How to Tell If Writing Is AI Generated Without a Detector

Signs Something Was Written by AI

No single test proves that artificial intelligence (AI) wrote a text. Look for clusters of recurring patterns, check drafts and version history, and treat detector scores as weak hints. The aim is to spot AI writing without being fooled by one sign or a faulty detector.

Wikipedia's Signs of AI writing page, a WikiProject AI Cleanup page, lists the patterns below. It calls them "observations, not rules," and warns that "Not all text featuring these indicators is AI-generated." The reason is that large language models (LLMs) are trained on human writing.

The documented patterns fall into these groups:

• Puffed-up significance: phrases such as "stands as," "is a testament," "pivotal," "setting the stage for," and "evolving landscape."
• Promotional wording: "boasts a," "vibrant," "rich," "nestled," "in the heart of," and "groundbreaking."
• Vague attributions: "Industry reports," "Observers have cited," and "Experts argue," none of which names a source.
• Trailing -ing phrases: add-ons such as "highlighting," "underscoring," and "emphasizing" that tack shallow analysis onto a sentence.
• Overused vocabulary: "delve," "crucial," "tapestry," "testament," "underscore," "showcase," "intricate," and "meticulous." The page says this vocabulary changes between model eras.
• Negative parallelisms and the rule of three: lines such as "Not just X, but also Y," and lists built in threes.
• Formatting habits: title case headings, heavy boldface, bulleted lists with bold inline headers, emoji used as formatting, and curly quotation marks.
• Chatbot phrasing left in by mistake: "I hope this helps," "Certainly!", "Would you like...", and "let me know."
• Knowledge-cutoff disclaimers: "as of my last knowledge update." The page ties this tell to 2022 and 2023 models. Refusals such as "as an AI language model" have become increasingly rare.
• Citation problems: broken links, invalid digital object identifiers (DOIs) or International Standard Book Numbers (ISBNs), DOIs that point to unrelated articles, and leftover tokens such as "oaicite" or "contentReference."
• Em dashes: LLMs use them more than nonprofessional human writers, often with spaces around them. The page calls this sign "most useful when taken in combination with other indicators." Some AI companies have tuned newer chatbots, including OpenAI's GPT-5.1, to suppress em dashes. The latest models, from September 2026 onward, are increasingly found not to use them at all.

The page also says LLM text tends to "regress to the mean." Specific, unusual facts get swapped for generic, positive, important-sounding wording. The subject then reads as both less specific and more exaggerated. For more on AI writing, see our guide to AI slop.

Why Gut Feel Fails

Wikipedia's page warns, "Do not rely too much on your own judgment." It cites a 2025 study in Advances in Simulation by Cheng et al. that found people could tell LLM text from human text "no better than random chance."

A separate 2025 preprint found that a subset of expert annotators who use LLMs heavily identified 90% of AI-generated text and misflagged 3.3% of human text. The page notes that this research "is still limited."

Neither a gut feeling nor a detector score settles the question alone.

Check Drafts and Version History

Drafts, edit history, and a conversation with the writer give you more to work with than a score.

If the text was written in Google Docs, its version history can be checked. Start with these steps.

1. Open the version history by clicking Last edit at the top right of the document.
2. Name the versions that matter. Click More, then choose Name this version.
3. Note the limit. Google Docs allows up to 40 named versions in one document.
4. Check for Show editors. Google says some Google Workspace business and education customers can select text, right-click, and choose Show editors to see who edited that part.
5. Ask the writer to walk through a passage and explain the choices in their own words.

How Accurate Are AI Detectors?

Turnitin reports a higher incidence of false positives when its tool detects less than 20% AI writing. It added an asterisk to those scores "to call attention to the fact that the score is less reliable." It also raised its minimum length for detection from 150 to 300 words, saying accuracy rises with more text.

Vanderbilt also said Turnitin gives "no detailed information as to how it determines" AI writing. It said the tool was enabled with less than 24 hours' notice. Its reasons included a concern that detectors are more likely to flag non-native English writers.

The Liang et al. study found detectors flagged non-native English writers at high rates. When ChatGPT enriched the vocabulary of the TOEFL essays, the average false positive rate fell from 61.22% to 11.77%. In the same study, simpler word choices raised the false positive rate for American essays from 5.19% to 56.65%.

