Best AI detector for news articles (tested)
Fast-DetectGPT is the best AI detector for news articles in our test. It caught 93.8% of AI-written news and flagged 2 of 24 human Wikinews reports (8.3%), a balanced accuracy of 92.7%. Desklib came second at 84.9% but flagged 5 of 24. With only 24 human reports, the 95% interval for Fast-DetectGPT's false positive rate still runs up to 25.8%.
Tested
AI detectors ranked for news articles
| # | Detector | Balanced accuracy | Catches AI text | Flags human text |
|---|---|---|---|---|
| 1 | Fast-DetectGPT (Qwen2.5-1.5B) | 92.7% | 93.8% | 8.3% |
| 2 | Desklib AI text detector | 84.9% | 90.6% | 20.8% |
| 3 | MAGE | 79.2% | 91.7% | 33.3% |
| 4 | Fakespot AI text detector | 64.6% | 100.0% | 70.8% |
| 5 | E5-small LoRA detector | 62.0% | 94.8% | 70.8% |
| 6 | OpenAI RoBERTa detector | 55.7% | 15.6% | 4.2% |
| 7 | TMR (Target Mining RoBERTa) | 50.0% | 100.0% | 100.0% |
Numbers behind this chart
| Item | Rate | Texts | 95% interval |
|---|---|---|---|
| Fast-DetectGPT (Qwen2.5-1.5B) | 8.3% | 2 of 24 | 2.3 to 25.8% |
| Desklib AI text detector | 20.8% | 5 of 24 | 9.2 to 40.5% |
| MAGE | 33.3% | 8 of 24 | 18.0 to 53.3% |
| Fakespot AI text detector | 70.8% | 17 of 24 | 50.8 to 85.1% |
| E5-small LoRA detector | 70.8% | 17 of 24 | 50.8 to 85.1% |
| OpenAI RoBERTa detector | 4.2% | 1 of 24 | 0.7 to 20.2% |
| TMR (Target Mining RoBERTa) | 100.0% | 24 of 24 | 86.2 to 100.0% |
Catch rates and false flags on news
Most detectors caught AI-written news better than any other kind of AI writing. Every detector except the OpenAI RoBERTa detector, which was built in 2019 for GPT-2, caught more than 90% of it. The human side went badly. TMR flagged all 24 human Wikinews reports, and Fakespot and E5-small LoRA flagged 17 each, so their near-perfect catch rates on AI news tell you very little.
The likely reason is the form itself. A wire-style report puts the main fact in the first sentence, attributes every claim (“police said”, “according to the report”), stays in the past tense and keeps opinion out. Ask a chatbot for a news story and you get the same shape, so a classifier that has learned what AI news looks like has little left to separate it from a reporter’s copy. A 167-word Wikinews report on the 67th Golden Globe Awards was flagged by six of the seven detectors, including both Fast-DetectGPT and Desklib.
Why Fast-DetectGPT wins on news
Fast-DetectGPT reached 92.7% balanced accuracy on news, the highest score any detector got in any genre. It flagged 2 of 24 human reports and caught 93.8% of AI ones. Two caveats apply. Its threshold was tuned on our calibration split, which none of the other detectors had. And with 24 reports the intervals overlap heavily: 2.3% to 25.8% for its false positive rate, 9.2% to 40.5% for Desklib’s.
News is also where Desklib loses its usual edge. Across all our human texts it flags the fewest of any useful detector, but here it flagged 5 of 24 reports (20.8%), more than Fast-DetectGPT and more than in any other genre. It still caught 90.6% of AI news. Desklib did much better on Wikipedia, which has a similar neutral tone; see the encyclopedia results.
Using a detector in a newsroom
Everything above is raw AI output. A story that an editor has reworked, or that a writer drafted with AI and then rewrote, is likely harder to catch. In our humanized text test, a light rewrite cut the catch rates of most detectors. The leaderboard ranks all seven detectors across every genre.
Questions people ask
Can AI detectors tell if a news article was written by AI?
For raw AI output, usually. Every detector except the OpenAI RoBERTa detector caught more than 90% of AI news articles. The trouble is human news, which Fakespot and E5-small LoRA flagged 17 times out of 24 and TMR flagged every time.
Why does real news get flagged as AI?
A likely reason is that news reports follow a fixed pattern, and AI models copy it closely when asked for a news story. News was Desklib's worst genre, with 5 of 24 human reports flagged, and Fakespot and E5-small LoRA each flagged 70.8%.
Should an editor run freelance copy through an AI detector?
Fast-DetectGPT is the most useful of the seven for news, but our 24 human reports cannot pin its false positive rate down closely. It could be anywhere from 2.3% to 25.8%, so use a flag only as a reason to ask the writer about the piece.