How readers spot AI writing: the structural tells

Sitefire, the company behind SlopShape, published how often each habit showed up in AI posts and in human posts:

All four come from one habit: the post makes its point three times, in the roadmap, in the summary and in the close. A clear thesis is often good writing. Stating it three times is the template.

Two more habits run against popular advice. AI posts framed the topic as an old way against a new way in 76% of cases, against 26% for human posts. And Sitefire reports AI posts escalating the stakes in 88% of cases against 52% (a figure from the earlier version of the paper). Adding a “the old way is broken” contrast to sound more human moves a post toward the AI default.

The paper also lists the features that mark human posts. Most are things a model writing from a 120-word brief couldn’t have:

• things to use: templates, checklists, calculators, code, links,
• a piece of standard advice the author says is wrong,
• the reader’s situation, described back to them,
• named outside people and sources, with clear attribution,
• comparisons between real options,
• claims from the author’s own work: “we tested”, “since 2016”,
• real dates and periods.

We think this is the most useful part of the research for anyone who edits. The tells are mostly absences. A post with a number from your own work, a named type of customer, a date and a link reads differently before anyone touches a sentence. Our guide on why AI content converges goes into where that material comes from.

Word-level tells have been measured:

• “delves” appeared at 28 times its expected rate in 2024 PubMed abstracts (Kobak et al., Science Advances 2025).
• GPT-4o uses present participial clauses, the “, highlighting the need for…” kind, at 5.3 times the human rate (Reinhart et al., PNAS 2025).
• In Graphite’s September 2026 study, “the _ is not simply” ran at 576 times the human rate and “matters because” at over 130 times.

They’re also the weakest signal to lean on. In Graphite’s data, between the two latest versions of each model family 55% to 72% of the tells were new, and the combined rate of nine well-known tells fell 41% to 86% from the earliest to the latest model in each family. People pick the words up as well: a preregistered study found people adopt chatbot vocabulary after brief exposure (Yakura et al.). One “delve” tells you very little. A paragraph where three or four stock phrases cluster tells you more.

Em dashes show how fast this moves. Graphite measured the newest GPT model using them 88% below the old human rate. Some readers still treat them as a sign, so follow your own habit and don’t build a check around them.

People who use AI tools heavily for their own writing are the strongest judges measured so far. In Russell, Karpinska and Iyyer (ACL 2025), a majority vote of five such readers misclassified only 1 of 300 articles, including articles that had been paraphrased or run through a humanizer. Their written explanations leaned on vocabulary, and also on formality, originality and clarity: the same mix of absence and habit described above.

Platforms now give readers a button for it. LinkedIn added a “seems like AI slop” report option on July 30, 2026 (TechCrunch). LinkedIn later said content its own classifiers label as slop gets about 40% fewer views; the company told Moneywise that figure comes from the classifiers, not from the button. On Substack, readers have been able to run a Pangram AI scan on posts longer than 100 words since July 21, 2026 (Mashable).

• SlopShape is a single-author preprint, not peer reviewed, from a company that sells an AI-content checker.
• Its human posts are company blog posts from 2008 to 2022. Its AI posts are single drafts from a 120-word brief with no editing. Nobody has tested these features on LinkedIn or X posts, on human writing from after 2023, or on drafts a person wrote with AI help.
• Between 12% and 52% of human posts show each “AI” habit, and the structural classifier called about 7% of human test posts AI.
• Version 3 of the paper (September 28, 2026) withdrew eleven features that were reacting to formatting, including the most quoted tip, that AI titles “promise the payoff”. That tip no longer has support from this study.

The table describes the default template. Depart from it on purpose, where it helps your reader.

Open the last post your team published and read its first and last paragraphs side by side. Then run the checklist and count what it fails. Fix the close first: it’s the cheapest change, and at 77% against 12% it’s the habit with the widest gap in the data.