OpenAI RoBERTa detector vs TMR, which AI detector is more accurate?
Neither is accurate enough to use. The OpenAI RoBERTa detector edges ahead on balanced accuracy, 52.4% against 50.0% for TMR, and it is the less harmful choice because it flagged 4.8% of human texts against 83.3%. But it caught only 9.5% of AI texts, while TMR's 83.3% catch rate comes from flagging nearly every multi-paragraph text.
Tested
OpenAI RoBERTa detector vs TMR (Target Mining RoBERTa): the numbers side by side
| Measure | OpenAI RoBERTa detector | TMR (Target Mining RoBERTa) |
|---|---|---|
| Rank on the leaderboard | 6 of 7 | 7 of 7 |
| Balanced accuracy | 52.4% | 50.0% |
| Catches AI text | 9.5% | 83.3% |
| Flags human text | 4.8% | 83.3% |
| Catches humanized AI text | 0.0% | 81.0% |
| AUROC | 0.579 | 0.605 |
| Flags famous public-domain texts | 3.3% | 93.3% |
| Catches Claude Fable 5.1 | 3.2% | 88.1% |
| Catches Claude Haiku 5.5 | 20.6% | 81.0% |
| Catches Claude Opus 5.5 | 2.4% | 83.3% |
| Catches Claude Sonnet 5.5 | 11.9% | 81.0% |
| Flags human text: Academic abstracts | 0.0% | 16.0% |
| Flags human text: Q&A answers | 10.7% | 100.0% |
| Flags human text: Encyclopedia articles | 4.0% | 100.0% |
| Flags human text: Fiction | 4.2% | 100.0% |
| Flags human text: News reports | 4.2% | 100.0% |
| Seconds per 1,000 words | 0.07 | 0.09 |
| Type | Trained classifier | Trained classifier |
| Licence | MIT | MIT |
One says human, the other says AI
The bottom two on our leaderboard err in different directions. The OpenAI RoBERTa detector, built in 2019 for GPT-2 output, says “human” to nearly everything. It caught 9.5% of AI texts, 2.4% of Claude Opus 5.5 texts and no humanized ones. TMR says “AI” to nearly everything with paragraph breaks. It flagged 83.3% of human texts, including every human news report, encyclopedia article, Q&A answer and fiction passage.
Their catch rates per Claude model show the same thing from the other side. The OpenAI detector ranged from 2.4% on Opus to 20.6% on Claude Haiku 5.5. TMR ranged from 81.0% to 88.1%, almost flat, likely because its verdicts depend so little on who wrote the text.
On 30 famous public-domain texts they agreed on one: Genesis 1, which every detector flagged. It was the only classic the OpenAI detector flagged. TMR flagged 28.
AUROC, speed and formatting
Both are RoBERTa-base classifiers using a cut-off of 0.5, and both were quick, at 0.07 and 0.09 seconds per 1,000 words. Their AUROC values, 0.579 for the OpenAI detector and 0.605 for TMR, sit close together and near chance. TMR’s scores separate AI from human text slightly better, but its fixed threshold throws that away.
Formatting moves TMR far more than the OpenAI model. With line breaks replaced by spaces, TMR flagged 23.8% of human texts and caught 38.7% of AI texts. The OpenAI detector barely moved, to 3.2% of human texts flagged and 7.9% of AI texts caught. The paragraph break guide covers all seven detectors.
What to use instead
Neither belongs in any decision about whether someone used AI. If a free online checker runs one of them, its answer mostly reflects the detector’s bias: “human” from the OpenAI model, “AI” from TMR for any text with paragraphs. Of the detectors we tested, Desklib flagged the fewest human texts while still catching a real share of AI writing, 6.3% and 58.3%, and Fast-DetectGPT caught the most AI text among detectors that pass most human writing. Fast-DetectGPT vs TMR compares TMR with the top of the leaderboard.
Questions people ask
Which is worse, the OpenAI RoBERTa detector or TMR?
They ranked sixth and seventh of seven. The OpenAI detector misses almost all AI text and TMR flags almost all human text, so each is wrong in a different way.
Did either detector catch humanized AI text?
TMR flagged 81.0% of humanized Claude Opus 5.5 texts, about as many as it flags of anything. The OpenAI RoBERTa detector flagged none.
Why does TMR have a higher AUROC than the OpenAI detector?
Its scores rank texts slightly better, 0.605 against 0.579, so its problem lies partly in where its threshold sits. Both are still close to the 0.5 of guessing.