How to Check If Text Is AI Generated When No Detector Can Prove It

No tool can prove whether text is AI generated. The reliable way to check if text is AI generated is to combine three kinds of evidence: the document's edit history, a check of every fact and citation, and the sentence-level results of two different detectors, such as GPTZero, Copyleaks or CrawlTide's free AI Content Detector. Treat the result as a reason to talk, not a verdict.

Say a detector reports a freelancer's draft as 87 percent AI, and the writer insists every word is theirs. Which do you believe? Detectors produce a probability from writing patterns, and both careful human writing and lightly edited AI text land on the wrong side of that line often enough to matter. The rest of this guide covers why, the evidence that holds up better, and how to read a detector result without overreading it.

Can you prove text was AI generated?

No. You can only collect evidence of different strengths and weigh it together. This table ranks the kinds you can actually get.

Evidence	What it shows	How much weight it carries	Limits
Edit or version history	Whether a person built the text over time	Strong, in both directions	Not every writer drafts in a tool that keeps history
Fact and citation check	Invented sources, wrong dates, quotes nobody said	Strong for quality, moderate for authorship	People make these mistakes too
Close reading for patterns	Stock transitions, same-length sentences, generic wording	Weak alone, useful in combination	Many people write this way
Detector score	How predictable or pattern-heavy the text is	Weak alone; read the flagged sentences, not the percentage	Misflags non-native writers; misses lightly edited AI text
Watermark check	Whether a known watermark is present	Strong when found	Google's public SynthID Detector lists images, video and audio, not text

To see how far a detector score can be trusted on known samples, look at the held-out test results for CrawlTide's AI Content Detector ("How did we test it?" on the tool page). They are covered below.

How accurate are AI detectors?

Less accurate than their precise percentages suggest. Many detectors lean on perplexity: how predictable each next word is to a language model. Text that a model finds very predictable scores as more machine-like. The trouble is that plenty of human writing is predictable too, especially from people writing carefully in a second language.

Even the company behind ChatGPT could not make this work reliably. OpenAI's own AI classifier, launched on January 31, 2023, correctly identified 26 percent of AI-written text as "likely AI-written" on its challenge set and wrongly labeled human-written text as AI-written 9 percent of the time. OpenAI said it was very unreliable on texts under 1,000 characters, and on July 20, 2023 withdrew it, citing its low rate of accuracy.

Do AI detectors flag non-native English writers?

Yes, at high rates in the best-known test. Liang and colleagues at Stanford tested this in GPT detectors are biased against non-native English writers, published in Patterns in July 2023. They ran seven widely used detectors on 91 TOEFL essays and 88 essays by U.S. eighth graders. The detectors classified the eighth-grade essays accurately. On the TOEFL essays they labeled more than half as AI-generated, with an average false-positive rate of 61.3 percent, and at least one detector flagged 97.8 percent of them (89 of 91).

Our own sample points the other way for one narrow case, and the size of that sample matters. CrawlTide's detector counts visible patterns rather than predictability, and its benchmark included six texts by writers for whom English was a second language (Joseph Conrad and Jacob Riis). All six came back with no AI patterns. Six published, edited literary texts say nothing about how any tool treats a student essay, so read it as a sign that the method matters, not as a clean bill of health.

Can lightly edited AI text pass a detector?

Often, yes. In the same Stanford study, essays generated with ChatGPT-3.5 were detected well at first, but after the researchers asked ChatGPT to revise its own text with one prompt ("Elevate the provided text by employing literary language"), detection rates fell to near zero.

Our benchmark shows the same weakness. CrawlTide built a labeled sample of 132 texts (public-domain books and pre-2020 Wikipedia for human writing, answers from two OpenAI models for AI writing, plus AI answers edited lightly by hand), set every threshold on one part, then scored the held-out part once. On that held-out part:

• Human writing: 0 of 46 texts were marked "Many AI patterns" (43 none, 3 some).
• Unedited AI writing: 18 of 24 were marked "Many AI patterns" (4 none, 2 some).
• Lightly edited AI writing: 10 of 10 came back with no AI patterns at all.

