You ordered website texts and, to be sure, ran them through an online detector. It said: “90% written by AI”. The first reaction is understandable: I was cheated, I won’t pay. Don’t rush to a conclusion. A detector’s percentage looks like a measurement, but in reality it is a program’s guess, and it can be wrong in both directions.
Below we cover what is known about the accuracy of such detectors, what Google says about AI-written text, and what a business owner can check alone in five minutes without any software. We checked the sources on 5 November 2026. Some are older than we would like, so we give the year next to each fact.
1. Even OpenAI shut down its own detector
OpenAI, the company behind ChatGPT, released its own “AI classifier” (a tool meant to recognise AI-written text) in January 2023. It also shut it down itself: on 20 July 2023, in its words, “due to its low rate of accuracy”.
The figures from its own test are these. On a “challenge set” of texts, the classifier correctly labelled only 26% of genuinely AI-written texts as “likely AI-written”. So it missed three of every four machine texts. And it wrongly labelled 9% of human-written texts as AI-written. The OpenAI page returns an error from our server, so we checked these figures against reports by Decrypt and PCWorld (July 2023), not the primary source. The test used English texts and was run in 2023: it says nothing about other languages or today’s detectors.
What it means for you. It is not proof that every detector is bad: today’s may be more accurate, or may not. But it is an honest signal: if the maker of the best-known model could not build a reliable detector, a percentage from a free service cannot be grounds to refuse payment. We also leave out the 2022 test of five detectors that circulates online: it used free versions that have long since changed.
2. Google looks at usefulness, not the author
On 8 February 2023, Google published an explainer in Search Central, “Google Search’s guidance about AI-generated content”. The key points:
- appropriate use of AI or automation is not against Google’s guidelines (“Appropriate use of AI or automation is not against our guidelines”);
- spam is automation used “with the primary purpose of manipulating ranking in search results”, that is, mass production of pages with no value;
- Google’s systems reward original, high-quality content made primarily for people, not for search engines.
So “Google punishes AI” is not quite accurate. What gets punished is not AI but pages nobody read or checked: identical, fact-free, made only for rankings. Google’s later guidance on generative AI on websites, and why such text must be checked by hand, is covered in our article “AI writes your website text? What Google says to check by hand”. It also has a section on detectors and the “percentage of edits”.
3. What to check yourself in five minutes
Your yardstick is simple: does the text contain what only someone who knows your business could write? Open the text and go through four points.
Prices and terms. Are there specific numbers: “from $30”, “in 2 days”, “12-month warranty”? Do they match your price list? Vague “affordable prices” and “fast” are worth nothing.
Examples from your work. Does it say what you actually do, for whom, in which city, with which materials? If the text could sit on any competitor’s website, it is fluff.
The “cover the name” test. Cover the logo and company name. If it is still unclear how you differ from others, the text is weak, even if an experienced person wrote it.
Facts. Pick three statements that can be checked and check them: addresses, delivery times, return conditions. AI, like a careless copywriter, happily invents what you do not have.
An imaginary example. A text about washing machine repair says: “Our experienced technicians will repair your appliance quickly and well at affordable prices”. A human or an AI could have written that sentence, and it is equally empty. But “A technician visits the same day you call, diagnosis $10, deducted from the repair cost” is already your text, and it is easy to check against your price list. The difference is visible without any detector.
If even one point fails, that is a reason to send the text back for revision. And it does not matter what the detector said: a detector easily confuses a bad human text with a good AI text.
4. How to talk to the copywriter
We wrote separately about paying for results instead of characters: “Copywriters get paid per character, you pay for fluff”. How to pick AI text for ads is covered in “80% of What AI Writes Is Better Thrown Away”. Here, only the “did you write this yourself?” dispute.
- Do not start with an accusation based on a percentage. Show specific places: “the prices here are not ours, this example is made up”.
- Ask where the numbers and facts came from and ask for a source. A person who wrote it themselves will answer at once.
- Put the requirements in the brief: your prices, terms, examples, fact-checking. This protects you whether or not the contractor uses AI.
- If it matters to you that texts are written by a human without AI, agree this in the contract, in advance. But remember: you will not be able to prove it with a detector.
From our practice
In our practice, a detector is useful only as a reason to read the text more carefully: if it shows a high percentage, we look for fluff and repeated phrases. But we decide on payment or replacement not by it but by the checklist above. We have not measured detector accuracy ourselves, so we give no figures like “our detector is wrong in N% of cases”.
And a disagreement, which we will show openly. Some practitioners advise running text through a detector and editing until the percentage drops. Others consider that pointless: a detector measures style, not usefulness. Google’s 2023 explainer has no official position on such tools: it speaks of quality and usefulness, not of checking services. We lean to the second view, but that is our position, not a rule.
Summary
A detector’s percentage is neither proof that a text is bad nor proof that AI wrote it. OpenAI shut down its own detector because it missed three quarters of machine text and wrongly “caught” humans. Google does not punish AI as such; it looks at whether the text is useful to a person. So look at the text:
- Are your prices and terms there, and are they correct?
- Are there examples from your work?
- Is it clear how you differ if you cover the name?
- Have the facts been checked?
If not, it is fluff, whoever wrote it. The Google and OpenAI figures in this article are from 2023 and may differ for modern tools. We did not find a recent independent test we could trust, so we do not cite one.
For how we run advertising, see search advertising and Performance Max.
Want a free audit of your Google Ads account? Send us a request — we will look at your account and tell you where money is being wasted.