The signs of AI writing, with numbers

The clearest signs in 2026 are leftover Markdown, unfilled placeholders, em dashes, a small set of overused words and "rather than" contrasts. We measured each one on 3,379 human and AI texts.

Signs that point to AI

The percentages show how many texts contain the sign at least once: first in unedited chatbot output, then in text people typed themselves (essays, emails, forum answers).

Markdown that came along with the copy

Chat windows format answers with Markdown. Paste the answer somewhere else and the double asterisks, # headings and bullet markers come with it. People typing an email or an essay do not write this way. 53% of chatbot texts, 0% of typed human texts.

Blanks nobody filled in

[Your Name], [Company Name], [Date]. A chatbot leaves these for the user, and a user in a hurry leaves them in. 30% of chatbot texts, 0% of typed human texts.

Em dashes

There is no key for the em dash, so it is rare in typed text and common in chatbot text. It is also common in edited, published writing, which is why it only counts together with other signs. 65% of chatbot texts, 0.5% of typed human texts.

A small set of overused words

Current models lean on genuinely, crucial, fundamentally, conversely, effectively, broader, landscape and ultimately. Each is an ordinary word. What gives the text away is how many of them turn up together. 78% of chatbot texts, 9% of typed human texts.

"Rather than" and other contrasts

Models like to define a thing by what it is not: "X rather than Y", "not just X, but Y", "it's not X, it's Y". 10% of chatbot texts, 3% of typed human texts.

Lists of three

"Clear, concise, and effective." People write lists of three too, but less often. 91% of chatbot texts, 34% of typed human texts.

Stock phrases

"Plays a crucial role", "relies heavily on", "lies in", "driven by", and in emails "I hope this email finds you well". 32% of chatbot texts, 3% of typed human texts.

Even sentence rhythm

People mix short sentences with long ones. Models keep sentences closer to one length. This is a matter of degree and only shows over a few paragraphs.

Signs that point to a person typing

Models do not make typing slips unless they are told to. These turn up in typed human text and almost never in AI text.

Sign	Typed human texts	AI texts
A lowercase "i" for the pronoun.	12%	0.2%
A contraction without its apostrophe: dont, cant, im.	15%	0.2%
A common misspelling: definately, recieve, alot.	8%	0%
A hyphen with spaces around it, used as a dash.	10%	0.6%
A sentence that starts in lowercase.	25%	5%

Things that look like signs but are not

Curly quotes. They are in 79% of chatbot texts, but Word, Google Docs and phones produce them for everyone.
Good grammar. Plenty of people write carefully. Clean text is not evidence.
Formal vocabulary. Academic and business writing uses many of the same words. In our published human texts, 19% contain at least one word from the overused list.

How to use this list

No single sign settles anything. Each one appears in some human writing, and all of them can be edited out of AI writing. What carries weight is several signs together in a text that was supposedly typed by hand. To count them in a text of your own, paste it into the checker. For how well that count works, see the accuracy page.