Direct answer first. Paste Arabic into TextSight and we will tell you we cannot score it. The system identifies the script, reports the language as Arabic, and marks the result unsupported rather than returning a number the model was never trained to produce. That is a deliberate choice, and on this page you can see the evidence behind it.
The evidence is not comfortable reading for any detector. In the largest Arabic study of its kind, 16,400 samples across 10 rewriting models, the best commercial detector scored 92% on unmodified Arabic and 12% once the human articles had been lightly polished. A number that moves 80 points on an editing pass is not a measurement. It is a coin flip with a decimal point.
Checkable in about ten seconds, which is the point: this is a product behaviour, not a marketing position.
The scoring endpoint runs a script check before anything else. Arabic script has its own Unicode range, so text in it is identified as Arabic and the response carries language_supported: false. The interface surfaces that rather than presenting a percentage as if it meant something.
This matters because the alternative is what most tools do. A detector built for English will happily return 87% on an Arabic paragraph, because the model produces a number for any input you give it. The number is meaningless, but nothing on screen says so, and someone will act on it.
We would rather decline than hand you a confident answer from a model that was never trained for the language. The same applies to Chinese, Japanese, Korean, Hindi and Russian: named, then declined.
There is now a serious Arabic-specific study, and it is the most useful thing anyone can hand you on this subject.
Researchers built two Arabic datasets. The first held 800 articles, half AI-generated and half human-authored, used to rank 14 large language models and commercial AI detectors on straightforward Arabic detection. The eight best were carried into the real test.
The second dataset, Ar-APT, took 400 human-authored Arabic articles and had 10 different LLMs lightly polish them across four settings, producing 16,400 samples. Every one of those articles was written by a person. The only question was whether a detector would still say so after an editing pass.
It largely would not. The best commercial detector reached 92% accuracy on unpolished Arabic and fell to 12% on articles polished by Mistral or Gemma-3. The best LLM-based detector, Claude-4 Sonnet, went from 83.51% to 57.63% after LLaMA-3 polishing. The authors' conclusion is that all the detectors tested "incorrectly attribute a significant number of articles to AI."
| Detector | Unpolished Arabic | After light polishing |
|---|---|---|
| Best commercial detector | 92% | 12% |
| Claude-4 Sonnet (best LLM) | 83.51% | 57.63% |
Source: "AI Text Detectors and the Misclassification of Slightly Polished Arabic Text" (arXiv:2511.16690). Figures as reported in the paper's abstract.
It is the normal case, not an edge case. A student runs a paragraph through an assistant to fix grammar. A journalist tightens a paragraph. A researcher asks a model to smooth an abstract. None of that is AI authorship, and all of it is now routine.
A detector that flags that work as machine-written is not catching cheating; it is producing false accusations against people who wrote their own material. In Arabic the published figures put that failure at a scale nobody should be acting on.
A large share of it, because the language of submission across the region is frequently English rather than Arabic.
Higher education across the Gulf runs substantially in English. Engineering, medicine, business and computer science programmes at universities in the UAE, Saudi Arabia, Qatar and Egypt commonly set coursework, theses and journal submissions in English, and a great deal of regional freelance, marketing and technical writing is produced in English for international clients.
That work we can score, with the same detector, the same sentence-level highlights and the same published false-positive benchmark as anywhere else. We also have country pages that go into the institutional detail for the UAE, Saudi Arabia and Egypt.
If Arabic is your first language and you write academic English, you are in the group detectors treat worst. Second-language English has lower perplexity and lower lexical variance than native prose, which is the same statistical signature detectors read as machine-written. A Stanford study measured an average 61.3% false positive rate across seven detectors on non-native TOEFL essays, against near-zero on comparable native writing.
We include second-language academic prose in training to reduce that, and no detector eliminates it. Our own published false-positive measurement is 1,180 academic papers at 5.85%, with the per-document dataset downloadable so you can check the arithmetic rather than take the figure.
Sources: Liang et al. (2023), Patterns (Cell Press); our published benchmark.
We are not the tool for it today, so here is what we would actually tell a colleague to do.
Given a documented 92% to 12% swing on an editing pass, no Arabic detector output should carry a disciplinary or commercial decision on its own. If an institution is using one that way, the Ar-APT paper is the document to put in front of them.
