Signs of AI writing: what readers notice and how to fix each one

Before you start

The signs in this lesson come from Wikipedia's Signs of AI writing, a field guide that the site's volunteer editors wrote to find undisclosed AI text on Wikipedia. It was read for this lesson on 6 November 2026. The lesson picks eight signs that carry over to other text. They are not the guide's own list of the most common signs on Wikipedia as of October 2026, which names Markdown, AI edit summaries and hedging disclaimers about sources, among others.

Its authors set three limits. Some signs may not apply outside Wikipedia. Not all text with these signs is AI-generated, because the models learned from human writing. And the list is descriptive, which means it records what editors observe and sets no rules for how to write.

You need a draft written with AI help, such as one from the first two lessons, and a chat app. The aim is to read your own text as an editor would, then put the fact back where a sentence skipped it.

Why AI text sounds the same

A language model, the kind of program behind ChatGPT, guesses which word should come next from a large body of training text. The guide says the result tends to regress to the mean, which means it lands on the most likely wording, the wording that fits the most cases. A later round of training on human feedback then steers the output toward prose that most readers find agreeable.

The effect shows in facts. The models learned from web text where famous people get positive, important-sounding praise, so they drop rare specific facts and write general praise in their place. The guide's example is a phrase like "inventor of the first train-coupling device", which can turn into a revolutionary titan of industry.

Eight signs and a fix for each

The signs below are from the guide, and the fixes are plain ones. One sign proves nothing. A text that carries several of them, again and again, is the one a reader starts to doubt.

The phrases in the middle column are the guide's own examples to watch for, and they sit in code marks so that you can search your draft for them.

A worked example: keep the fact, drop the significance

The humanizer skill, an open-source set of instructions that tells a chat app how to remove these signs, includes a before and after pair. Its first text condenses a September 2024 revision of Wikipedia's article on the Statistical Institute of Catalonia, which the guide shows as a sample of puffed-up significance.

Before:
The Statistical Institute of Catalonia was officially established in 1989, marking a pivotal moment in the evolution of regional statistics in Spain. This initiative was part of a broader movement across Spain to decentralize administrative functions and enhance regional governance.

After:
The Statistical Institute of Catalonia was established in 1989, part of a wider decentralization of administrative functions in Spain.

• The year and the institute stay, because they are the facts. The skill's rule for this sign is "Keep the fact and drop the significance."
• The claim that the founding marked a pivotal moment goes, and so does officially.
• part of a broader movement becomes part of a wider decentralization, which says what the movement was.
• The last clause, about enhancing regional governance, goes too. It states a hope, and the text names no event for it.
• The skill's next pair does the same to a stock section. Two sentences about the area of Korattur, built on Despite these challenges, become Korattur has recurring traffic congestion and water shortages.

Fix your own draft in four steps

The skill works in four steps: mark the tells, draft the rewrite, check the draft and write the final version. Its rule for a rewrite is to keep every supported claim and to add no fact, name, number, date, quote or citation that did not come from the source or from you.

The prompt below covers the first two steps. It is ours, built on the skill's steps, and the sample line applies the skill's rule that a writing sample is read first and overrides its patterns.

Step 1. Mark every phrase in the text below that matches one of these signs: a crowd of stock words, claims of significance, "serves as" in place of "is", a trailing -ing comment, vague authorities, not-only-but-also pairs, a closing challenges paragraph, leftovers from the chat. List each one with its sentence. Do not rewrite yet.

Step 2. After I reply, rewrite. Keep every claim the text supports. Add no fact, name, number, date, quote or citation. If a sentence has no fact left, delete it. Match my sample's sentence length, word choice and punctuation.

My sample: [paste three paragraphs you wrote without AI]
The text: [paste your draft]

1. Run step 1 and read the list. Strike any phrase you want to keep, because a person can use any of these on purpose.
2. Reply with the numbers to rewrite, then run step 2.
3. Compare the rewrite with your draft. Treat any new fact, name, number or date as an error, as the skill treats an unsupported addition.
4. Read the final text aloud, and vary sentence length wherever every sentence runs the same length.

