AI text detector

Wikipedia’s Signs of AI writing — checked automatically

Wikipedia editors maintain a field guide to spotting chatbot-written text. GPTTrace turns it into an automatic checker: paste text and each sign it finds is highlighted, with a link to the guide’s section.

Free · no sign-up · your file never leaves your device

0 words0 characters

Why Wikipedia’s guide matters

Wikipedia was one of the first places to face a flood of AI-written text: new articles about obscure towns, companies and people, generated in seconds and full of confident, unverifiable claims. Volunteer editors responded by documenting what that text looked like. Their guide is unusual among AI-detection advice because it comes from people who have read and cleaned up thousands of real examples, and because it explains why each pattern appears — usually as a side effect of how models are trained to sound helpful, neutral and complete.

The signs, in brief

Language and tone

Chatbots inflate significance (“stands as a testament to”, “left an indelible mark”), slip into promotional language (“nestled in the heart of”, “boasts a diverse array”), editorialize (“it’s important to note”) and stack connectives (“Moreover… Furthermore… Additionally”).

Style and structure

They end sentences with shallow participle phrases (“…, highlighting the importance of community”), group ideas in threes, build “not only X but also Y” contrasts and avoid plain “is” in favour of “serves as” or “boasts”.

Content

Claims are attributed to unnamed experts and studies, superlatives go unsupported, and sections end with a summary of what was just said.

Communication left in

Text pasted from a chat keeps phrases aimed at the user — “I hope this helps”, “Certainly! Here’s…”, “As of my last knowledge update” — and placeholders such as “[Your Name]”.

Markup and characters

Markdown bold and headings appear where they don’t render; chat-interface citation tokens like “oaicite” and “turn0search0” survive copying; and some tools leave invisible Unicode characters behind. The invisible character detector focuses on that last one.

GPTTrace implements each of these as a pattern with a measured weight, adds statistics the guide describes informally (uniform sentence length, even paragraphs), and highlights every match in your text with a link back to the relevant section of the guide.

How to use the guide well

The guide’s own advice is to treat these signs as reasons to look closer, not as grounds to accuse anyone, and to focus on substance: unsupported claims, invented citations and promotional tone are problems in any text, whoever wrote it. GPTTrace follows the same principle — it highlights the patterns so you can see them in context and decide.

How accurate is it? Our measured numbers

We test GPTTrace on labelled text samples and publish the results, including where it does badly. It is tuned to keep false accusations rare, so it misses some AI content rather than flag real work.

standard text check: AUC 0.795 (cross-validated)

2,467 labelled samples (1,258 AI, 1,209 human), run 2026-10-08. At the “Likely AI” line it caught 34% of AI samples and wrongly flagged 5% of human ones.

SourceTruthSamplesResult at “Likely AI”
claude-opus-5AI1250% caught
gpt-4.1AI12100% caught
gpt-oss-120bAI12100% caught
claude-haiku-4-5AI1283% caught
qwen-3.8-27bAI863% caught
pd-literatureHuman3061% wrongly flagged
human-pmc-eslHuman1835% wrongly flagged
human-wikipedia-pre2022Human1558% wrongly flagged
hc3-open_qaHuman714% wrongly flagged
chatgpt-3.5AI120232% caught
hc3-wiki_csaiHuman5586% wrongly flagged

text check with deep scan: AUC 0.91 (cross-validated)

2,467 labelled samples (1,258 AI, 1,209 human), run 2026-10-08. At the “Likely AI” line it caught 56% of AI samples and wrongly flagged 5% of human ones.

SourceTruthSamplesResult at “Likely AI”
claude-opus-5AI1242% caught
gpt-4.1AI12100% caught
gpt-oss-120bAI12100% caught
claude-haiku-4-5AI12100% caught
qwen-3.8-27bAI875% caught
pd-literatureHuman3060% wrongly flagged
human-pmc-eslHuman1831% wrongly flagged
human-wikipedia-pre2022Human1551% wrongly flagged
hc3-open_qaHuman70% wrongly flagged
chatgpt-3.5AI120254% caught
hc3-wiki_csaiHuman5589% wrongly flagged

Data sources and method: methodology & accuracy.

Frequently asked questions

What is Wikipedia’s “Signs of AI writing” page?
A guide written by Wikipedia editors, part of the WikiProject AI Cleanup effort, that catalogues patterns common in text produced by large language models. It was written to help editors find and fix AI-generated additions to articles, and it has become one of the most cited practical references on the topic.
Does matching a sign mean the text is AI-written?
No, and the guide itself says so. Each sign also appears in human writing. What matters is the density and the combination of signs, which is what GPTTrace’s weighted score measures.
Which signs does GPTTrace implement?
Language and tone (undue significance, promotional wording, editorializing, connective overuse), structure (superficial -ing analyses, rule of three, negative parallelisms, avoiding “is”), content (vague attributions, superlatives, formulaic summaries), technical artifacts (chatbot phrases, disclaimers, chat-interface markup, placeholders, Markdown, title case, emoji) and typography (invisible Unicode, em-dash density, curly punctuation).
Is the em dash really a sign?
It appears in the guide because language models use it far more than most people type it. On its own it is weak — plenty of writers love em dashes — so GPTTrace gives it a small weight based on density.
Do you update the signs when the guide changes?
Yes. The catalogue is reviewed against the guide periodically, and each sign’s weight is refitted on labelled data so that signs that stopped working lose influence.