E5-small LoRA vs OpenAI RoBERTa detector, which is more accurate?
E5-small LoRA edges it on balanced accuracy, 55.6% against 52.4%, but neither is worth using. E5 flagged 50.0% of human texts and caught 61.1% of AI texts; the OpenAI RoBERTa detector flagged 4.8% and caught 9.5%. Their AUROC scores, 0.582 and 0.579, show that neither one's scores separate AI writing from human writing well.
Tested
E5-small LoRA detector vs OpenAI RoBERTa detector: the numbers side by side
| Measure | E5-small LoRA detector | OpenAI RoBERTa detector |
|---|---|---|
| Rank on the leaderboard | 5 of 7 | 6 of 7 |
| Balanced accuracy | 55.6% | 52.4% |
| Catches AI text | 61.1% | 9.5% |
| Flags human text | 50.0% | 4.8% |
| Catches humanized AI text | 47.6% | 0.0% |
| AUROC | 0.582 | 0.579 |
| Flags famous public-domain texts | 20.0% | 3.3% |
| Catches Claude Fable 5.1 | 49.2% | 3.2% |
| Catches Claude Haiku 5.5 | 66.7% | 20.6% |
| Catches Claude Opus 5.5 | 64.3% | 2.4% |
| Catches Claude Sonnet 5.5 | 64.3% | 11.9% |
| Flags human text: Academic abstracts | 36.0% | 0.0% |
| Flags human text: Q&A answers | 75.0% | 10.7% |
| Flags human text: Encyclopedia articles | 52.0% | 4.0% |
| Flags human text: Fiction | 12.5% | 4.2% |
| Flags human text: News reports | 70.8% | 4.2% |
| Seconds per 1,000 words | 0.03 | 0.07 |
| Type | Trained classifier | Trained classifier |
| Licence | MIT | MIT |
Nearly the same AUROC, opposite verdicts
The OpenAI RoBERTa detector flagged 6 of 126 human texts and caught 48 of 504 AI texts. E5-small LoRA flagged 63 human texts and caught 308 AI texts. One says “human” to almost everything, the other says “AI” to half of all human writing.
Yet their AUROC scores are almost the same, 0.582 for E5 and 0.579 for the OpenAI model, the two lowest in our test. AUROC ignores the threshold and asks only whether AI texts get higher scores than human ones. Both models barely manage that, so moving either threshold would trade one kind of mistake for the other without making the detector useful.
Where each one fails
The OpenAI detector was released in 2019 to spot GPT-2 output and misses almost all Claude text. Its best result was Claude Haiku 5.5, at 20.6%, and it flagged none of the 126 humanized Claude Opus 5.5 texts. E5 caught 47.6% of those.
E5’s failures fall on human writing. It flagged 75.0% of human forum answers and 70.8% of human news reports. Fiction was its safest genre, at 12.5%.
Forum answers are bad for both. The OpenAI detector flagged 3 of 28 human answers there and caught none of the AI answers, so every flag it gave in that genre was wrong.
Famous texts follow the overall pattern. The OpenAI model flagged only Genesis 1, which E5 also flagged. E5’s other five were three novel openings, the US Constitution and The Art of War.
Formatting and speed
The OpenAI detector moved a little when we removed line breaks, to 3.2% of human texts flagged and 7.9% of AI texts caught. E5 gave the same verdicts either way. Both are small and fast, at 0.03 and 0.07 seconds per 1,000 words, so speed is no reason to choose either.
Is either worth using?
No. A pass from the OpenAI detector tells you nothing about a modern chatbot’s text, and a flag from E5 is wrong too often to act on. Desklib vs the OpenAI RoBERTa detector shows what a low false positive rate looks like when it comes with a real catch rate. The full results are on the E5-small LoRA and OpenAI RoBERTa pages.
Questions people ask
Which is better, E5-small LoRA or the OpenAI RoBERTa detector?
E5-small LoRA scores slightly higher on balanced accuracy, 55.6% against 52.4%, but both rank near the bottom of our leaderboard, fifth and sixth of seven. We would not use either to judge a text.
Would a different threshold fix either detector?
Probably not. AUROC measures how well a detector's scores separate AI from human text at every possible threshold, and both scored close to 0.5, the level of random guessing, at 0.582 and 0.579.
Which one is faster?
E5-small LoRA, at 0.03 seconds per 1,000 words against 0.07. Both are among the fastest detectors we tested.