# Desklib vs the OpenAI RoBERTa detector, which is more accurate?

Source: https://toolbenchmarks.com/compare/desklib-vs-openai-roberta/  
Tested: 30 October 2026. Publisher: Tool Benchmarks. Data licence: CC BY 4.0.

> Desklib, by a wide margin. The OpenAI RoBERTa detector flagged slightly fewer human texts, 4.8% against 6.3%, but it caught only 9.5% of AI texts against Desklib's 58.3%, and 2.4% of Claude Opus 5.5 texts against 62.7%. It was built in 2019 to spot GPT-2 output and passes almost all writing from current chatbots.

| Measure | Desklib AI text detector | OpenAI RoBERTa detector |
|---|---|---|
| Balanced accuracy | 76.0% | 52.4% |
| Catches AI text | 58.3% | 9.5% |
| Flags human text | 6.3% | 4.8% |
| Catches humanized AI text | 32.5% | 0.0% |
| AUROC | 0.877 | 0.579 |
| Flags human text: Academic abstracts | 8.0% | 0.0% |
| Flags human text: Q&A answers | 3.6% | 10.7% |
| Flags human text: Encyclopedia articles | 0.0% | 4.0% |
| Flags human text: Fiction | 0.0% | 4.2% |
| Flags human text: News reports | 20.8% | 4.2% |

## Two low false positive rates with different meanings

The OpenAI RoBERTa detector and Desklib are the two most cautious detectors in our test. The OpenAI model flagged 6 of 126 human texts and Desklib 8, which taken alone makes the OpenAI model look safer.

The catch rates undo that. Desklib caught 294 of 504 AI texts; the OpenAI detector caught 48. A detector that calls almost everything human will always look good on human writing, and AUROC exposes the difference: 0.877 for Desklib, the best in the test, and 0.579 for the OpenAI model, the lowest.

Forum answers show the problem most clearly. The OpenAI detector flagged 3 of 28 human Q&A answers and caught none of the AI-written ones, so every flag it gave in that genre landed on a person. Desklib flagged 1 human answer and caught 50.0% of the AI answers.

## Built for a different generation of models

OpenAI released this RoBERTa-base model in 2019 to show that GPT-2 text could be detected. It learned what GPT-2 output looks like, and the likely reason it misses so much is that Claude writes differently. Its best result on any Claude model was 20.6% of Claude Haiku 5.5 texts. On Opus it caught 2.4% and on Fable 5.1 3.2%, where Desklib caught 62.7% and 32.5%.

On famous public-domain texts the OpenAI model flagged only Genesis 1, which Desklib flagged too, along with three others. The OpenAI detector also moved slightly when we removed line breaks, flagging 3.2% of human texts instead of 4.8%, while Desklib gave identical verdicts.

Its one practical edge is speed, at 0.07 seconds per 1,000 words against Desklib's 0.57. Both are released under the MIT licence.

## Which one to trust

Use Desklib. A pass from the OpenAI RoBERTa detector says nothing about text from a current chatbot, and it still flags some human writing. If a free online checker turns out to run this model, treat its "human" verdicts as meaningless.

Desklib is far from proof either way. It passed most Claude Fable 5.1 texts and most humanized text, and it still flagged 5 of 24 human news reports. The [Desklib](/ai-detectors/desklib/) and [OpenAI RoBERTa](/ai-detectors/openai-roberta/) pages have the full results, and [how accurate are AI detectors](/guides/how-accurate-are-ai-detectors/) puts both in context.

## Questions people ask

### Which AI detector flags the fewest human texts?

The OpenAI RoBERTa detector, at 4.8%, with Desklib next at 6.3%. The OpenAI detector gets there by calling almost everything human, since it caught only 9.5% of AI texts.

### Can the OpenAI RoBERTa detector catch humanized AI text?

No. It flagged 0 of 126 humanized Claude Opus 5.5 texts. Desklib caught 41 of them, or 32.5%.