AI 寫作檢測器如何運作:為何同一篇文章可得 100% 和 0%
How AI Detectors Work: Why One Essay Can Score 100% and 0%
AI 寫作檢測器按各自模型分析完成稿中的寫作模式,而非寫作過程,因此同一篇文章可在一個檢測器得 100%,在另外兩個得 0%。Turnitin 的文件百分比是超過門檻的句子比例,100% 並不代表結果確定。文章亦提到 Adelphi 案中,紐約法官於 2026 年 1 月因程序及證據問題撤銷違規裁定,而非裁定檢測器準確性;若被標記,可提交草稿和版本紀錄。
How AI detectors work: each one scores your finished text, not how you wrote it, with its own model and its own line, so one essay can get 100% AI from one detector and 0% from two others. We start with a real case: at Adelphi University a professor's report said Turnitin scored an essay 100%, the student said two detectors his parents ran showed 0%, and a New York judge annulled the violation in January 2026, over process and evidence, not detector accuracy.
From there we follow the idea: perplexity and burstiness, why GPTZero called part of the US Constitution likely AI in 2023, OpenAI's own classifier (26% of AI text caught in its test, 9% of human writing wrongly flagged, then pulled for low accuracy), and the deep learning models GPTZero and Turnitin describe now. Then we cover what a Turnitin percentage means, what recent tests found on false positives and missed AI, and what to bring if you're flagged: drafts, version history and your school's rule on tools like Grammarly.
Cheat sheet
An AI detector scores your finished text for patterns that resemble AI writing, not how you wrote it. The result is a statistical estimate. Turnitin and GPTZero say it is not a verdict.
What the detector actually measures
- Predictability (early tools). Perplexity is how surprising text is to a language model. Low perplexity leaned AI. Ars Technica's example: "a cup of coffee" is low perplexity, "a cup of spiders" is high.
- Burstiness. GPTZero defined it as how much writing patterns and perplexity vary across a document.
- Learned patterns (now). GPTZero says it moved to deep learning in autumn 2023, though the same page still calls both one of seven indicators. Turnitin says its model is not explicitly programmed for either, but still explains detection through word probability.
- A threshold. Turnitin scores each sentence from 0 to 1. The document percentage is the share over its threshold.
What the evidence says
| Myth | Evidence |
|---|---|
| ChatGPT's maker has a working detector | OpenAI's classifier caught 26% of AI text and wrongly flagged 9% of human text in its 2023 test, then was withdrawn that July for low accuracy. |
| Detectors mostly flag human writing | A 2025 working paper put GPTZero's false positive rate at about 0.7% and Pangram's near zero. Not peer reviewed, no student essays, and its authors expect an arms race. |
| Few false alarms means AI gets caught | In a small, peer-reviewed 2026 test, Turnitin, GPTZero and Copyleaks failed on fully AI papers from ChatGPT Deep Research. |
| A 1% error rate is harmless | Vanderbilt estimated Turnitin's claimed 1% would have meant about 750 of Vanderbilt's 75,000 papers from 2022 wrongly labelled: a projection, not a count. |
What it cannot tell
- Predictable human writing. In 2023, GPTZero said part of the US Constitution was likely written entirely by AI. People write predictably too, especially in formal styles such as law.
- Non-native writers. In a 2023 Stanford study, seven detectors averaged about 61% false positives on 91 TOEFL essays. That finding is contested, and GPTZero self-reports 1.1% false positives on TOEFL texts.
- Short or mixed text. Detection weakens sharply on short texts and joint human-and-AI essays.
- The trade-off. Within one detector, fewer false positives means more missed AI text, and the reverse.
If your work is flagged
- Turnitin says its percentage should not be the sole basis for action and does not determine misconduct.
- Researchers say a flag should be paired with drafting history, so keep drafts and version history.
- Detectors can disagree on the same essay, which you can raise in a dispute.
One essay, 100% and 0%: the Adelphi case
Your essay gets 100 percent AI from one detector, zero from two. That happened at Adelphi University: the professor's report said Turnitin scored it 100 percent, and accused the student of using Grammarly as generative AI. He denied writing it with AI, and said two detectors his parents ran showed zero. A New York judge annulled the violation in January 2026, over process and evidence, not detector accuracy.
