Desklib vs TMR, which AI detector is more accurate?
Desklib. TMR caught more AI text, 83.3% against 58.3%, but it flagged 83.3% of human texts as well, against Desklib's 6.3%. TMR flagged all 101 human texts in our test that had more than one paragraph; Desklib flagged 5.9% of them. TMR's balanced accuracy of 50.0% is what a detector that flags everything would score.
Tested
Desklib AI text detector vs TMR (Target Mining RoBERTa): the numbers side by side
| Measure | Desklib AI text detector | TMR (Target Mining RoBERTa) |
|---|---|---|
| Rank on the leaderboard | 2 of 7 | 7 of 7 |
| Balanced accuracy | 76.0% | 50.0% |
| Catches AI text | 58.3% | 83.3% |
| Flags human text | 6.3% | 83.3% |
| Catches humanized AI text | 32.5% | 81.0% |
| AUROC | 0.877 | 0.605 |
| Flags famous public-domain texts | 13.3% | 93.3% |
| Catches Claude Fable 5.1 | 32.5% | 88.1% |
| Catches Claude Haiku 5.5 | 74.6% | 81.0% |
| Catches Claude Opus 5.5 | 62.7% | 83.3% |
| Catches Claude Sonnet 5.5 | 63.5% | 81.0% |
| Flags human text: Academic abstracts | 8.0% | 16.0% |
| Flags human text: Q&A answers | 3.6% | 100.0% |
| Flags human text: Encyclopedia articles | 0.0% | 100.0% |
| Flags human text: Fiction | 0.0% | 100.0% |
| Flags human text: News reports | 20.8% | 100.0% |
| Seconds per 1,000 words | 0.57 | 0.09 |
| Type | Trained classifier | Trained classifier |
| Licence | MIT | MIT |
Why TMR’s catch rate means little
TMR caught more AI text than any detector we tested, including 88.1% of Claude Fable 5.1 texts, a model Desklib caught only 32.5% of. Those numbers come with a matching false positive rate. TMR flagged every human news report, encyclopedia article, forum answer and fiction passage in the test split, and 28 of 30 famous public-domain texts, from the Gettysburg Address to Walden.
The cause is formatting. TMR flagged all 101 multi-paragraph human texts. When we replaced the line breaks with spaces and scored the same words again, it flagged 26 of them, and its catch rate on AI text fell to 38.7%. Desklib gave identical verdicts with and without line breaks, flagging 6 of the 101 multi-paragraph human texts both times. The paragraph break guide has the full test.
Academic abstracts, where formatting is not a factor
Abstracts are mostly single paragraphs, so they show TMR without its line-break problem. It flagged 16.0% of human abstracts, against Desklib’s 8.0%, and caught 16.0% of AI abstracts, against Desklib’s 50.0%. With its quirk out of the way, TMR flags human and AI abstracts at the same rate, while Desklib separates them.
The same holds on humanized Claude Opus 5.5 texts. TMR’s 81.0% looks strong next to Desklib’s 32.5%, but TMR flagged 83.3% of genuine human writing too, so a flag from it carries almost no information.
Same benchmark, opposite results
Both detectors have roots in the RAID benchmark. TMR was trained on RAID with a method meant to keep false positives low, and Desklib’s model card says it led the RAID leaderboard when it was released. On our texts, written by current Claude models and by people before ChatGPT, only Desklib kept its false positives low. TMR is the faster model, at 0.09 seconds per 1,000 words against Desklib’s 0.57.
Is TMR ever the right choice?
No. TMR has no practical use as published, and even with line breaks stripped it flagged more human texts than Desklib (23.8% against 6.3%) while catching fewer AI texts (38.7% against 58.3%). Use Desklib.
Desklib is not enough on its own for fiction, where it caught 20.8% of AI passages, or for Fable 5.1 output. More detail is on the Desklib and TMR detector pages.
Questions people ask
TMR catches more AI text than Desklib. Is it better?
No. It caught 83.3% of AI texts and flagged 83.3% of human ones, so its verdict barely depends on who wrote the text. Desklib flagged 6.3% of human texts.
Does removing line breaks fix TMR?
It cuts TMR's false positives to 23.8% of human texts, but its catch rate falls to 38.7%. Desklib gave the same verdicts either way, flagging 6.3% and catching 58.3%.
Which famous texts did TMR pass?
Two of the 30, Hamlet's "To be, or not to be" soliloquy and The Federalist No. 10. Desklib flagged 4 of the 30.