AI detectors estimate how likely it is that a piece of text was written by AI. They do this by looking for statistical patterns, not by proving who wrote it. They can be useful as one signal, but they make mistakes, including false positives that flag human writing as AI.
If your email, proposal or LinkedIn post ever gets flagged, the answer is not to “beat” the detector. It is to write clearly in your own voice and be able to show your work.
Key Takeaways
- AI detectors give a probability, not proof. Every AI score is an estimate.
- Detection methods include predictability measures (perplexity and burstiness), trained classifiers and watermarks.
- OpenAI shut down its own AI text classifier in July 2023 because of its low rate of accuracy.
- Stanford researchers found that popular detectors flagged more than half of essays by non-native English writers as AI generated.
- Vendor accuracy claims (Copyleaks, for one, claims over 99%) come from the vendor’s own tests. Treat them as claims, not guarantees.
- The best defense is your writing process: drafts, version history and your own words.
What Is an AI Detector?
An AI detector tool scans text and estimates whether it was written by a person, by AI, or by a mix of both. You paste text or upload a document, and the tool returns an AI score or a label such as “likely human” or “likely AI.”
AI detectors are not plagiarism checkers. A plagiarism checker looks for matching text in other sources. An AI detection tool looks at how the text is written. Some vendors offer both plagiarism checks and AI detection in one product, but they answer different questions.
AI detectors also cannot see your screen, your drafts or your chat history. They only see the final text. That is why their results are always an educated guess about AI use, not a record of it.
How Do AI Detectors Work?
Most AI detectors work by comparing your text with what they have learned about human writing and AI generated writing. There are three main approaches.
Perplexity and Burstiness
Large language model tools write by predicting the next likely word. That makes AI generated text tend to be predictable.
- Perplexity measures how predictable the words in a text are to a language model. Low perplexity means the word choice is very expected. Human writing often has more unexpected word combinations, so it tends to score higher.
- Burstiness measures how much that predictability varies across a document. People mix short and long sentences and change sentence structure. AI text is often more even.
GPTZero helped make these two terms popular. Its own article now says it no longer uses perplexity and burstiness for detection and has moved to a deep learning approach. That shift suggests these two signals alone were not enough.
Classifiers Trained on Examples
Many detectors use a classifier. That is a machine learning model trained on large sets of human written content and AI generated samples. It learns the specific characteristics of each and then predicts which group your text looks closer to.
Turnitin, for example, says its model splits a submission into segments, gives each segment a score between 0 and 1, and then combines those scores into one result for the document.
Classifiers depend on their training data. If a new AI model writes in a style the detector has not seen, or if your writing style is unusual for the training data, the result can be wrong. Detectors also have to keep pace with evolving language models.
Watermarks
A watermark is a hidden signal an AI company adds to text while its model generates it. Google’s SynthID Text is one example. Google says it adjusts the model’s word choices in a pattern that a detector can later check. Google has open sourced SynthID Text for developers.
Watermarks have limits too. Google says SynthID Text is less effective on factual responses, and that detector confidence can drop a lot when text is heavily rewritten or translated. A watermark also only helps if the AI tool that wrote the text added one in the first place.
Detection Methods Compared
Treat every method below as a signal, not proof.
| Method | What it looks at | Where it helps | Weak spots |
|---|---|---|---|
| Perplexity and burstiness | How predictable the words are, and how much that varies | Quick signal on long, unedited AI text | Plain, formal or simple human writing can look “too predictable” |
| Trained classifier | Patterns learned from human and AI training data | Used by commercial AI checker tools such as Turnitin | Only as good as its training data; new AI models and edits can confuse it |
| Watermark | A hidden pattern added when the text was generated | Strong signal when the AI provider added it | Only works on text from tools that watermark; heavy rewrites and translation weaken it |
How Accurate Are AI Detectors?
Short answer: less accurate than their marketing suggests, and accuracy depends heavily on the text you test. Here is what the evidence shows.