Why detectors disagree, and the bet on predictability
How can one essay get 100 and zero? Each detector has its own model and line, and scores the finished text, not how you wrote it. Early detectors bet on predictability: AI picks the likely next word, and people are quirkier.
Perplexity and burstiness (a cup of spiders)
"I'd like a cup of coffee" surprises nobody. "A cup of spiders" does. That surprise is perplexity: low leaned AI. GPTZero's 2023 dashboard also scored burstiness, how much it varies.
Why the US Constitution got flagged as AI
In 2023, GPTZero said part of the US Constitution was likely written entirely by AI. GPTZero's creator, Edward Tian, called it "a text fed repeatedly into the training data." People write predictably too, especially in formal styles like law.
OpenAI's own classifier: 26% caught, then pulled
OpenAI's own trained classifier caught 26 percent of AI text in its 2023 test, and wrongly flagged 9 percent of human writing. By July 2023, it was gone, for low accuracy. So can a detector tell?
From perplexity to deep learning
GPTZero says it moved to deep learning in 2023, and Turnitin says its model isn't explicitly programmed for perplexity, though it still explains detection through word probability.
What a 100% score means (Turnitin, GPTZero)
Today, Turnitin's model learns patterns and scores each sentence from 0 to 1. Your percentage is the share over its line, so 100 doesn't mean certain. GPTZero's demo calls its number a probability, meaning how similar a text is to AI writing. Similar, not proven.
False positives on human writing
Fully human writing looks safer: in a 2025 working paper, GPTZero wrongly flagged about 0.7 percent, Pangram almost none. The catch: no student essays, no peer review, and its authors expect an arms race.
Where detectors miss AI text
Catching AI is the weak spot: in a small 2026 test, three tools got no fully AI paper right. On 40 mixed papers, scores landed in the true AI range zero to 37 times, depending on the tool.
False alarms vs missed AI (Vanderbilt's 750)
Detectors pick a line: fewer false alarms usually means more missed AI. Vanderbilt figured Turnitin's claimed 1 percent could have meant about 750 of its 75,000 papers from 2022 wrongly labeled. That was one concern when it switched the detector off in 2023.
If you're flagged: what to bring
If you're flagged, bring drafts, version history and your school's rule on tools like Grammarly, and point out detectors can disagree. Even Turnitin says its score shouldn't be the sole basis for action. A detector scores how your text looks, but your drafts show how you wrote it.
Sources
- 2026 NY Slip Op 26021 (Supreme Court, Nassau County, January 28, 2026) — New York State Law Reporting Bureau
- Why AI writing detectors don't work — Ars Technica
- GLTR: Statistical Detection and Visualization of Generated Text — arXiv
- Perplexity, burstiness, and statistical AI detection — GPTZero
- GPTZero Technology — GPTZero
- How to Use GPTZero | GPTZero Tutorial — GPTZero on YouTube
- What is GPTZero? | Demo Video 2025 — GPTZero on YouTube
- New AI classifier for indicating AI-written text — OpenAI
- Turnitin's AI writing detection capabilities FAQs — Turnitin
- How should I review the AI Writing report? — Turnitin
- Understanding the false positive rate for sentences of our AI writing detection capability — Turnitin
- Understanding the false positive rate for sentences within our AI writing detection — Turnitin on YouTube
- Artificial Writing and Automated Detection — NBER
- Do AI Detectors Work Well Enough to Trust? — Chicago Booth Review
- Who wrote this? Evaluating the reliability of AI detection tools in higher education — International Journal for Educational Integrity
- Guidance on AI Detection and Why We're Disabling Turnitin's AI Detector — Vanderbilt University
- Detecting AI-Generated Essays in Writing Assessment: Responsible Use and Generalizability Across LLMs — arXiv
- Why AI Detection Fails for Academic Integrity — arXiv
- Who Wrote This? Turing, Total Variation, and the Mathematics of AI-Text Detection — arXiv
- AI Detectors Don't Work. Here's What to Do Instead. — MIT Sloan Teaching & Learning Technologies
Read the full transcript, with every source, on the page for this video. More explainers like this one are on the Kush_EXE YouTube channel.
This article is built from the video's script and published sources. The video uses AI illustrations.
來源:DEV Community · dev.to