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

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

> Fast-DetectGPT. With a threshold tuned on our calibration split, it caught more AI text than the E5-small LoRA detector, 68.1% against 61.1%, and flagged far fewer human texts, 15.9% against 50.0%. E5 is quicker, at 0.03 seconds per 1,000 words against 0.99, but its AUROC of 0.582 shows its scores barely separate AI from human writing.

| Measure | E5-small LoRA detector | Fast-DetectGPT (Qwen2.5-1.5B) |
|---|---|---|
| Balanced accuracy | 55.6% | 76.1% |
| Catches AI text | 61.1% | 68.1% |
| Flags human text | 50.0% | 15.9% |
| Catches humanized AI text | 47.6% | 47.6% |
| AUROC | 0.582 | 0.815 |
| Flags human text: Academic abstracts | 36.0% | 40.0% |
| Flags human text: Q&A answers | 75.0% | 21.4% |
| Flags human text: Encyclopedia articles | 52.0% | 8.0% |
| Flags human text: Fiction | 12.5% | 0.0% |
| Flags human text: News reports | 70.8% | 8.3% |

## Speed against false positives

E5-small LoRA is the smallest and fastest model in the test, scoring 1,000 words in 0.03 seconds on an Apple M4 Pro. Fast-DetectGPT runs the Qwen2.5-1.5B language model over every token and took 0.99 seconds, the slowest of the seven. For a single essay both finish in about a second or less.

The slower method makes far fewer false accusations. Fast-DetectGPT flagged 2 of 24 human news reports, where E5 flagged 17, and 2 of 25 human encyclopedia articles, where E5 flagged 13. On forum answers it flagged 21.4% against 75.0%. Its AUROC of 0.815 against 0.582 says the same across all thresholds.

One caveat applies. E5 used the standard probability threshold, while Fast-DetectGPT's threshold was chosen on our 144 calibration texts, which gave it a small advantage. The AUROC comparison does not depend on that choice.

## Short and long texts

Text length splits them sharply. Fast-DetectGPT caught 74.2% of AI texts under 200 words, its best length band, and E5 caught 30.9%. On texts of 300 words or more E5 pulled ahead, 81.0% against 69.0%, but it also flagged 64.3% of human texts that long, against Fast-DetectGPT's 14.3%. If you check short answers or paragraphs, E5 misses most AI text while still flagging 39.5% of short human texts.

## Where E5 holds up

Academic abstracts trouble both. Fast-DetectGPT flagged 40.0% of human abstracts and E5 36.0%, though E5 caught only 21.0% of AI abstracts against 84.0%. On Claude Fable 5.1 and Opus 5.5 E5 caught a little more, 49.2% and 64.3% against 46.0% and 63.5%, and both caught 47.6% of humanized Opus texts.

On 30 famous public-domain texts E5 was more careful, flagging 6 against Fast-DetectGPT's 17. Fast-DetectGPT was the only detector to flag Hamlet's soliloquy and The Federalist No. 10.

Fast-DetectGPT is also the one of the two that reacts to formatting. With line breaks removed it caught 57.9% of AI texts instead of 68.1%, while E5 gave the same verdicts.

## Is E5's speed worth it?

For nearly every use, no. Pick Fast-DetectGPT. E5's speed does not matter at the scale most people work at, and a detector that flags 50.0% of human texts cannot support a judgment about anyone. Neither is safe on academic abstracts; [Desklib vs Fast-DetectGPT](/compare/desklib-vs-fast-detectgpt/) covers the detector that flagged 2 of 25 there. Read the [Fast-DetectGPT](/ai-detectors/fast-detectgpt/) page for how its threshold was set, and the [E5-small LoRA](/ai-detectors/e5-lora/) page for its genre results.

## Questions people ask

### Is Fast-DetectGPT better than E5-small LoRA?

Yes, on almost every measure. Its balanced accuracy was 76.1% against 55.6%, and it flagged 20 of 126 human texts where E5-small LoRA flagged 63.

### Which one is better on short texts?

Fast-DetectGPT. It caught 74.2% of AI texts under 200 words, E5-small LoRA 30.9%.

### Is there anything E5-small LoRA does better?

Speed, and famous texts. It flagged 6 of 30 famous public-domain texts against Fast-DetectGPT's 17, and it caught slightly more Claude Fable 5.1 texts, 49.2% against 46.0%.