How to Tell If Something Was Written by AI

There is no reliable way to prove that text was written by AI just by reading it or running it through a detector. AI detectors can be wrong, including flagging writing that a person wrote. The most dependable approach is to look at several clues together: factual errors or invented sources, generic phrasing, a mismatch with the writer's usual work, and the process behind the writing. Treat any single signal as a reason to look closer, not as proof.

Why there is no foolproof test

AI writing tools are built on large language models, which generate text by predicting likely next words based on patterns learned from enormous amounts of human writing. The result is fluent text that resembles the way people write, because that is what the models learned from.

That creates three problems for anyone trying to tell the difference:

• People and AI can write in similar ways. Clear, tidy, generic prose is not unique to AI. Plenty of people write that way, especially in formal settings.
• Most writing is not purely one or the other. Someone might use AI to outline, rewrite a paragraph or fix grammar, then edit the result. Mixed writing has no clean line to detect.
• The tools keep changing. Many detectors are built around the AI writing of a particular period. New models and simple editing can change the patterns they look for.

Even AI developers say detection is unreliable. OpenAI released its own AI text classifier in January 2023 and withdrew it in July 2023 because of its low accuracy. In OpenAI's tests, it correctly identified only 26% of AI-written text and incorrectly labeled human writing as AI-written 9% of the time. OpenAI's current guidance for educators answers the question "Do AI detectors work?" with: "In short, not in our experience."

It also warns that asking ChatGPT whether it wrote something does not work. ChatGPT has no record of what content was AI-generated and may simply make up an answer.

Signs of AI writing: clues, not proof

These patterns can suggest text was generated by AI. None of them proves it, and each can appear in human writing too.

1. Invented or unverifiable sources. Citations, quotes, statistics or studies that do not exist, or that do not say what the text claims.
2. Confident factual errors. Wrong dates, names or details stated with complete assurance.
3. Generic content that could fit any topic. Smooth paragraphs that make broad points without specific examples, experiences or evidence.
4. Leftover chatbot phrasing. Lines like "As an AI language model" or "Certainly! Here is your essay" that were pasted in by mistake.
5. A mismatch with the writer. A sudden change in vocabulary, tone or skill compared with the person's previous work.
6. Answers to a slightly different question. Content that misses the specific details of the assignment, brief or conversation.
7. Repetitive structure. Every section built the same way, the same transitions used again and again, or a summary that restates what was just said.
8. Experiences that do not add up. First-person stories with details the writer could not have, or that conflict with what you know about them.

The first two clues are the most useful because you can check them. A fabricated source is a verifiable problem, whoever created it. AI tools are known to produce convincing but false information, a problem explained in What Is an AI Hallucination?.

The style-based clues are the weakest. Generic, well-organized prose is exactly what many people are taught to write. Using style alone to decide authorship risks accusing someone who did nothing wrong.

How accurate are AI detectors?

AI detectors are not accurate enough to be used as proof on their own. Independent research has found error rates high enough to wrongly flag human writing, and simple editing or paraphrasing can make AI text go undetected.

A detector can make two kinds of mistakes:

• False positive: labeling human writing as AI-generated. This is the error that leads to unfair accusations.
• False negative: missing text that really was AI-generated.

What the evidence shows:

