Fast-DetectGPT vs MAGE, which AI detector is more accurate?
Fast-DetectGPT is more accurate on both catch rate and false positives. It caught 68.1% of AI texts against 46.4% for MAGE, and flagged fewer human texts, 15.9% against 19.0%. MAGE wrongly flags fewer academic abstracts and Q&A answers, but it also catches little AI text of those kinds. Fast-DetectGPT's threshold was set on our calibration split.
Tested
Fast-DetectGPT (Qwen2.5-1.5B) vs MAGE: the numbers side by side
| Measure | Fast-DetectGPT (Qwen2.5-1.5B) | MAGE |
|---|---|---|
| Rank on the leaderboard | 1 of 7 | 4 of 7 |
| Balanced accuracy | 76.1% | 63.7% |
| Catches AI text | 68.1% | 46.4% |
| Flags human text | 15.9% | 19.0% |
| Catches humanized AI text | 47.6% | 23.8% |
| AUROC | 0.815 | 0.666 |
| Flags famous public-domain texts | 56.7% | 20.0% |
| Catches Claude Fable 5.1 | 46.0% | 28.6% |
| Catches Claude Haiku 5.5 | 88.9% | 66.7% |
| Catches Claude Opus 5.5 | 63.5% | 45.2% |
| Catches Claude Sonnet 5.5 | 73.8% | 45.2% |
| Flags human text: Academic abstracts | 40.0% | 12.0% |
| Flags human text: Q&A answers | 21.4% | 7.1% |
| Flags human text: Encyclopedia articles | 8.0% | 40.0% |
| Flags human text: Fiction | 0.0% | 4.2% |
| Flags human text: News reports | 8.3% | 33.3% |
| Seconds per 1,000 words | 0.99 | 0.39 |
| Type | Zero-shot | Trained classifier |
| Licence | MIT (method), Apache-2.0 (model) | Apache-2.0 |
Two approaches from 2024
Both detectors come from 2024 research papers, but they work in very different ways. MAGE is a Longformer classifier, trained on human text and the output of many earlier language models. Fast-DetectGPT is zero-shot: it has no detector training, and it asks a language model, here Qwen2.5-1.5B, how predictable each word is. A method that never learned from old models’ output may cope better with new ones. On our Claude texts, Fast-DetectGPT caught more on every model, including Claude Fable 5.1, where it caught 46.0% against 28.6% for MAGE.
One caveat runs through this comparison. The Fast-DetectGPT method has no threshold for this language model, so we chose one on our calibration split. MAGE used the cut-off from its authors’ own code.
The kinds of writing each gets wrong
The two have almost mirror-image false positives. Fast-DetectGPT flagged 40.0% of human academic abstracts and 21.4% of Q&A answers, where MAGE flagged 12.0% and 7.1%. MAGE in turn flagged 40.0% of human encyclopedia articles and 33.3% of news reports, against 8.0% and 8.3% for Fast-DetectGPT.
MAGE’s low numbers on abstracts and answers come with a catch. It caught only 12.0% of AI abstracts and 21.4% of AI answers, so in those genres it mostly says “human” to everything. Fast-DetectGPT caught 84.0% and 43.8%.
Famous public-domain texts are where Fast-DetectGPT is the more trigger-happy one. It flagged 17 of 30 classics and MAGE 6, including all of MAGE’s picks. A language model has likely read texts like the Declaration of Independence many times, which would make them look predictable to Fast-DetectGPT.
Short and edited texts
Length was MAGE’s biggest weakness. It caught 20.6% of AI texts under 200 words, against 74.2% for Fast-DetectGPT, which did about as well on short texts as on long ones. After a humanizer rewrite of the Claude Opus 5.5 texts, Fast-DetectGPT caught 47.6% and MAGE 23.8%.
Formatting moved Fast-DetectGPT and not MAGE. With line breaks removed, Fast-DetectGPT’s catch rate fell to 57.9%, still above MAGE’s 46.4%, which did not change.
Which detector to choose
Fast-DetectGPT, for nearly everyone. The exception is someone who checks only academic abstracts and fears a false accusation above all. MAGE flags fewer human abstracts, though it also lets most AI ones through. Both flag too much human writing to serve as evidence on their own. For a detector with fewer false positives, see Desklib or the false positives page.
Questions people ask
Is MAGE or Fast-DetectGPT better for academic writing?
Neither works well. MAGE flagged 12.0% of human abstracts but caught only 12.0% of AI ones. Fast-DetectGPT caught 84.0% of AI abstracts but flagged 40.0% of human ones.
Which is better on short texts?
Fast-DetectGPT, by a wide margin. Under 200 words it caught 74.2% of AI texts and MAGE 20.6%.
Which one is faster?
MAGE, at 0.39 seconds per 1,000 words against 0.99 for Fast-DetectGPT on an Apple M4 Pro.