# Desklib vs E5-small LoRA, which AI detector is more accurate?

Source: https://toolbenchmarks.com/compare/desklib-vs-e5-lora/  
Tested: 30 October 2026. Publisher: Tool Benchmarks. Data licence: CC BY 4.0.

> Desklib is the better detector for almost everyone. It flagged 6.3% of human texts in our test, while the E5-small LoRA detector flagged 50.0%, exactly half. E5 caught slightly more AI text, 61.1% against 58.3%, but its AUROC of 0.582 against Desklib's 0.877 shows that its scores barely separate AI writing from human writing.

| Measure | Desklib AI text detector | E5-small LoRA detector |
|---|---|---|
| Balanced accuracy | 76.0% | 55.6% |
| Catches AI text | 58.3% | 61.1% |
| Flags human text | 6.3% | 50.0% |
| Catches humanized AI text | 32.5% | 47.6% |
| AUROC | 0.877 | 0.582 |
| Flags human text: Academic abstracts | 8.0% | 36.0% |
| Flags human text: Q&A answers | 3.6% | 75.0% |
| Flags human text: Encyclopedia articles | 0.0% | 52.0% |
| Flags human text: Fiction | 0.0% | 12.5% |
| Flags human text: News reports | 20.8% | 70.8% |

## Where E5's extra catches come from

On paper the E5-small LoRA detector wins on catch rate. It caught 61.1% of AI texts to Desklib's 58.3%, 49.2% of Claude Fable 5.1 texts to 32.5%, and 47.6% of the Claude Opus 5.5 texts we ran through a humanizer, where Desklib caught 32.5%.

By kind of writing, E5 caught more AI text than Desklib in forum answers, encyclopedia articles, fiction and news, and in each of those it also flagged more human writing. It called 21 of 28 human Q&A answers AI-written, where Desklib flagged 1. It flagged 13 of 25 human encyclopedia articles; Desklib flagged none. The likely reading is that E5 reacts to the style of a genre, so it flags AI and human text in the same places.

AUROC, which measures how well a detector's scores rank AI text above human text at any threshold, agrees. Desklib's 0.877 was the best of the seven and E5's 0.582 was second to last.

## Famous texts and formatting

Desklib flagged 4 of 30 famous public-domain texts and E5 flagged 6. They agreed on Genesis 1, the US Constitution and The Art of War. E5's other three were all novel openings, Alice's Adventures in Wonderland, Pride and Prejudice and Dracula, while Desklib passed every novel and flagged the Declaration of Independence instead.

Removing line breaks changed nothing for either detector, so both give the same verdict however the text was pasted.

## Which one to use

E5's advantage is speed. At 0.03 seconds per 1,000 words it was the fastest detector we ran, and it is the smallest model in the test. Desklib, the largest trained classifier here, took 0.57 seconds, which is still quick on a laptop.

For checking whether someone else wrote a text, use Desklib. A detector that flags 63 of 126 human texts cannot back up an accusation, and E5 has no use we can recommend.

Desklib has limits of its own. It caught 20.8% of AI fiction and 32.5% of Fable 5.1 texts, so a pass on a story or on Fable output proves little. The [Desklib page](/ai-detectors/desklib/) and the [E5-small LoRA page](/ai-detectors/e5-lora/) have the full breakdowns, and the [false positives page](/false-positives/) shows how often each detector flagged genuine human writing.

## Questions people ask

### Which catches more AI text, Desklib or E5-small LoRA?

E5-small LoRA, by a small margin, 61.1% against 58.3%. It also flagged 50.0% of human texts against Desklib's 6.3%, so a flag from E5 tells you much less.

### Is E5-small LoRA faster than Desklib?

Yes. It took 0.03 seconds per 1,000 words on an Apple M4 Pro, the fastest of the seven detectors we tested. Desklib took 0.57 seconds.

### Do Desklib and E5-small LoRA flag the same famous texts?

Partly. Both flagged Genesis 1, the US Constitution and The Art of War. E5-small LoRA also flagged the openings of Alice's Adventures in Wonderland, Pride and Prejudice and Dracula, and Desklib also flagged the Declaration of Independence.