AI Detector False Positives: Why Human Writing Can Be Flagged as AI-Generated

Jump to Contents

• Uses of AI detectors
• What AI detectors measure
• What is an AI detector false positive?
• How common are AI false positives?
• What to do if you are falsely accused of AI writing?

Uses of AI detectors

AI detectors are used to flag potentially AI-generated text in the following situations:

1. Universities use it to screen essays and coursework, to catch academic dishonesty
2. Employers use it to screen job applications or test materials for positions that involve writing skills
3. Academic journals and publishers use them to spot potential fake papers and paper mills

What AI detectors measure

AI detectors measure how predictable the text is. They cannot inspect authorship, which AI tool or model produced the text, or whether the text was copied directly from another published source (i.e., plagiarized). AI detectors generate a score based on two factors:

Perplexity: this means how predictable a sequence of words is. AI output is usually much lower in perplexity than human writing because AI models are trained to select high-probability words.

Burstiness: this means how much sentence length and structure varies within and across paragraphs. Human writing is a lot more bursty than AI writing because humans get tired, distracted, lose focus, or get bored of the writing task. AI sentences are uniformly smooth and structured.

What is an AI detector false positive?

An AI detector false positive is when the detector flags a genuinely human-written piece of text as AI written. The opposite is an AI false-negative: an AI-generated text is not flagged as AI-written and is instead labelled human.

Common reasons for AI false positives

1. Short text: AI detection is less accurate when the text to be assessed is comparatively short. That’s why journal or conference abstracts, brief emails, or text messages
2. Text meant to be structured and predictable: Recipes, training manuals, performance reports, legal reports, etc. use standard phrasing and are highly structured by nature. Pangram, a highly popular AI detector, acknowledges on its website that Pangram performs best when “[t]he text contains more creative input, and is less formulaic.”

How common are AI false positives?

You’ve seen AI detector companies claiming “99% accuracy”, but also social media tirades about how “an AI detector flagged my essay from 2019 as AI-written!”. Whom to believe? There’s no independent published record of how often AI detectors get it wrong, but independent research has found more AI false-positives in the following cases:

1. Writing by non-native English speakers: Liang et al. (2023) found that the design of many AI detectors “inherently discriminates against non-native authors, particularly those exhibiting restricted linguistic diversity and word choice”. Simply put, when English isn’t your first language, you have a smaller vocabulary and don’t mess around with complex sentences, preferring simple structures for which you know the rules. In Liang et al.’s study, AI detectors had high false positives for essays written by non-native English speakers and were much more accurate on essays written by native speakers.
2. Scientific writing: Journal articles, theses, dissertations, conference papers, lab reports, etc. all follow a predefined structure and format. There are standard phrases and expressions used to describe procedures, analyses, and results. That means perplexity is low and so is burstiness. You can be creative and spontaneous while reviewing your latest travel coffee mug on social media, but not while describing the primer sequences you used in RT-PCRT. Tsigaris et al. (2026) put it most bluntly in their article on using AI to detect AI in academic writing: most AI detector findings are false.

One thing to keep in mind is that many vendors claim that their AI detectors are “suitable for academic writing” but those tools are much more accurate on essays rather than research papers. Undergraduate coursework is naturally a lot more “bursty” than graduate writing because undergraduates are still learning academic writing conventions and field-specific vocabulary, and are bound to slip up on some expressions. And admissions essays or statements of purpose are nearly always meant to be free from scientific terminology and reflect the personality of the writer.

What to do if you are falsely accused of AI writing?

• Take a breath before you respond. It’s super stressful, but firing off an angry email at 2 am won’t help. Give yourself a few hours to calm down so you can respond clearly and professionally.
• Find out exactly what the accusation is based on. Politely ask the journal editor or your supervisor/instructor about what they’re basing their accusation on. Was it a detection tool like Pangram, Turnitin, or GPTZero, or something about your writing style? You can’t defend yourself properly without knowing this.
• Dig up your version history. If you wrote in Google Docs or Word (with OneDrive), you can pull up the edit history showing your paper, dissertation, or essay being built bit by bit over time. This is honestly your strongest evidence, so screenshot or export it ASAP.
• Gather your messy behind-the-scenes stuff. Lab notebooks, fieldwork notes, rough drafts, outlines, brainstorming notes, handwritten scribbles, browser history of your research, saved sources in your reference manager, output from your analysis software, email exchanges with co-authors, all show the real process of how you worked.
• If you’re a student, offer to talk through your work. Ask to meet your instructor and explain your argument, how you found your sources, and how you got to your conclusions. Someone who actually wrote the paper can usually discuss it easily, and this can clear things up fast.
• Offer samples of your previous writing. Older articles or essays, especially ones written before 2022, can show that this is just how you write. If you’ve always used sophisticated vocabulary or a formal tone, that’s your style.
• If you’re a student, read your school’s academic integrity policy. Many schools have specific rules about whether AI detector results can be used as evidence.
• Protect yourself going forward. Get in the habit of writing in tools like Google Docs that save version history and keeping your drafts. It’s annoying that you have to, but it makes proving non-use of AI way easier next time.

And remember: many journals and publishers actually do permit AI use, so always read the target journal’s guidelines carefully before submission and frame a suitable AI disclosure statement if required.

References

Liang W et al. 2023. GPT detectors are biased against non-native English writers. doi: 10.1016/j.patter.2023.100779

Tsigaris P et al. 2026. AI detecting AI in academic writing: Why most AI detector findings are false. https://www.sciencedirect.com/science/article/abs/pii/S305047592600093X?utm