OpenAI Retired Its Own Detector
In January 2023, OpenAI released a classifier to flag AI written text. On OpenAI’s own English test set, it correctly identified 26% of AI written text as “likely AI written.” It also wrongly labeled human written text as AI written 9% of the time.
OpenAI also said the classifier was very unreliable on short texts under 1,000 characters, and that AI text could be edited to evade it. On July 20, 2023, OpenAI shut it down, citing its low rate of accuracy.
Stanford Research Found Bias Against Non-Native English Writers
A Stanford study, with professor James Zou as senior author, tested seven popular detectors. It was published in the journal Patterns in July 2023. The researchers ran 91 essays written by non-native English speakers for the TOEFL English test through those detectors.
- The detectors flagged 61.22% of those human written essays as AI generated.
- 89 of the 91 essays (97.8%) were flagged by at least one detector.
- The same detectors were near perfect on essays by U.S.-born eighth graders.
The reason ties back to perplexity. The researchers note that non-native speakers tend to score lower on measures like lexical richness and lexical diversity, the range of words they use. That makes their writing look more predictable to a detector. If you write at work in your second language, this matters to you.
What Vendors Claim
Detector companies publish their own numbers. These are vendor claims based on their own testing, not independent proof. The same caution applies to other AI detection tools:
- Turnitin says it aims to keep false positives under 1% for documents with over 20% AI writing. It needs at least 300 words of prose. For scores from 1% to 19%, it shows an asterisk instead of a number, to reduce false positives.
- Copyleaks says it has over 99% accuracy and a 0.03% false positive rate, and that it supports over 30 languages.
- Pangram says it is over 99% accurate, puts its false positive rate at 1 in 10,000, and benchmarks against 26 AI models.
- Scribbr offers free checks of up to 1,200 words per submission with no account. It says no AI model can guarantee 100% accuracy, and its own research found the highest accuracy was 84% for a premium tool and 68% for the best free tool.
Even a small false positive rate adds up. If a detector has a 1% false positive rate and your company scans 1,000 human written proposals, about 10 of them would be wrongly flagged.
Why Results Vary So Much
- Length: Short text gives the detector less to work with. A two-line LinkedIn comment is hard to judge.
- Editing: AI text that a person edits, and human text that an AI tool polishes, both blur the line.
- Style: Formal, plain or template-based writing can look predictable even when a person wrote every word.
- New models: Detectors must keep up with evolving AI technologies. A detector trained on older AI models may miss newer ones, whatever high accuracy it claims.
There is no single most accurate AI detector for every kind of text. Results change with length, language and how much editing happened.
Why Human Writing Gets Flagged as AI
False positives are not random bad luck. They come from the way detection works. Your writing is more likely to be flagged as AI if:
- It follows a tight template, such as a standard follow-up email or a proposal boilerplate.
- It uses safe, common phrases (“I hope this email finds you well,” “Please let me know if you have any questions”).
- You write in English as a second language and keep word choice simple.
- It is very short.
- You ran it through a tool that rewrote whole sentences.
None of that means you used AI. It means your text looks statistically predictable.
Before and After: Generic Phrasing vs. Your Own Voice
The goal here is not to trick a detector. It is to sound like a real person who knows the reader. Specific, personal details do both.
| Before (generic, predictable) | After (specific, your voice) |
|---|---|
| “I hope this email finds you well. I wanted to follow up on our recent conversation.” | “Thanks again for walking me through your Q3 hiring plan on Tuesday.” |
| “Our solution can help streamline your processes and drive results.” | “You said your reps spend Friday afternoons updating the CRM. Here’s how two teams cut that time.” |
| “Please do not hesitate to reach out if you have any questions.” | “If the pricing page raises questions, text me. I’m at my desk until 4.” |
| “In conclusion, this proposal outlines a comprehensive approach.” | “Bottom line: we start with your West region in May, then decide on the rest together.” |
Notice the pattern. The “after” lines name a day, a detail from the call or a clear next step. Only you know those things.
What to Do If Your Email or Proposal Gets Flagged as AI
Getting flagged can feel like an accusation. Stay calm and work the problem.
