Desklib AI detector accuracy test on 900 texts
Desklib's AI detector flagged 6.3% of human-written texts in our test, the lowest false positive rate of any detector that catches a real share of AI writing, and it had the best AUROC (0.877). It caught 58.3% of AI texts, though, and only 32.5% of texts from Claude Fable 5.1.
Tested
Verdict: The most careful detector we tested and the best at telling AI from human text overall, but it misses about four in ten AI texts and most AI fiction.
Desklib AI text detector test results
| Rank | 2 of 7 |
|---|---|
| Balanced accuracy | 76.0% |
| Catches AI text | 58.3% (294 of 504; 95% interval 54.0 to 62.6%) |
| Flags human text | 6.3% (8 of 126; 95% interval 3.3 to 12.0%) |
| Catches humanized AI text | 32.5% (41 of 126) |
| AUROC | 0.877 |
| Speed on an Apple M4 Pro | 0.57 seconds per 1,000 words |
| Type | Trained classifier |
| Threshold | 0.5 (model card (predict_single_text, threshold=0.5)) |
| Licence | MIT |
| Model | desklib/ai-text-detector-v1.01 |
Numbers behind this chart
| Item | Rate | Texts | 95% interval |
|---|---|---|---|
| Claude Fable 5.1 | 32.5% | 41 of 126 | 25.0 to 41.1% |
| Claude Haiku 5.5 | 74.6% | 94 of 126 | 66.4 to 81.4% |
| Claude Opus 5.5 | 62.7% | 79 of 126 | 54.0 to 70.6% |
| Claude Sonnet 5.5 | 63.5% | 80 of 126 | 54.8 to 71.4% |
| Humanized Claude Opus text | 32.5% | 41 of 126 | 25.0 to 41.1% |
Numbers behind this chart
| Item | Rate | Texts | 95% interval |
|---|---|---|---|
| Academic abstracts | 8.0% | 2 of 25 | 2.2 to 25.0% |
| Q&A answers | 3.6% | 1 of 28 | 0.6 to 17.7% |
| Encyclopedia articles | 0.0% | 0 of 25 | 0.0 to 13.3% |
| Fiction | 0.0% | 0 of 24 | 0.0 to 13.8% |
| News reports | 20.8% | 5 of 24 | 9.2 to 40.5% |
Numbers behind this chart
| Item | Rate | Texts | 95% interval |
|---|---|---|---|
| under 200 words | 51.5% | 50 of 97 | 41.7 to 61.2% |
| 200 to 299 words | 63.5% | 158 of 249 | 57.3 to 69.2% |
| 300 words or more | 54.4% | 86 of 158 | 46.7 to 62.0% |
What Desklib is
Desklib is a company that sells study tools, and it published this detector as a free model. It is a fine-tuned DeBERTa-v3-large model, the largest classifier in our test, and its model card says it led the RAID benchmark leaderboard when it was released. We used the threshold of 0.5 from its model card.
The fewest false accusations
Desklib flagged no human encyclopedia articles and no human fiction. Its false positives were mostly news reports: 5 of 24, or 20.8%. Q&A answers and academic abstracts were rarely flagged, at 3.6% and 8.0%.
Its AUROC of 0.877 was the best in the test. AUROC measures how well a detector’s scores separate AI from human text across every possible threshold, so Desklib ranks texts better than any other detector here, even where its fixed threshold lets AI text through.
What it misses
It caught 58.3% of AI texts overall. Fiction was its weak spot: it caught only 20.8% of AI-written fiction passages. It also struggled with Claude Fable 5.1, catching 32.5% of its texts, against 74.6% for Claude Haiku 5.5.
A humanizer rewrite nearly halved its catch rate on Opus texts, from 62.7% to 32.5%. If you use Desklib, read a pass as “probably not raw AI output”, not as proof of human writing.
On an Apple M4 Pro it took 0.57 seconds per 1,000 words.
Human writing Desklib AI text detector flagged as AI
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...
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,...
Questions people ask
Is the Desklib AI detector accurate?
It is the most accurate of the seven we tested at separating AI from human text, with an AUROC of 0.877. At its default threshold it caught 58.3% of AI texts and flagged 6.3% of human texts.
Does Desklib flag human writing?
Rarely, but it happens. It flagged 8 of 126 human texts, and 5 of those were news reports, so news is where to be most careful with its verdict.
Can Desklib detect humanized AI text?
Less well. Its catch rate on Claude Opus 5.5 texts fell from 62.7% to 32.5% after a light rewrite with a humanizer.