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

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

> Fakespot is the less bad of the two. It flagged 34.9% of human texts against 50.0% for the E5-small LoRA detector, caught more AI text, 66.3% against 61.1%, and had a higher AUROC, 0.741 against 0.582. Both flagged 70.8% of human news reports, though, so neither should be used to decide whether a person wrote something.

| Measure | E5-small LoRA detector | Fakespot AI text detector |
|---|---|---|
| Balanced accuracy | 55.6% | 65.7% |
| Catches AI text | 61.1% | 66.3% |
| Flags human text | 50.0% | 34.9% |
| Catches humanized AI text | 47.6% | 47.6% |
| AUROC | 0.582 | 0.741 |
| Flags human text: Academic abstracts | 36.0% | 12.0% |
| Flags human text: Q&A answers | 75.0% | 39.3% |
| Flags human text: Encyclopedia articles | 52.0% | 40.0% |
| Flags human text: Fiction | 12.5% | 12.5% |
| Flags human text: News reports | 70.8% | 70.8% |

## The same problem with news

Both small classifiers fail on news. Each flagged 17 of 24 human news reports, all published before ChatGPT, and each caught almost all AI news, Fakespot 100.0% and E5-small LoRA 94.8%. Encyclopedia writing follows the same pattern: they flagged 40.0% and 52.0% of human Wikipedia excerpts while catching 95.0% and 92.0% of the AI versions. Both detectors flag human writing most in the genres where they catch the most AI text, which suggests they react to the style of news and encyclopedia writing more than to who wrote it.

Fiction is the other point of agreement. Both flagged 3 of 24 human fiction passages.

## Where they part ways

E5-small LoRA's worst genre is Q&A answers. It flagged 21 of 28 human forum answers, Fakespot 11. It also flagged 36.0% of human academic abstracts against Fakespot's 12.0%, while catching fewer AI abstracts, 21.0% against 33.0%.

Long texts are the riskiest for both. For human texts of 300 words or more, Fakespot flagged 57.1% and E5 64.3%, against 34.2% and 39.5% under 200 words. A long human text is more likely than not to be called AI by either one.

On AI text they are closer than the headline suggests. Each caught 47.6% of the humanized Claude Opus 5.5 texts, and on Claude Fable 5.1 they caught 50.0% and 49.2%. Fakespot's lead comes mostly from Claude Haiku 5.5, where it caught 81.0% against 66.7%.

## Famous texts go the other way

E5-small LoRA flagged 6 of 30 famous public-domain texts and Fakespot 13. Fakespot flagged openings of novels such as Moby-Dick, Jane Eyre and Little Women, along with The Wealth of Nations and On the Origin of Species. E5 flagged a few novel openings too, but far fewer texts overall.

Neither detector changed its verdicts when we removed line breaks. Fakespot cleans its input in a way that drops them, and E5's tokenizer ignores them.

## If you have to pick one

Fakespot makes fewer mistakes. Neither should be used for news or long texts, or to judge a person's work at all, because a flag from either is wrong too often. [Desklib vs Fakespot](/compare/desklib-vs-fakespot/) shows a detector that flagged 6.3% of human texts. Per-genre tables for each are on the [E5-small LoRA](/ai-detectors/e5-lora/) and [Fakespot](/ai-detectors/fakespot/) pages.

## Questions people ask

### Which is more accurate, E5-small LoRA or Fakespot?

Fakespot, with a balanced accuracy of 65.7% against 55.6%. Both still flag a large share of human writing, 34.9% and 50.0%.

### Which flags fewer famous texts?

E5-small LoRA. It flagged 6 of 30 famous public-domain texts and Fakespot 13, one of the few measures where E5 comes out ahead.

### Which one is faster?

E5-small LoRA, at 0.03 seconds per 1,000 words, the fastest in our test. Fakespot took 0.07 seconds, so both are effectively instant.