Those counts describe one sample, not a general accuracy rate. The lesson is the last row: light editing removes the patterns a detector looks for, which is why a low score never clears a text.

How do you check if text is AI generated? A five-step method

Collect different kinds of evidence in this order. One weak signal proves little; three pointing the same way is worth a conversation.

1. Ask for provenance first. Request the drafts, notes and version history: Google Docs version history or Word's tracked changes will do. A document that arrived as one paste with no history is a stronger signal than any detector percentage, and three days of visible edits is strong counter-evidence.
2. Read for patterns, sentence by sentence. Look for runs of sentences of nearly the same length, the same stock transition opening several sentences, generic words where a specific term belongs, and every paragraph built on an identical structure. One of these means nothing. A page full of them is a pattern to write down.
3. Check every fact and citation. Search the exact title of each cited paper and the exact wording of each quote. Sources that do not exist, dates that are wrong and quotes nobody said are among the most checkable signs of unedited model output, and they are a problem in the text whoever wrote it.
4. Run two different detectors and read the sentence-level output. Ignore the headline percentage and look at which sentences each tool flags. If two tools disagree, that disagreement is your answer: the result is inconclusive. Treat short texts with extra suspicion, given OpenAI's own warning about texts under 1,000 characters.
5. Do not count on a watermark for text. Google opened its SynthID Detector to everyone on October 7, 2026, but it checks whether an image, video or audio file was made with AI from Google or partners such as OpenAI, NVIDIA and Kakao. Text is not among the formats Google lists, so for writing, steps 1 to 4 are what you have.

How should you read a detector result?

It depends on what else you know about the text. Match your situation to one of these cases.

• High AI score, rich edit history: trust the history. Version history shows a person working; a pattern score only shows a style.
• High AI score, no history, citations that do not check out: this is the strongest combined case you will get. It still justifies a conversation, not a verdict.
• Low AI score: this does not clear the text. As the benchmark above shows, light editing can remove every pattern.
• Writer working in a second language: discount the detector score heavily. That is exactly the group the Stanford study found detectors misclassify most.

Does Google penalize AI-generated content?

Not for being AI-generated. In its February 8, 2023 post, Google Search's guidance about AI-generated content, Google wrote that "Appropriate use of AI or automation is not against our guidelines." What breaks its spam policies is using automation, AI included, "with the primary purpose of manipulating ranking in search results." Google's current spam policies name this as scaled content abuse: many pages generated mainly to manipulate rankings rather than help users, however they were produced.

For anyone publishing on the web, that shifts the useful question from "who wrote this?" to "is this any good?" Generic, unsourced, pattern-heavy prose is a quality problem whether a person or a model produced it, and the fixes are the same: specific claims, checked facts, a real point of view. Those same qualities are what make a page worth quoting in AI answers, as covered in how to optimize your website for AI search. Whether AI search engines then cite your pages is a separate measurement, explained in what AI visibility is and how to check it.

What should you do when the stakes are high?

Write the decision down, and never let a detector score be the deciding evidence for grades, contracts or hiring. The Stanford authors put it plainly: "we strongly caution against the use of GPT detectors in evaluative or educational settings, particularly when assessing the work of non-native English speakers."

If you do have to make a call, record what you relied on: the edit history you saw, the citations you could not verify, the detectors you ran and what each flagged. A decision built on documented evidence can be explained and reviewed. A decision built on one percentage cannot.

What an AI detector cannot tell you

A detector cannot tell you who wrote a text, how much of it a person changed, or whether the ideas are original. It cannot see a draft that was written by a person and polished by a model, or the reverse. It reports patterns in the words in front of it, and those patterns overlap between people and models. Use it to find the sentences worth a closer look, then let the other evidence decide.

Use the score to start a conversation, not to end one

The practical routine to check if text is AI generated is short. Ask for history, read closely, verify the facts, and treat any detector as one input among several. When the patterns are there, the most useful outcome is usually a better draft rather than an accusation.

To see exactly which sentences carry the patterns, paste the text into CrawlTide's free AI Content Detector. It explains what it noticed in each flagged sentence, suggests rewrites for the strongest ones, and does not store the text you submit.