Desklib vs Fakespot, which AI detector is more accurate?
Desklib, for most people, because it flags far less human writing, 6.3% of human texts in our test against 34.9% for the Fakespot AI text detector. Fakespot caught more AI text, 66.3% against 58.3%, and did much better on AI fiction, but it called 70.8% of human news reports AI-written, where Desklib flagged 20.8%.
Tested
Desklib AI text detector vs Fakespot AI text detector: the numbers side by side
| Measure | Desklib AI text detector | Fakespot AI text detector |
|---|---|---|
| Rank on the leaderboard | 2 of 7 | 3 of 7 |
| Balanced accuracy | 76.0% | 65.7% |
| Catches AI text | 58.3% | 66.3% |
| Flags human text | 6.3% | 34.9% |
| Catches humanized AI text | 32.5% | 47.6% |
| AUROC | 0.877 | 0.741 |
| Flags famous public-domain texts | 13.3% | 43.3% |
| Catches Claude Fable 5.1 | 32.5% | 50.0% |
| Catches Claude Haiku 5.5 | 74.6% | 81.0% |
| Catches Claude Opus 5.5 | 62.7% | 65.1% |
| Catches Claude Sonnet 5.5 | 63.5% | 69.0% |
| Flags human text: Academic abstracts | 8.0% | 12.0% |
| Flags human text: Q&A answers | 3.6% | 39.3% |
| Flags human text: Encyclopedia articles | 0.0% | 40.0% |
| Flags human text: Fiction | 0.0% | 12.5% |
| Flags human text: News reports | 20.8% | 70.8% |
| Seconds per 1,000 words | 0.57 | 0.07 |
| Type | Trained classifier | Trained classifier |
| Licence | MIT | Apache-2.0 |
News is where both go wrong
Human news reports were the genre each detector flagged most. Fakespot called 17 of 24 of them AI-written and Desklib 5, every one published before ChatGPT existed. Both also caught nearly all AI-written news, so in news the two seem to respond to the register of the writing more than to who wrote it. Desklib’s 20.8% is low next to Fakespot, but it is still the weak spot in an otherwise careful detector.
Outside news the gap is still wide. Fakespot flagged 40.0% of human encyclopedia articles and 39.3% of human forum answers. Desklib flagged no encyclopedia articles and 1 forum answer out of 28.
Fiction is Fakespot’s best case
Fakespot caught 67.7% of AI fiction passages against Desklib’s 20.8%, and flagged 12.5% of human fiction against none. Balanced accuracy, the average of the share of AI caught and the share of human text passed, was 77.6% for Fakespot on fiction and 60.4% for Desklib. In every other genre Desklib scored higher.
Famous novels fared worse with Fakespot. Of 30 famous public-domain texts it flagged 13, including the openings of Pride and Prejudice, Moby-Dick and Jane Eyre. Desklib flagged 4, none of them novels.
Fakespot also caught more of the text Desklib finds hardest, with 50.0% of Claude Fable 5.1 texts against 32.5%, and 47.6% of humanized Claude Opus 5.5 texts against 32.5%.
Speed and formatting
Fakespot is a RoBERTa-base model from Mozilla’s Fakespot team and ran at 0.07 seconds per 1,000 words. Desklib is the larger DeBERTa-v3-large model and took 0.57 seconds. Neither changed its verdicts when we removed line breaks.
Which one fits your use
Pick Desklib when the result could count against a person, such as a student essay or a freelance article. Fakespot is the better screen for AI fiction when a false flag costs little, for example when sorting a large pile of story submissions before a human reads them.
Neither is good enough for news writing, where Fakespot’s flags are mostly noise and Desklib still flagged 5 human reports, or for edited AI text. Each detector has its own page, Desklib and Fakespot, and E5-small LoRA vs Fakespot pits Fakespot against a detector with the same news problem.
Questions people ask
Which flags fewer human texts, Desklib or Fakespot?
Desklib. It flagged 8 of 126 human texts (6.3%), Fakespot 44 of 126 (34.9%). Fakespot flagged more human writing in every genre we tested, fiction included, where it flagged 12.5% and Desklib none.
Is Fakespot better than Desklib for fiction?
In our test, yes. It caught 67.7% of AI-written fiction against Desklib's 20.8%, while flagging 3 of 24 human fiction passages. Its balanced accuracy on fiction was 77.6% against Desklib's 60.4%.
Why do both detectors flag human news reports?
We cannot see inside either model. Both caught almost all AI news (100.0% and 90.6%) and both flagged human news more than any other genre, which suggests the news style itself pushes them toward an AI verdict.