In 2026 the signs of AI writing are long sentences chained with “and”, nouns built from verbs, metaphor where a plain statement would do, and short closing lines written for effect. “Delve” has mostly gone. This page lists each tell with where it was measured or reported, and how far to trust it.
Two cautions before the list. A single tell proves little, since people write lists of three and closing lines too, and it’s several together that make a reader suspicious. And none of this is a way to accuse someone. Shan, Lee, and Hao (EMNLP 2026) found that text a model edits has a much weaker footprint than text a model generates, so evidence about generated text shouldn’t be turned on a person’s edited draft.
Last checked on 10 October 2026. The entries about specific Claude versions date fastest.
These come from published corpus comparisons. They are the strongest evidence on this page, and they still have limits: The Economist measured news journalism, and the Reinhart figures are from 2024 models.
"the implementation of", "utilization", "significant", "increasingly"
Polysyllables, Latinate suffixes, and nouns made from verbs. All four models in the study do it; Gemini and Claude do it most. Reinhart et al. put nominalisations at 1.5 to 2 times the human rate.
"And" is the most overused word in the study, and paragraphs are rarely broken by a short statement.
Fewer commas and semicolons than human writers, and almost no parentheses.
"It's not just a defensive measure — it's a foundational component."
ChatGPT and Claude use these most per 1,000 sentences.
"..., underscoring its importance"
Measured at 2 to 5 times the human rate. It attaches significance to a fact without giving evidence for it.
Of the four models, only Claude uses more em dashes than human writers. ChatGPT uses fewer than any other writer in the study. The article gives no figure for the size of the gap.
"studies show", "experts say"
The models do not quote named people.
Anthropic describing its own model. It is not a measurement, and it compares the model with its predecessor, not with people.
"a dial worth turning" for "a parameter worth varying"
Metaphor and flourish in place of direct statement. Anthropic's guide names the habit and gives this example in its suggested prompt.
Sentences run longer and there are fewer paragraph breaks than in the previous model.
These come from two blog posts summarising forum threads about Claude Opus 5. Nobody has counted them, and one widely shared post can make a habit look more common than it is.
"load-bearing", "carries the argument"
None of these is a banned word. "Load-bearing" is the right term for a wall.
"worth stating plainly", "honestly", "frankly"
A sentence that says the next one will be direct.
"Not a detail. A design decision."
A short sentence that repeats the one before it for effect.
A rule built on any of these is aimed at 2023.
These sit on Wikipedia's list for 2023 to mid-2024. Its list for mid-2025 onward is four words: "emphasizing", "enhance", "highlighting", "showcasing". The page records that "delve" fell sharply in 2025.
Version 1.x called this the strongest signal, on the strength of one 2024 study of Mistral, Falcon, and LLaMA. A 2026 test of 284 features across 27 models found most proposed indicators depended heavily on context.
Pangram, whose detector does best in independent tests, argues that both are unreliable and does not use them.
Avoiding these tells does not get text past an AI detector, because detectors aren’t reading for them. Current commercial detectors are trained classifiers. Pangram’s technical report describes a model trained on paired human and synthetic text, and the product now has a separate head for text that has been through a humanizer. Pangram reports catching the output of nineteen humanizer tools more than 90% of the time, which is the vendor’s own figure.
Xu et al. (2026) found that GPTZero and Pangram often classify text from base models as human and text from the instruction-tuned versions of the same models as AI. Their conclusion is that detectors “are tracking artifacts of instruction tuning and local context”. Getting past them took a fine-tuned paraphraser applied repeatedly, and the published attacks of that kind use the detector’s own score to guide the rewrite.
A style guide read by an instruction-tuned model doesn’t change that model’s tuning. So humanizing-writing makes no claim about detector scores. The reader it’s written for is a person — and practised people are good at this. Russell, Karpinska, and Iyyer (2025) found that the majority vote of five frequent LLM users misclassified 1 of 300 articles.