Desklib vs MAGE, which AI detector is more accurate?
Desklib wins clearly. It caught more AI text than MAGE, 58.3% against 46.4%, and flagged fewer human texts, 6.3% against 19.0%. It also caught more texts from each of the four Claude models and more humanized text, 32.5% against 23.8%. MAGE's advantages are speed, AI fiction and long texts.
Tested
Desklib AI text detector vs MAGE: the numbers side by side
| Measure | Desklib AI text detector | MAGE |
|---|---|---|
| Rank on the leaderboard | 2 of 7 | 4 of 7 |
| Balanced accuracy | 76.0% | 63.7% |
| Catches AI text | 58.3% | 46.4% |
| Flags human text | 6.3% | 19.0% |
| Catches humanized AI text | 32.5% | 23.8% |
| AUROC | 0.877 | 0.666 |
| Flags famous public-domain texts | 13.3% | 20.0% |
| Catches Claude Fable 5.1 | 32.5% | 28.6% |
| Catches Claude Haiku 5.5 | 74.6% | 66.7% |
| Catches Claude Opus 5.5 | 62.7% | 45.2% |
| Catches Claude Sonnet 5.5 | 63.5% | 45.2% |
| Flags human text: Academic abstracts | 8.0% | 12.0% |
| Flags human text: Q&A answers | 3.6% | 7.1% |
| Flags human text: Encyclopedia articles | 0.0% | 40.0% |
| Flags human text: Fiction | 0.0% | 4.2% |
| Flags human text: News reports | 20.8% | 33.3% |
| Seconds per 1,000 words | 0.57 | 0.39 |
| Type | Trained classifier | Trained classifier |
| Licence | MIT | Apache-2.0 |
Desklib wins almost everywhere
Desklib flagged less human writing than MAGE in each of the five genres we tested. The largest gap was encyclopedia articles, where MAGE flagged 10 of 25 human Wikipedia excerpts and Desklib none. In news, Desklib’s weakest genre, MAGE flagged 33.3% and Desklib 20.8%.
On AI text the gap was widest for Claude Opus 5.5 and Sonnet 5.5. MAGE caught 45.2% of each, against 62.7% and 63.5% for Desklib. MAGE was trained on text from earlier language models, which may explain why current Claude writing slips past it more often.
MAGE also flagged more famous texts, 6 of 30 against 4. Both flagged Genesis 1, the US Constitution and the Declaration of Independence. MAGE added the openings of The Communist Manifesto, The Prince and Civil Disobedience, while Desklib’s other flag was The Art of War.
Where MAGE holds its own
Fiction is where MAGE beats Desklib. It caught 43.8% of AI fiction passages against Desklib’s 20.8%, and flagged 1 human fiction passage out of 24, where Desklib flagged none.
Length matters more for MAGE. On AI texts under 200 words it caught only 20.6%, while Desklib caught 51.5%. At 300 words or more MAGE edged ahead, 57.6% against 54.4%, but it also flagged 14.3% of human texts that long, where Desklib flagged none.
MAGE is a Longformer model from an ACL 2024 paper, and we used the decision rule from the authors’ own code. It ran at 0.39 seconds per 1,000 words, Desklib at 0.57. Neither changed its verdicts when we stripped line breaks.
Which to pick
For checking essays, articles or applications, use Desklib. It flags fewer human texts in every genre and catches more AI text overall. MAGE is mainly of interest to researchers comparing against the paper.
Neither handles AI fiction well. If that is your main concern, Desklib vs Fast-DetectGPT shows a detector that caught more AI fiction while flagging no human fiction. The MAGE results include its length breakdown, and the Desklib results its misses by genre.
Questions people ask
Is MAGE more accurate than Desklib?
No. Desklib had a higher catch rate (58.3% against 46.4%), a lower false positive rate (6.3% against 19.0%) and a higher AUROC (0.877 against 0.666).
Does MAGE do anything better than Desklib?
It caught more AI fiction, 43.8% against 20.8%, slightly more AI news, 91.7% against 90.6%, and more AI texts of 300 words or more, 57.6% against 54.4%. It also ran faster, at 0.39 seconds per 1,000 words against 0.57.
Which works better on short texts?
Desklib. On AI texts under 200 words it caught 51.5% and MAGE 20.6%.