Fakespot vs OpenAI RoBERTa detector, which is more accurate?
Fakespot is more accurate for anyone checking text from current chatbots. It caught 66.3% of AI texts against 9.5% for the OpenAI RoBERTa detector, and 65.1% of Claude Opus 5.5 texts against 2.4%. The OpenAI detector flagged fewer human texts, 4.8% against 34.9%, but only because it calls nearly everything human.
Tested
Fakespot AI text detector vs OpenAI RoBERTa detector: the numbers side by side
| Measure | Fakespot AI text detector | OpenAI RoBERTa detector |
|---|---|---|
| Rank on the leaderboard | 3 of 7 | 6 of 7 |
| Balanced accuracy | 65.7% | 52.4% |
| Catches AI text | 66.3% | 9.5% |
| Flags human text | 34.9% | 4.8% |
| Catches humanized AI text | 47.6% | 0.0% |
| AUROC | 0.741 | 0.579 |
| Flags famous public-domain texts | 43.3% | 3.3% |
| Catches Claude Fable 5.1 | 50.0% | 3.2% |
| Catches Claude Haiku 5.5 | 81.0% | 20.6% |
| Catches Claude Opus 5.5 | 65.1% | 2.4% |
| Catches Claude Sonnet 5.5 | 69.0% | 11.9% |
| Flags human text: Academic abstracts | 12.0% | 0.0% |
| Flags human text: Q&A answers | 39.3% | 10.7% |
| Flags human text: Encyclopedia articles | 40.0% | 4.0% |
| Flags human text: Fiction | 12.5% | 4.2% |
| Flags human text: News reports | 70.8% | 4.2% |
| Seconds per 1,000 words | 0.07 | 0.07 |
| Type | Trained classifier | Trained classifier |
| Licence | Apache-2.0 | MIT |
Same architecture, different training
Fakespot and the OpenAI detector are built on the same base model, RoBERTa-base, and they ran at the same speed, 0.07 seconds per 1,000 words. What separates them is what they were trained on. OpenAI released its detector in 2019 to spot GPT-2 text. Fakespot’s comes from the Mozilla team that built tools for spotting fake reviews, and its results suggest it learned from text much closer to what today’s chatbots write.
Fakespot caught 66.3% of our AI texts and the OpenAI RoBERTa detector 9.5%. On Claude Opus 5.5, the OpenAI model caught 2.4%; on Q&A answers written by any Claude model it caught none. Humanized text went the same way, with 47.6% caught by Fakespot and 0.0% by the OpenAI detector.
What Fakespot’s catch rate costs
Fakespot pays for its catch rate with false positives. It flagged 34.9% of human texts, and 70.8% of human news reports, all written before ChatGPT existed. The OpenAI detector flagged 4.8%, and its worst genre, Q&A answers, reached only 10.7%.
The OpenAI detector’s low rate comes from calling almost everything human, so it passes human and AI writing alike. Its AUROC of 0.579 shows its scores barely separate AI from human text, where Fakespot reached 0.741. On 30 famous public-domain texts it flagged only Genesis 1, which all seven detectors flagged. Fakespot flagged 13, among them the US Constitution and the opening of Pride and Prejudice.
Who should use either one
Nobody checking modern AI writing should rely on the OpenAI RoBERTa detector. A “human” result from it says almost nothing, and a free web tool built on it would pass most Claude text. Fakespot does catch most raw AI output, but it flagged 44 of 126 human texts, so its verdict needs a second opinion before anyone acts on it.
For that second opinion, a detector that flags less human writing is the better choice. Desklib flagged 6.3% of human texts and caught 58.3% of AI texts (see the Desklib page), and Fakespot vs Fast-DetectGPT compares Fakespot with the detector at the top of our leaderboard.
Questions people ask
Why does the OpenAI detector miss so much AI text?
It was trained in 2019 to spot GPT-2 output, and modern chatbot writing likely looks nothing like what it learned from. It caught 9.5% of our AI texts and none of the humanized ones.
Is Fakespot's false positive rate a problem?
Yes. It flagged 44 of 126 human texts, including 70.8% of human news reports, so a Fakespot flag on its own proves little.
Which is faster?
Neither has an edge. Both took 0.07 seconds per 1,000 words on an Apple M4 Pro.