How to tell if something was written by AI

Every claim below is cited to a measured source. Updated August 2026.

The short version: stop hunting for magic words and start reading structure. The famous giveaways of 2023 and 2024, "delve" above all, are being tuned out of new models, but the way AI organizes sentences has proven much harder to hide. Wikipedia's editors, who clean up more AI text than anyone on earth, open their field guide with a blunt warning that automated detectors are basically useless; trained eyes reading for patterns do better.

The loudest current sign: "It's not just X, it's Y"

Writing researchers call it negative parallelism: defining a thing by first denying something about it.

A Washington Post analysis of 328,744 real ChatGPT messages found a variant of this construction in roughly 6 percent of them, and Wikipedia's guide names it among the most prominent signs. Humans write this sentence occasionally, for actual emphasis. Models reach for it constantly, as decoration.

Punctuation: the em dash, with a 2026 twist

For two years the em dash was the meme tell, and the data backed the meme: GPT-4.1 produced about 10.6 em dashes per 1,000 words against a measured human baseline of 3.2, per SlopDetector's corpus work. The twist is that the labs listened. OpenAI's newest models are now below the human rate, and a July 2026 analysis found only one major chatbot still out-dashing professional writers: Claude. So a dash-heavy text today points to a narrower set of suspects than it did a year ago, and a dash-free text clears nobody.

Vocabulary: real, but fading at the source

The word-frequency evidence is spectacular: "delves" appeared at 28 times its expected rate in 2024 biomedical abstracts (Kobak et al., Science Advances), "underscores" at nearly 14 times, and stock phrases like "plays a significant role in shaping" run 182 times more common in AI text than human text by GPTZero's count. But treat word lists as dated evidence. Once a word becomes a meme, the next model generation avoids it, while the published record and even human speech keep absorbing it. Our full word list tracks which ones are rising and which are being retired.

The structural signatures that persist

These are harder for labs to tune away, because they live in how models compose rather than which words they pick. The rule of three: triplets of near-synonyms used for rhythm ("innovative, transformative, and groundbreaking"). False ranges: "from X to Y" constructions that sound specific and say nothing ("from intimate gatherings to global movements"). Vague attribution: "experts say" and "studies show" with no expert and no study. Trailing participles that fake analysis: a fact, then a comma, then "highlighting the importance of..." And the compulsive summary ending that restates the piece you just read. One stylometric study found 82 percent of AI posts share a uniform sentence-cadence fingerprint regardless of which model wrote them; human rhythm is lumpier.

A caution about accusing people

None of these signs is proof. Humans wrote "delve" before ChatGPT did, and research shows AI vocabulary is now leaking into human speech: Max Planck researchers measured a significant post-2022 rise of GPT-favored words in podcasts and unscripted talks. Meanwhile academic stylometry can attribute text to a specific model with high accuracy in lab conditions, yet the consumer detector tools built on similar ideas produce false accusations regularly. Read for the cluster of signs, not any single one, and hold the conclusion loosely.

Sources: Wikipedia: Signs of AI writing · Washington Post message dataset (via Decrypt) · SlopDetector em-dash data · Kobak et al., Science Advances 2025 · GPTZero AI vocabulary · Yakura et al., Max Planck · Sun et al., CMU stylometry

Part of HumanSounding, the live index of AI writing tells. Figures quoted from their sources without adjustment; the trend data behind this site updates weekly.