Do AI detectors flag human writing? 117 of 126 human texts were flagged
Yes, often. We ran seven free AI detectors on 126 human texts that were all published before ChatGPT launched on 30 November 2022. 117 of them (92.9%) were flagged as AI by at least one detector. The most careful detector, Desklib, flagged 6.3%; the worst, TMR, flagged 83.3%.
Tested
Numbers behind this chart
| Item | Rate | Texts | 95% interval |
|---|---|---|---|
| OpenAI RoBERTa detector | 4.8% | 6 of 126 | 2.2 to 10.0% |
| Desklib AI text detector | 6.3% | 8 of 126 | 3.3 to 12.0% |
| Fast-DetectGPT (Qwen2.5-1.5B) | 15.9% | 20 of 126 | 10.5 to 23.2% |
| MAGE | 19.0% | 24 of 126 | 13.1 to 26.8% |
| Fakespot AI text detector | 34.9% | 44 of 126 | 27.2 to 43.6% |
| E5-small LoRA detector | 50.0% | 63 of 126 | 41.4 to 58.6% |
| TMR (Target Mining RoBERTa) | 83.3% | 105 of 126 | 75.9 to 88.8% |
Which kinds of human writing get flagged?
| Detector | Academic abstracts | Q&A answers | Encyclopedia articles | Fiction | News reports |
|---|---|---|---|---|---|
| Fast-DetectGPT (Qwen2.5-1.5B) | 40.0% | 21.4% | 8.0% | 0.0% | 8.3% |
| Desklib AI text detector | 8.0% | 3.6% | 0.0% | 0.0% | 20.8% |
| Fakespot AI text detector | 12.0% | 39.3% | 40.0% | 12.5% | 70.8% |
| MAGE | 12.0% | 7.1% | 40.0% | 4.2% | 33.3% |
| E5-small LoRA detector | 36.0% | 75.0% | 52.0% | 12.5% | 70.8% |
| OpenAI RoBERTa detector | 0.0% | 10.7% | 4.0% | 4.2% | 4.2% |
| TMR (Target Mining RoBERTa) | 16.0% | 100.0% | 100.0% | 100.0% | 100.0% |
How we know these texts are human
Every text on this page was published on Wikipedia, Wikinews, Stack Exchange, arXiv or Project Gutenberg before 30 November 2022, the day ChatGPT came out. For Wikipedia and Wikinews we used the article as it stood before that date, and for Stack Exchange we dropped any answer edited after it. AI writing tools existed before ChatGPT, but they were rare in this kind of writing.
The detectors disagree about which texts look like AI
Only 9 of the 126 human texts passed every detector. Each detector has its own blind spot. Fast-DetectGPT flagged 40.0% of academic abstracts and no fiction. Fakespot and E5-small LoRA flagged most news reports. MAGE flagged 40.0% of encyclopedia articles. Because the false flags fall on different texts, every extra detector you run adds new false positives.
If you check a text with several free tools and treat any “AI” result as a warning, these seven detectors would have flagged 92.9% of genuine human writing.
Detectors with the most false positives
TMR flagged 83.3% of human texts, including every news report, encyclopedia article, forum answer and fiction passage. E5-small LoRA flagged 50.0%, and Fakespot 34.9%. Each of them catches plenty of AI text too, but a detector that flags this much human writing cannot tell the two apart. The leaderboard compares all seven.
Real human writing that detectors called AI
It is extra payments that you apply to the principal. Here is an example. Let's say that you have a mortgage with a monthly payment of $1000 per month (principal and interest), which you pay on the 1st of the month. Now, let's say that, in addition to your $1000 regular payment, you have an extra $1000 that you would like to pay to reduce your debt. So after you...
Last Sunday evening at the Beverly Hilton Hotel in Beverly Hills, California were the 67th Annual Golden Globe Awards hosted by English comedian Ricky Gervais. The choice of Gervais as host was a move to increase viewership which declined to 15 million viewers last year from 27 million in 2004. Canadian director James Cameron—who previously won in 1997 for his film Titanic—and his high-grossing 3D epic Avatar seemed to dominate...
