# Can AI detectors flag writing that a human wrote?

Source: https://texthumanizer.com/forum/discuss/can-ai-detectors-flag-writing-that-a-human-wrote/
Site: Text Humanizer Forum, category Reading AI detector scores
Published: 2026-11-16. Updated: 2026-11-16. Replies: 0.

## Question by Desmond Achterberg (staff), 2026-11-16

Yes. Detectors make false positives: they sometimes label human writing as AI-written. The vendors admit it, and published research found detectors misclassify non-native English writing more often than native writing. Short, formal or formulaic text is also at risk. A flag is a reason to ask questions, not evidence of anything.

## The vendors say so themselves

You do not have to take a critic's word for this. The detector makers publish it.

- Turnitin's [own explanation of false positives](https://www.turnitin.com/blog/understanding-the-false-positive-rate-for-sentences-of-our-ai-writing-detection-capability) says its sentence-level false positive rate is around 4%, meaning a highlighted sentence has a 4% likelihood of being human-written. It says this is more common in documents that mix human and AI writing, and that it cannot "mitigate the risk of false positives completely." Its advice to instructors is to use highlights "to initiate a conversation, not to draw a conclusion."
- GPTZero's [FAQ](https://gptzero.me/faq) says "there always exist edge cases with both instances where AI is classified as human, and human is classified as AI." It adds that the classifier "can sometimes flag other machine-generated or highly procedural text as AI-generated."
- OpenAI [retired its own classifier](https://openai.com/index/new-ai-classifier-for-indicating-ai-written-text/) in July 2023 for low accuracy. While it was live, the company warned that human-written text is sometimes "incorrectly but confidently" labeled as AI-written.

## The bias finding

The study most often cited is Liang, Yuksekgonul, Mao, Wu and Zou, ["GPT detectors are biased against non-native English writers"](https://arxiv.org/abs/2304.02819) (arXiv 2304.02819, later published in the journal Patterns). The authors evaluated several widely used detectors on writing samples from native and non-native English writers. Their abstract reports that the detectors "consistently misclassify non-native English writing samples as AI-generated, whereas native writing samples are accurately identified."

Their explanation is that detectors can penalize writers with "constrained linguistic expressions", meaning a narrower range of vocabulary and sentence patterns. That is common for anyone writing carefully in a second language. The authors call for caution about using detectors "in evaluative or educational settings." The measured rates are in the paper itself, so read it there rather than relying on a number repeated second-hand.

## Which human writing gets flagged

From the sources above, these are the situations where a false positive is more plausible:

| Situation | Why it is risky |
|---|---|
| Writing in a second language | Narrower vocabulary and safer sentence patterns (Liang et al.) |
| Very short samples | OpenAI called its classifier very unreliable below 1,000 characters |
| Lists, templates, procedures | GPTZero says procedural text can be flagged |
| Heavily formulaic prose | Predictable text is hard to classify either way (OpenAI) |
| Documents that mix human and AI text | Turnitin says false positives cluster near real AI passages |

## What a flagged writer can do

First, do not rush to rewrite for the detector. Rewriting to chase a number is a poor use of an afternoon and it does not prove anything. The useful moves are about evidence and clarity.

1. Keep your process: notes, outlines, drafts, tracked changes, version history in Google Docs or Word.
2. Be ready to explain your argument and your choices out loud. A writer can do that; a score cannot.
3. If you did use an AI tool for part of the work, say so, and check your policy first. See [do I have to disclose AI use](/forum/discuss/do-i-have-to-disclose-that-i-used-ai-to-help-write-something/).
4. Ask what the detector report actually shows. See [what an AI detector score means](/forum/discuss/what-does-an-ai-detector-score-actually-mean/).

## A plain-prose example

Careful second-language writing often looks like this, and it is fine:

> The city has many problems. The first problem is traffic. The second problem is the price of housing. These problems are important for the people.

A reader would say it is clear but flat. The edit for a reader is about specificity, not about hiding anything:

> Two problems stand out in the city: slow traffic and expensive housing. Rush-hour traffic adds a long wait to every commute, and rents have pushed many workers to the edge of town.

Use your own facts there, not mine. The second version is better writing, and that is the reason to make the edit. For the robotic patterns editors look for, see [why AI writing sounds robotic](/forum/discuss/why-does-ai-writing-sound-robotic-and-what-patterns-give-it-away/). Do not add errors on purpose. That helps nobody.

---
Cite as: Text Humanizer Forum, "Can AI detectors flag writing that a human wrote?", https://texthumanizer.com/forum/discuss/can-ai-detectors-flag-writing-that-a-human-wrote/, accessed 2026-11-17.