• Independent testing found no tool reliable. A 2023 peer-reviewed study in the International Journal for Educational Integrity tested 14 detection tools, including Turnitin, and concluded they were "neither accurate nor reliable." All scored below 80% accuracy. When AI text was edited by hand, about half went undetected. When it was paraphrased by software, overall accuracy fell to 26%. The tools were more likely to miss AI text than to wrongly flag human writing.
• Non-native English writers were flagged more often. A 2023 Stanford-led study published in Patterns ran seven detectors on 91 essays written by people for the TOEFL English exam. The detectors misclassified more than half of them as AI-generated, with an average false positive rate of 61%. The same detectors were nearly perfect on essays by U.S. eighth-graders. The sample was small, but the pattern is a serious fairness concern.
• Paraphrasing defeats many kinds of detection. University of Maryland researchers showed that repeatedly rewording AI text with another AI model significantly reduced detection rates across several detector types, including watermark-based ones, often with little loss in quality.
• Vendor-reported rates are lower, with caveats. Turnitin said in June 2023 that for documents in which its detector identified at least 20% AI writing, its document-level false positive rate was under 1%, and its sentence-level false positive rate was around 4%. These are the company's own figures, not independent audits. Turnitin does not display a percentage for scores below 20%, because it says low scores are less reliable, and it states that its AI writing score "should not be used as the sole basis for adverse actions against a student."

Even a low error rate adds up. When Vanderbilt University disabled Turnitin's AI detector in August 2023, it noted that it had submitted 75,000 papers in 2022. At a 1% false positive rate, around 750 student papers could have been wrongly labeled as partly AI-written.

Why edited AI text is harder to identify

Once a person edits AI-generated text, attributing it becomes much harder, and arguably less meaningful.

People combine their own writing with AI in many ways:

• asking AI for ideas, then writing from scratch
• writing a draft, then asking AI to tighten it
• generating a draft with AI, then rewriting most of it
• using AI only to fix grammar and spelling

Each produces a different blend of human and machine contributions. A detector returns a single score, but the real question is often about the process: what was allowed, what was done and whether it was disclosed.

Watermarks and provenance: a different approach

Instead of guessing from style, some companies mark AI content when it is created.

Text watermarks. Google DeepMind published a text watermarking method called SynthID Text in Nature in October 2024 and reported that it was watermarking responses in its Gemini app. A watermark like this subtly shapes word choices during generation so that a matching detector can later recognize it. It has clear limits: it only identifies text from tools that apply that specific watermark, Google says it is less effective on factual responses, and detection confidence can drop sharply when text is thoroughly rewritten or translated.

Not every AI company watermarks text. OpenAI said in August 2024 that it had developed a text watermarking method but had not released it. It cited how easily the method could be bypassed, for example by rewording text with another AI model, and concern that it could stigmatize non-native English speakers who use AI as a writing aid.

Content Credentials. The C2PA standard attaches provenance information, such as origin and edits, to digital media including images, video and audio. It is mainly used for media rather than text copied from a chatbot. Its own documentation notes that provenance information alone cannot tell you whether content is true, and this kind of metadata can be removed.

Watermarks and provenance help in specific cases, but they are not a universal test for AI writing.

What to do if you suspect AI writing

Whether you are a teacher, editor, manager or reviewer, a fair process matters more than a detection score.

1. Check the facts and sources first. Verify citations, quotes and statistics. Fabricated or misrepresented sources are a concrete problem you can document.
2. Compare with previous work. Look for sharp changes in style, vocabulary or knowledge, while allowing for growth and effort.
3. Look at the writing process. Drafts, notes, outlines and document version history can show how the work developed.
4. Have a conversation. Ask the writer to explain their reasoning, summarize their main points or walk through how they approached the task. Someone who wrote or deeply engaged with the work can usually discuss it.
5. Treat detector scores as one weak input. If you use a detector, never rely on its score alone, and be cautious with short texts and writers whose first language is not English.
6. Check the rules. Was AI use allowed, restricted or banned? Was disclosure required? Many disputes come from unclear expectations.

Why this matters

Getting this wrong has consequences in both directions. A false accusation can damage a student's record, a job applicant's chances or a writer's reputation. Missing undisclosed AI use can undermine trust in schoolwork, journalism and professional work.

A more useful question is often not "Did AI write this?" but "Is this accurate, is it the person's own thinking where that matters, and did they follow the rules for using AI?" Those questions can usually be answered with evidence.

For a grounding in how AI tools work, and why their writing sounds the way it does, see What Is AI? Real Examples You Already Use.