detector can misclassify individual documents in either direction and that the patterns it found are only meaningful across large collections of text, not any single page. Statistical detectors report a likelihood; only a planted watermark or provenance credential is close to decisive.
The most-cited include "delve," "testament," "tapestry," "underscore," and "navigate," along with stock phrases like "it's worth noting," "in today's fast-paced world," and "certainly! here's." Pew found AI-favored vocabulary more than doubled across sampled web pages since ChatGPT's release. But any one word is weak evidence, and the list shifts as models are tuned away from the famous ones.
Fix the content, not the fingerprints. Bring a specific point of view, a first-hand example, and sourced facts a model cannot invent; avoid the default look and voice; and put a human quality check between generation and publishing. Scrubbing em dashes and swapping out "delve" only makes generic output slightly less detectable — the durable move is to say something specific enough that no model could have produced it for you.