E5-small LoRA vs MAGE, which AI detector is more accurate?
MAGE, on balance. It flagged far fewer human texts than the E5-small LoRA detector, 19.0% against 50.0%, and had the higher balanced accuracy, 63.7% against 55.6%. E5 caught more AI text, 61.1% against 46.4%, and more humanized text. Neither is accurate enough to judge whether a real document was written by a person.
Tested
E5-small LoRA detector vs MAGE: the numbers side by side
| Measure | E5-small LoRA detector | MAGE |
|---|---|---|
| Rank on the leaderboard | 5 of 7 | 4 of 7 |
| Balanced accuracy | 55.6% | 63.7% |
| Catches AI text | 61.1% | 46.4% |
| Flags human text | 50.0% | 19.0% |
| Catches humanized AI text | 47.6% | 23.8% |
| AUROC | 0.582 | 0.666 |
| Flags famous public-domain texts | 20.0% | 20.0% |
| Catches Claude Fable 5.1 | 49.2% | 28.6% |
| Catches Claude Haiku 5.5 | 66.7% | 66.7% |
| Catches Claude Opus 5.5 | 64.3% | 45.2% |
| Catches Claude Sonnet 5.5 | 64.3% | 45.2% |
| Flags human text: Academic abstracts | 36.0% | 12.0% |
| Flags human text: Q&A answers | 75.0% | 7.1% |
| Flags human text: Encyclopedia articles | 52.0% | 40.0% |
| Flags human text: Fiction | 12.5% | 4.2% |
| Flags human text: News reports | 70.8% | 33.3% |
| Seconds per 1,000 words | 0.03 | 0.39 |
| Type | Trained classifier | Trained classifier |
| Licence | MIT | Apache-2.0 |
Which way each one errs
E5-small LoRA and MAGE sit fourth and fifth on our leaderboard. E5 leans toward calling text AI. It flagged exactly half of the human texts and caught 61.1% of the AI ones. MAGE leans the other way, flagging 19.0% of human texts and catching 46.4% of AI texts.
Which mistake matters depends on what you do with the verdict. If a flag leads to a conversation with a student or a writer, E5’s false positives are the bigger cost. If a pass means a text goes out unchecked, MAGE’s misses are.
Forum answers show the split most sharply. E5 flagged 21 of 28 human Q&A answers; MAGE flagged 2. Both struggle with encyclopedia writing, flagging 52.0% and 40.0% of human Wikipedia excerpts, and with news, at 70.8% and 33.3%.
Same count of famous texts, different texts
Each flagged 6 of the 30 famous public-domain texts we tested. They agreed only on Genesis 1 and the US Constitution. E5’s other four were the openings of Alice’s Adventures in Wonderland, Pride and Prejudice and Dracula, plus The Art of War. MAGE’s were The Communist Manifesto, the Declaration of Independence, The Prince and Civil Disobedience. We cannot see why each model picked the texts it did, but the split between mostly novels and political writing is clear.
Length, editing and speed
Both are weak on short texts. Under 200 words E5 caught 30.9% of AI texts and MAGE 20.6%. A humanizer pass cut MAGE’s catch rate on Claude Opus 5.5 texts from 45.2% to 23.8%, and E5’s from 64.3% to 47.6%.
MAGE is a Longformer model from an ACL 2024 research paper and ran at 0.39 seconds per 1,000 words. E5 is the smallest and fastest model in our test, at 0.03 seconds. Neither changed its verdicts with line breaks removed.
Should you use either?
If you have to choose between these two for checking other people’s writing, MAGE makes fewer false accusations. In practice neither is good enough for that: MAGE still flagged 10 of 25 human encyclopedia articles and missed most AI texts from Claude Opus 5.5 and Sonnet 5.5. The famous texts page shows how all seven detectors handled the classics. Each detector’s own results are on the E5-small LoRA and MAGE pages.
Questions people ask
Which flags fewer human texts, MAGE or E5-small LoRA?
MAGE, at 24 of 126 human texts (19.0%). E5-small LoRA flagged 63 of 126 (50.0%).
Which catches more AI text?
E5-small LoRA, 61.1% against 46.4%. It also caught 47.6% of humanized Claude Opus 5.5 texts, where MAGE caught 23.8%.
Do they flag the same famous texts?
Only two of the same. Each flagged 6 of 30, and both flagged Genesis 1 and the US Constitution. E5-small LoRA's other flags were mostly novel openings; MAGE's were political and philosophical writing.