The words change with every model

The guide splits stock vocabulary into periods and says these are not hard cutoffs, only a rough idea of how earlier and later output reads. It says delve was famously overused in 2023 and early 2024, then became less frequent later in 2024 and dropped off sharply in 2025. The humanizer skill draws the practical point: word habits change with each model release, while the structural habits persist, so fix the structure before you chase single words.

Punctuation moves too. The guide files overuse of the em dash under historical indicators, which it describes as "much less frequent in newer models". This sign is most useful in combination with other indicators and not by itself. Some AI companies have tried to make newer chatbots suppress it, most notably OpenAI's GPT-5.1, and the latest models, from September 2026 on, have more and more often been found to avoid em dashes entirely.

Readers are poor judges

The guide says humans are "notoriously bad" at telling human and AI text apart. It cites a 2025 study that found human ability no better than random chance. A second 2025 study, on German theses, found that readers recognised 57% of the AI texts and 64% of the human ones. In a 2025 preprint, a group of expert annotators did better, and participants who rarely used language models did only slightly better than chance.

What is not a sign

The guide lists indicators it calls ineffective, because they show up in human writing too. Leave these alone when you edit, and do not hold them against another writer.

• Perfect grammar. The guide says many editors are skilled writers or come from professional writing backgrounds.
• Formal or fancy prose. The guide says the pattern is specific words and phrases, and does not extend to all formal prose.
• A mix of casual and formal language. It can point to a person in a technical field, a young writer, or several authors on one page.
• Transition words on their own. Only a few are known to be overused, and essay writers use them too.
• From the guide's section on signs of human writing: human-written Wikipedia articles use short there is a and it has a sentences more often than AI text does. The same goes for plain verbs such as wrote instead of authored. Keep them when an edit offers to smooth them out.

Fix the cause as well as the sign

The guide tells its own readers that the signs are potential signs of a problem and not the problem itself, and that the deeper problems are the ones to address. In your text, the deeper problem is a sentence that holds no fact, or a fact you have not checked.

Two checks go past wording. Compare every name, date and figure with your notes, because a smooth sentence can be wrong. And say where AI helped when your reader needs to know. Anthropic's AI Fluency cheat sheet defines Transparency Diligence as "Being honest about AI's role in your work with everyone who needs to know". It defines Deployment Diligence as "Taking responsibility for verifying and vouching for the outputs you use or share".

Common mistakes

• Swapping single words and leaving the sentence as it was. It still says nothing, and a reader sees the same hollow sentence.
• Treating one sign as proof. The guide says even the em dash is best read together with other indicators.
• Accusing a colleague or a student on the strength of these signs. The guide warns that false accusations of AI use can drive away new editors and spread suspicion.
• Sanding the signs off and stopping there. A text without its tells can still hold claims nobody has checked.
• Fixing only the words on last year's list. Word habits change with each model, and the structure persists.

Try it yourself

• Take a draft from lesson 1 or 2 and run step 1 of the prompt. Do not rewrite yet.
• Count the sentences that hold a name, a date or a number, and mark those that hold none.
• Rewrite five of the marked sentences by hand, each with one fact from your notes, and delete any that have no fact to add.
• Run step 2 with a sample of your own writing, compare the result with your hand rewrite, and keep whichever sounds like you.

Next: do AI detectors work

Spotting signs by eye is one thing, and a detector program is another. Do AI detectors work? covers OpenAI's withdrawn classifier, the studies on false positives and watermarking.

Frequently asked questions

Sources

1. Wikipedia: Signs of AI writing (read November 6, 2026)https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing
2. GitHub: the humanizer skill (blader/humanizer) (read November 6, 2026)https://github.com/blader/humanizer/blob/main/SKILL.md
3. Anthropic: AI Fluency key terminology cheat sheet (read November 6, 2026)https://www-cdn.anthropic.com/4396730ed190e691a3712cf2fd6bfe35509deca2.pdf