Drafts, revision history, notes and outlines are language-independent and cannot be manufactured by any rewriter. In Arabic, where the detection signal is demonstrably weak, that evidence is not just stronger than a score, it is most of what should count.
A lot of disputes described as "Arabic" are really about an English submission written by an Arabic speaker. That is the case we handle, and it is worth establishing which one you actually have before arguing about tools.
Including about our own roadmap, because a date we miss is worse than no date.
We do not detect Arabic today. Not partially, not "with reduced accuracy". The endpoint reports it as unsupported, and any page telling you otherwise about our product is wrong.
We are not announcing a date. Arabic support is something we want, and we will say it is here when it is here and not before. If you are reading this page after that changes, it will say so plainly rather than in a footnote.
We do not translate. Pasting Arabic to get an English score is not a workflow we support, and it would not be reliable if we did: machine-translated documents scored about 75% accuracy with roughly 11% false positives in the 14-tool Weber-Wulff study, against about 2% on ordinary human text.
We publish no headline accuracy percentage, in any language. What we publish is a false-positive rate with its corpus attached, and accuracy in bands tied to length.
Source for the translation figures: Weber-Wulff et al. (2023), 756 tests across 14 tools.
Not today. Paste Arabic and the system identifies the script, reports the language as Arabic and marks the result unsupported rather than returning a score. That is deliberate: a model trained for English will still output a confident-looking number for Arabic input, and that number means nothing. We would rather decline than let someone act on it. English text, including English written by Arabic speakers, we score normally.
Much less accurate than their marketing suggests, and there is now Arabic-specific evidence. A study built on 16,400 samples found the best commercial detector scoring 92% on unmodified Arabic and 12% on the same human-written articles after light polishing by an LLM. The best LLM-based detector fell from 83.51% to 57.63%. The authors concluded that all the detectors tested incorrectly attribute a significant number of articles to AI.
The same mechanism as in English, amplified by less training data. Detectors read how predictable and how varied the writing is, and any editing pass that smooths the prose moves those properties toward the machine-written range. The Ar-APT study isolated exactly this: 400 genuinely human-authored Arabic articles, lightly polished by 10 different models, with detector accuracy collapsing as a result. Light editing is normal behaviour, not evidence of authorship.
You can paste it, and we will tell you it is unsupported rather than give you a rough idea. There is no useful "rough idea" here: the published spread on Arabic runs from 92% to 12% depending on whether someone ran a grammar pass. A number produced under those conditions is not a weak signal, it is noise with a percentage sign on it.
Statistically the risk is higher, and it is an industry-wide problem rather than one tool's defect. Second-language English tends to have lower perplexity and lower lexical variance than native prose, which is the signature detectors read as machine-written. Stanford measured an average 61.3% false positive rate across seven detectors on non-native TOEFL essays. We include second-language academic prose in training to reduce it, publish our own false-positive rate at 5.85% on 1,180 papers with the dataset attached, and surface a low-confidence flag on borderline text rather than rounding it to a verdict.
No, and it can make things worse. Machine-translated documents scored around 75% accuracy with roughly 11% false positives in the 14-tool Weber-Wulff study, against about 2% on ordinary human text. Translated work gets flagged more often regardless of who wrote it, so the translation step introduces exactly the risk you were trying to measure.
We are not putting a date on it. Arabic support is something we want, and announcing a quarter we then miss is worse than saying nothing. What we will do is state it plainly on this page when it ships, rather than quietly widening what the detector claims. Until then, the honest answer is that Arabic is identified and declined.
AUS, Khalifa, UAEU and NYU Abu Dhabi, where the medium of instruction is English.
See the UAE page →The Saudi institutional landscape and the Vision 2030 shift toward English output.
See the Saudi page →Egyptian universities, Cairo tech and the English-language business shift.
See the Egypt page →The second-language problem in full, and the protocol when a flag lands on real work.
Read the guide →1,180 academic papers, 5.85% measured rate, dataset downloadable so you can check it.
Check our numbers →Cites the published research rather than arguing about a percentage.
Use the template →If the work in question is in English, including English written by Arabic speakers, you can check it now: sentence-level highlights, 3 checks a day free, no signup and no card.