- Ask what tool was used and what the score was. “AI generated” from one tool is not the same as a 15% score from another. Turnitin, for example, does not show scores from 1% to 19%.
- Remember what the score means. It is a probability estimate, not a finding of fact. Turnitin itself says its AI percentage should not be the sole basis for action, and OpenAI said its old classifier should not be a primary decision-making tool.
- Show your writing process. In Google Docs, click Last edit at the top right to open version history and show how your draft changed over time. Save drafts and notes for important documents.
- Be honest about AI use. If you used AI to brainstorm or tighten a paragraph, say so. Follow your company’s AI policy. If your team allows AI as a helper, make sure a person owns the final message.
- Offer to walk through it. Explain why you made the points you made. Nobody knows your deal, your client or your reasoning better than you.
Why “Beating” AI Detectors Is the Wrong Goal
Search for “how to humanize AI text” and you will find tools that promise to make AI writing pass any AI checker. That is the wrong target for three reasons.
- Detectors change. What slips past a detector today may get flagged next month.
- Readers notice. Buyers, managers and recruiters can tell when a message is generic, whatever a detector says.
- Trust is the real score. If you rely on hidden tricks and get caught, the damage to your reputation is worse than any flag.
A better goal: use AI responsibly, then make the final message yours. Add what only you know. Cut what any tool could have written.
How to Write in Your Own Voice: Checklist
To make AI-written emails sound like you, add one detail only you know, swap stock phrases for words you would say out loud, vary your sentence length, and rewrite the draft in your own words before you send.
Run through this before you send an important email, proposal or blog post:
- [ ] Did you mention one detail from your last conversation with the reader?
- [ ] Did you replace stock phrases with words you would say out loud?
- [ ] Did you vary sentence length, with a mix of short and longer sentences?
- [ ] Did you state one clear next step with a day or time?
- [ ] If AI helped, did you rewrite the result in your own words and check every fact?
- [ ] Would a teammate recognize this as your writing?
Catch Mistakes Before You Hit Send
FlyMSG does not make an AI detector, and you do not need one to write well. What helps is clean, human writing: your ideas, in your own voice, without typos.
The FlyMSG AI Grammar Checker is a free Chrome and Edge browser extension that fixes the slips and leaves the voice to you. It checks grammar, spelling and punctuation as you type on any website with a text field, including email, LinkedIn and your CRM. It supports 80+ languages, which helps if you write at work in more than one language.
FlyMSG has a free plan. If you upgrade, use code 20PercentYr1 for 20% off your first year on an annual Growth or Business Pro plan.
Want to use AI tools well without losing your voice? Read our guide to writing good AI prompts and our overview of the AI writing assistant.
Quick Quiz: Test What You Know About AI Detection
- An AI detector proves whether AI wrote a text. True or false? False. It gives a probability based on patterns.
- Why did OpenAI shut down its AI text classifier? Its low rate of accuracy.
- Which writers did the Stanford study find were flagged most often? Non-native English writers.
- What is the best response if your proposal gets flagged? Ask for the tool and score, show your drafts and version history, and be honest about any AI use.
FAQ
Is 20% AI detection bad?
Not by itself. The meaning depends on the tool. Turnitin says its percentage reflects the share of qualifying text its model predicts was AI generated, not the whole submission, and it hides scores from 1% to 19%. Other tools report a probability. A 20% score is a reason to look closer, not proof of AI use.
Can Turnitin detect ChatGPT?
Turnitin says its English model can detect AI generated content from OpenAI’s GPT models, the models behind ChatGPT, plus other large language model families such as Gemini and Claude. It needs at least 300 words of long-form prose. Like every detector, it can make mistakes, and Turnitin says its percentage should not be the sole basis for action.
Does Google penalize AI-generated content?
Not for being AI-generated. Google’s Search guidance says it rewards helpful, high-quality content however it is made. Using automation, including AI, mainly to manipulate search rankings breaks its spam policies. For a blog post, what counts is whether the content is accurate, original and useful to the reader.
