AI Writing Detector Scores and What They Really Tell You
A freelancer sends you a blog post, you paste it into a free checker, and it comes back "87% likely AI generated".

A freelancer sends you a blog post, you paste it into a free checker, and it comes back "87% likely AI generated". Now you have to decide whether to pay the invoice, accuse someone of cheating, or ignore a number you do not fully understand. An AI writing detector gives you a score that looks precise, and that precision is exactly what makes it risky to act on.
This guide explains what these tools measure, why the same paragraph can score 12% on one site and 90% on another, and a fairer way to check work when you have doubts.
What an AI Writing Detector Actually Measures
Most detectors do not compare your text against a database of chatbot output. They estimate how predictable the text is. A language model reads each word and asks how likely it was given the words before it. Text full of highly probable word choices scores as "machine-like". Text with surprising turns scores as "human".
Many tools also look at variation between sentences, sometimes called burstiness. People tend to mix a four-word sentence with a forty-word one. Chatbots, left to their defaults, produce steadier rhythm. A detector combines these signals into a single probability and shows it as a percentage.
That percentage is not "the share of this text written by AI". It is the tool's confidence in a guess, based on statistical patterns. A score of 87% does not mean 87% of the sentences came from a chatbot. It means the model thinks the text resembles its idea of AI writing.
Why Scores Disagree So Much
Run the same 400-word sample through three popular checkers and you can easily get three very different answers. There are a few reasons for that:
- Different reference models. Each tool judges predictability using its own model, trained on its own data, so they disagree about what "likely" means.
- Different thresholds. One vendor flags anything above 50%, another waits for 80%. The underlying numbers are not on the same scale.
- Short text. Below about 250 words, scores swing wildly. A single paragraph is close to a coin toss.
- Editing. Light human edits to AI text, or light AI polishing of human text, both push scores toward the middle where they mean very little.
Formal and formulaic writing is the bigger problem. Legal summaries, product descriptions, instructions and anything written by someone working in their second language all tend to use common, predictable phrasing. Those are exactly the texts most likely to be flagged even when a person wrote every word.
The Common Advice to "Just Run It Through a Detector" Is Wrong
Plenty of small business owners, teachers and editors have been told that a detector is a quick, objective first check. I disagree. A tool that is wrong one time in ten sounds fine until you remember that a manager screening 30 freelancer pieces a month will then wrongly flag about three of them every month. Over a year, that is dozens of false accusations against people who did the work.
Even vendors with good published numbers usually test on clean samples: pure chatbot text against pure human text. Real submissions are messier. A writer might use a chatbot to suggest a headline, a spell checker to fix grammar and their own brain for everything else. No detector can sort that out, and none claims to.
There is a relationship cost too. A good freelancer who gets accused once, on the basis of a number, will often stop working with you. Replacing a reliable writer can take weeks of trial pieces and edits.
The more useful question is not "was AI involved?" but "is this work accurate, original and what I paid for?" That question you can answer without a percentage.
A Fairer Way to Check Work You Are Unsure About
If something feels off, these steps take about 15 minutes and give you evidence a score never will:
- Check the facts. Chatbot drafts often include invented statistics, quotes or product features. Pick three specific claims and verify them. Made-up details matter regardless of who wrote them.
- Look for specifics only a person would know. Does the piece mention your actual customers, local details or the brief you gave? Generic text that could fit any company is a quality problem first.
- Ask for drafts or notes. Most writers keep outlines or version history in Google Docs or Word. A two-minute look at revision history says more than any detector.
- Have a conversation. Ask the writer to talk you through one section. People who wrote something can explain their choices easily.
- Set clear rules up front. If you do not mind AI help with outlines but want final copy written by a person, say so in the brief. Most disputes come from expectations nobody wrote down.
If you are writing those rules for the first time, it helps to think about what tools your team already uses and which ones can read your documents. This guide to AI browser extension permissions is a good place to start, since many writing helpers run as extensions.
When a Detector Score Is Still Useful
I am not saying detectors are worthless. Used as one weak signal among several, they can point you toward pieces worth a closer read. A publisher handling 500 guest submissions a week might use scores to decide which ones an editor reviews first. That is triage, not judgement.
A few habits keep the risk low. Only test texts longer than 300 words. Use the same tool every time so at least the scale stays consistent. Treat anything between 30% and 70% as "unknown" rather than a lean in either direction. And never let a score be the only reason for rejecting work, refusing payment or giving a student a lower grade.
If you are choosing which AI tools to bring into a small business at all, start with the job you need done rather than the tool's marketing, the same way you would when choosing the right chatbot platform for a small business. You can find more practical comparisons in our AI tools section.
FAQ
Can an AI writing detector prove someone used ChatGPT?
No. A detector produces a statistical guess, not proof. Even the vendors behind these tools advise against using their scores as the sole basis for decisions about students or employees.
Why did my own writing get flagged as AI?
Clear, formal and predictable writing often scores as machine-like. Non-native English writers, technical writers and people following a strict template are flagged more often, even when every word is their own.
Do detectors work on other languages?
Usually worse than in English. Most tools are trained mainly on English text, so scores for Spanish, German or Indonesian content are less reliable and should be treated with extra caution.
Is there a score that means a text is definitely human?
No score guarantees it. Light editing can push chatbot text to a very low score, and a careful human writer can score high. Low and high numbers both deserve a second look before you act.
More in Automation Tools
Automation Tools
Automation Tool Pricing: Tasks vs Operations Explained
A reader wrote to me in September because she had signed up for a plan with 750 tasks a month, built two small workflows, and...
Automation Tools
Choosing a No Code Database for Small Team Automations
Every automation I have helped someone build eventually hits the same wall.
Automation Tools
Scheduling Link Tools and How to Pick One Without Overpaying
The email thread started on Monday. "Does Thursday work?" "Thursday is tight, how about Friday morning?" "Friday I can do after...