Despite the prevalence of community detection algorithms, relatively less work has been done on understanding whether a network is indeed modular and how resilient the community structure is under perturbations. To address this issue, we propose a new vertex-based metric called "permanence", that can quantitatively give an estimate of the community-like structure of the network. The central idea of permanence is based on the observation that the strength of membership...
Labeling data is an important step in the supervised machine learning lifecycle. It is a laborious human activity comprised of repeated decision making: the human labeler decides which of several potential labels to apply to each example. Prior work has shown that providing AI assistance can improve the accuracy of binary decision tasks. However, the role of AI assistance in more complex data-labeling scenarios with a larger set of labels...
Unlike many other countries, US airports do not have any form of physical immigration controls when you are departing the country on an international flight. In fact, in most airports there isn't even a concept of an "International" terminal/gate, with the same gates frequently being used for international flights and domestic flights. As a result of this they can't enforce the concept of a 'transit' passenger - once you're in...
A United Nations report released Thursday states that recent attacks carried out on four villages in Darfur by the Sudanese military and armed miltias were a violation of international law. The attacks, which involved aerial bombardment in two of the villages, left a total of 115 people dead and over 30,000 forcibly displaced, according to the report. Damage to civilian property was extensive; homes, schools, and shops were systematically destroyed,...
A 16-year-old boy has been arrested after he hijacked the Washington Township, New Jersey Wal-Mart's public announcement system to make racial comments. The boy was arrested Friday night and then released into his parents' custody. He was charged with harassment as a bias crime. The boy, whose name was not released, faces up to a year in a youth detention center if convicted. Police say that they are not sure...
In a report released in The New England Journal of Medicine on Wednesday, scientists studying patients in Zhuhai in China's Guangdong Province reported that the newly emerged SARS-CoV-2, the coronavirus that causes COVID-19, may spread in a manner similar to influenza — suggesting, experts noted, that this outbreak, which began in the city of Wuhan in Hubei Province in December 2019, may call for containment measures similar to those for...
The hot Jupiter WASP-43b has now joined the ranks of transiting hot Jupiters HD 189733b and HD 209458b as an exoplanet with a large array of observational constraints on its atmospheric properties. Because WASP-43b receives a similar stellar flux as HD 209458b but has a rotation rate 4 times faster and a much higher gravity, studying WASP-43b serves as a test of the effect of rotation rate and gravity on...
On the topic of the Aztecs, an intriguing book on this subject is Aztec Warfare: Imperial Expansion and Political Control, by Ross Hassig. The Aztecs were an extremely war-like civilization, that were constantly attacking and subjugating their neighbors. Interestingly, though, their style of warfare was quite different from what we are familiar with from western european history. For the Aztecs it was all about taking prisoners. The more prisoners you...
Through the course of World War 2; in September 1939 battleships were still considered more important than aircraft carriers, and by VJ Day carriers were considered utterly decisive at sea. This process took a good few steps, however, and really examining it in detail could take an entire book. You also need to consider the application of carrier air power to two main problems: 1) the decisive battle 2) the...
Vote using physical tokens (like stones in pots) Deciding who has the most votes is then a matter of simultaneously reducing all the piles of votes by 1 token each. Eventually you’ll reach a point where only one pile has any tokens left. This is the winner. To expand this out to multiple voting sites you can either take the remaining tokens in the pile (which represent the margin of...
Questions people ask
Which AI detector has the fewest false positives?
The OpenAI RoBERTa detector flagged the fewest human texts, 4.8%, but it also misses almost all AI text. Among detectors that catch a real share of AI writing, Desklib had the fewest false positives at 6.3%.
What kind of human writing gets flagged as AI most?
News reports. Fakespot and E5-small LoRA each flagged 70.8% of human news reports, and TMR flagged all of them. Fiction was flagged least by most detectors.
Can a teacher prove a student used AI with a detector?
Not with these detectors. Even Desklib, the most careful one we tested, flagged 6.3% of human texts, and different detectors flag different texts, so running several makes a false flag more likely, not less.