How to Tell if an Article Was Written by AI (and Why Word Lists Don't Work)

Updated October 15, 2026. 8 min read

Search for how to spot AI writing and you will find the same advice everywhere: look for words like "delve" and "tapestry", count the em dashes, notice the tidy rhythm. We measured those signs on real websites while building our own checker. Most of them did not separate AI from human writing at all. Here is what we found, and what works better.

What we measured

While calibrating Bot or Built, we analysed article text from three groups of websites:

• 50 long-running human sites: personal blogs, food blogs, newsrooms, documentation and company sites with years of human-written archives.
• AI content farms that are still online after being named in investigations by NewsGuard and Nieman Lab.
• Sites generated by AI website builders such as Durable and Lovable, where the first draft of the copy is written by a model.

For each site we counted three things: how often a list of about 80 "AI-favourite" words and phrases appeared, how many em dashes appeared per 1,000 words, and how much sentence length varied. It is a small sample, so read the numbers as a reality check rather than a scientific study.

Finding 1: word lists only catch unedited marketing copy

On the human sites, the AI-favourite words appeared between 0 and about 2 times per 1,000 words. On the content farms, the range was almost exactly the same: 0 to about 2.5. A word-list detector would have called nearly every farm human.

The builder-generated sites were different. Several had 6 to 10 such words per 1,000, three to five times the top of the human range, with words like "crucial", "robust" and "game-changer" doing a lot of work.

The pattern makes sense. Builder copy is often a model's first draft about a generic business, published without edits. Content farms are optimized operations that have learned to avoid the famous words, either by editing them out or by prompting the model to skip them. So word lists can flag lazy, unedited AI marketing copy. They are close to useless against anyone who is trying not to be caught.

There is a second problem: human marketing writers use many of these words too. "Seamless", "robust" and "cutting-edge" were marketing clichés long before chatbots existed.

Finding 2: em dashes are a writer's habit, not a model's signature

Several human-written sites in our set used more than 8 em dashes per 1,000 words, and one well-known essay blog used more than 15, as many as the most dash-heavy AI builder pages we found. The content farms sat in between. The ranges overlap so much that an em-dash count can't tell you anything about a single page.

Finding 3: sentence rhythm overlaps too

Language models do tend to write sentences of similar length, while many people mix short and long ones. We measured this as the variation in sentence length. Human sites ranged widely, from quite even to very varied. The content farms clustered toward the even end, but inside the human range. A newsroom with a strict house style looks just as "even" as a farm.

Why AI text detectors struggle as well

Dedicated AI text detectors use more sophisticated statistics than word counts, but they face the same underlying problem: good human writing and good machine writing are statistically close. Two facts are worth knowing before you trust any detector score:

• OpenAI, the company behind ChatGPT, built its own AI text classifier and withdrew it in July 2023 because of its low accuracy. On its own test set, it identified only 26% of AI-written text correctly and wrongly flagged 9% of human text.
• A Stanford study, "GPT detectors are biased against non-native English writers" (Liang et al., 2023), found that popular detectors consistently misclassified essays by non-native English speakers as AI-generated, and that simple prompting could get AI text past them.

A detector score is a probability with a real error rate. It should start a conversation, never end one, and it should never be the only basis for accusing a person of using AI.

What works better: read like an editor

The signals below are harder to fake because they depend on knowledge and accountability, not on wording.

Specific, first-hand detail

Human writers who know a subject include things only someone who was there would know: the part that broke, the price last spring, the name of the person who helped, the mistake they made the first time. Generated text describes what usually happens. Ask of every paragraph: could this have been written by someone who never did the thing?

Claims you can check

Pick three specific claims, quotes or statistics and look them up. Invented quotes, studies that don't exist and numbers with no source are among the most reliable signs of unsupervised AI writing, because models produce plausible details with full confidence.

Consistency across the page

Does the article contradict itself? Does the date in the intro match the date in the conclusion? Does the product have the same price in two sections? Generated text is fluent paragraph by paragraph but often doesn't hold together across a long piece.

An actual opinion

Many AI articles cover every side evenly and commit to nothing: "it depends on your needs", "both options have their pros and cons". Human experts usually take a position and explain why, sometimes awkwardly. A piece with no point of view isn't proof of AI, but it is a clue.

Leftovers

Phrases such as "as an AI language model", "I cannot fulfil this request" or "here is a rewritten version" still turn up in published articles. Search for them. When they appear, the question is settled.

Who stands behind it

Click the author. A real byline links to other work, a profile or a history. If the author has published 40 articles today on 40 unrelated topics, or doesn't exist outside the site, the article almost certainly wasn't written the way it pretends to be.

Judge the site, not just the sentence

The strongest evidence about a single article often comes from the site around it. How many articles does it publish a week? Did the domain belong to someone else a few years ago? Was the site generated by an AI builder? Those questions are easier to answer than "was this paragraph written by a model?", and the answers are more reliable. Our guide to spotting AI content farms and our 12-sign checklist cover them in detail.

That is also how Bot or Built approaches the problem. It gives little weight to wording, more to publishing patterns and the most to technical fingerprints, and it shows every clue so you can decide for yourself.