The study dates from 2023, so it does not test today's detectors. Its authors note that several of the detectors used GPT-2 as a backbone. They "strongly caution against the use of GPT detectors in evaluative or educational settings, particularly when assessing the work of non-native English speakers."

OpenAI said its classifier was "unreliable on text with fewer than 1,000 characters." OpenAI pulled the tool in July 2023, citing "its low rate of accuracy."

Is GPTZero Accurate?

GPTZero's homepage claims 99% accuracy, and that figure is GPTZero's own claim. Its website FAQ says GPTZero detected 95.7% of AI texts on the RAID benchmark while flagging 1% of human texts as AI. The same FAQ gives 96.5% on mixed documents.

GPTZero's October 2025 accuracy post says GPTZero ran that evaluation itself. The post is not an independent audit.

An independent 2025 preprint by Dik, Erdem and Dik tested GPTZero. It tested 28 AI-generated essays and 50 human-written essays.

Most AI essays were detected, at 91 to 100% AI probability. Results on human essays "fluctuated," with a handful of false positives. The authors conclude that GPTZero's "reliability in distinguishing human-authored texts is limited."

Wikipedia's page says detectors such as GPTZero "perform better than random chance" but "have non-trivial error rates." It warns, "Do not solely rely" on them.

Watermarks for Text

A watermark helps only when the model's developer added one and offers a checker. SynthID Text, from Google DeepMind, adjusts token probabilities while the text is generated.

Google says SynthID Text is "less effective on factual responses." Detection confidence can drop when text is "thoroughly rewritten" or translated. Google also says the tool is "not designed to directly stop motivated adversaries."

Google DeepMind's 2024 announcement says the watermark holds up to "cropping pieces of text, modifying a few words and mild paraphrasing." It works best on longer, varied responses.

OpenAI said in August 2024 that it had built a text watermark that is "highly accurate" against localized tampering such as paraphrasing. It said the method is "less robust against globalized tampering," such as translation.

OpenAI also warned that text watermarking could "stigmatize use of AI as a useful writing tool for non-native English speakers." As of that statement, OpenAI had not released the tool.

Google said in May 2025 that text detection would roll out later in its SynthID Detector. On 7 October 2026 Google opened that detector to everyone in English. It checks image, video and audio files made with AI from Google or its partners, and the announcement does not list text.

How to Prove You Didn't Use AI

A detector score alone is weak evidence in either direction. A written record of how the text was made gives a reader more to check.

1. Keep drafts and document history from the start. In Google Docs, click Last edit at the top right to open the version history.
2. Name the versions that matter, using More, then Name this version. Google Docs allows up to 40 named versions.
3. Check whether Google Docs can show who changed each part. Some Google Workspace business and education customers can use Show editors.
4. Be ready to explain the text in your own words, passage by passage.

Non-native English writers face a specific risk here, so keep drafts and version history as a record. The Liang et al. study found detectors flagged TOEFL essays at high rates. Vanderbilt cited the same risk of false positives for non-native writers.

Who This Is Not For

Anyone who needs a detector score to settle a grade or a hiring decision should not rely on one alone. The published record does not support treating a score as proof. If that is your situation, start with the process steps above and talk with the writer. Our guide to whether AI is banned in schools covers the school side of the question. Readers checking pictures rather than text should read our guide to how to tell if an image is AI generated.

Sources

• Wikipedia: Signs of AI writing
• Liang et al., GPT detectors are biased against non-native English writers
• Decrypt, OpenAI quietly shutters its AI detection tool
• Vanderbilt University, Guidance on AI Detection and Why We're Disabling Turnitin's AI Detector
• Turnitin, AI writing detection update
• Turnitin, false positive rate for sentences
• GPTZero, accuracy stats
• Dik, Erdem and Dik, Assessing GPTZero's Accuracy in Identifying AI vs. Human-Written Essays
• Weber-Wulff et al., Testing of Detection Tools for AI-Generated Text
• Dugan et al., RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors
• Google Docs Editors Help, Check activity and file versions
• Google AI for Developers, SynthID
• Google DeepMind, Watermarking AI-generated text and video with SynthID
• TechCrunch, OpenAI on its approach to detecting ChatGPT writing