Spanish AI Detector and AI Detection in Other Languages

Scanner AI is an AI text detector and humanizer: it flags AI-written fragments in GPT, Claude and Gemini output and rewrites them while keeping keywords, headings and links intact. Working across those models makes one thing plain — detection quality is not a property of a tool but of each language it was trained on. This article gathers what detector vendors publish about languages other than English, and marks who measured what.

A Spanish AI detector is not the English tool with a language switch: it needs Spanish data behind it, and its published error rates differ from the English ones. Every number below comes from a detector company's own evaluation, because no independent test of multilingual detection appears in this search result.

Yes, and the vendors' own figures show it working less well than in English. Originality.ai puts this in writing: its English-only Lite and Turbo models "still outperform the current multilingual model", which it reports at 97.8% accuracy and a 2.4% false-positive rate against 99% for English Turbo.

Copyleaks publishes the same asymmetry per language. Its help page gives English 99.98% on human text and 98.40% on AI text, Spanish 99.85% and 98.02%, Portuguese 99.95% and 93.08% — a five-point spread on AI recall between two Romance languages in one table, published by the vendor that chose the samples and the moment of measurement.

Coverage counts are not comparable, because each vendor counts something different. Originality.ai names 30 languages after its Multi Language 2.0.0 release, Pangram advertises more than 20, GPTZero announced nine "most prevalent" languages while publishing a table for 20-plus, and Copyleaks documents accuracy for six. A claimed count says what marketing counted; only a published table says whether anyone measured your language.

Language-specific training is the whole difference. GPTZero describes language-aware pipelines built because "the concept of words, spacing and sentence boundaries vary across various languages and linguistic families" — an English tokenizer applied to Chinese or Arabic measures the wrong units before it measures anything else.

Scale is the second variable. Originality.ai reports evaluating its multilingual model on 127,150 samples across 30 languages, split evenly between 63,575 human-written and 63,575 AI-generated texts from 12 models. One benchmark divided across 30 languages is a thin slice per language.

Spanish is the best-served non-English case because institutions asked for it. Turnitin's press release describes launching AI writing detection for Spanish submissions inside Turnitin Originality, and states the model was then trained to detect GPT-3.5 and GPT-4 content — a reminder that a generator's release date and a detector's training date rarely match.

• Ask which languages the model was trained and evaluated on, and when.
• Treat a dialect claim — Castilian against Latin American Spanish, say — as unverified until a figure comes with it.

The most consequential errors happen in English, to people writing in a second language. Pangram states the mechanism plainly: non-native speakers are "more likely to be unfairly flagged by less accurate AI detectors because of their phrasing". The writing is human; it just reads the way these models learned to expect machines to write.

Phrasly's own FAQ lists what triggers a false positive on its detector, and the list reads as a portrait of second-language academic writing: a formal style with consistent sentence structure, predictable phrasing, standardised technical terminology, and being "a non-native English speaker writing in simple, clear sentences".

• Consistent sentence structure and formal register.
• Common expressions and predictable transitions.
• Standardised terminology in technical topics.
• Simple, clear sentences from a second-language writer.

QuillBot describes the same failure from the other side: formulaic human writing such as academic definitions and legal copy "can score higher than expected for any AI writing detector". None of these traits is evidence of AI use — they are what a pattern classifier cannot tell apart from it.

Translation is the quietest source of confusion here, and the least documented. Machine translation regularises wording: idiom flattens, sentence length evens out, vocabulary narrows towards the most probable option. The output is machine-produced language wrapped around a human's ideas, and a detector cannot separate those layers.

Vendors acknowledge it obliquely rather than measure it. Pangram lists "Detecting Double-Translation" as a use case, and a Spanish-focused tool markets spotting "translated-style" phrasing. Not one page in this search result publishes an error rate for translated text, in any language pair.

• Keep the original-language draft and the translation history.
• Note which passages were machine-translated and which were edited afterwards.
• Expect a translated human text to read as more predictable than the original.

The gap hits the people already at risk. A student who drafts in a first language and translates, or a writer who runs their English through a translation tool, ends up with machine regularity in the text through no misconduct of their own — and with no published figure anywhere that says how often a detector mistakes it for generated writing.

The published tables thin out as languages get smaller, and vendors say so themselves. GPTZero reports ranking first in almost every language it tested except Croatian and Czech, and names low-resource languages as the work still ahead. Its own table gives Czech 96.4% AI recall and Ukrainian 96%, against 100% for Chinese.

Below the documented languages there is no published number at all. Copyleaks' table stops at six; the counts of 20 and 30 elsewhere are lists, not measurements. Outside every table, a percentage on screen is the model extrapolating from what it learned elsewhere.

• Check whether your language appears in a published table, not just a list.
• Check the script and the regional variety, not only the language name.
• Treat an unlisted language as an unmeasured one.

Length comes first, because it decides whether the score means anything. QuillBot requires at least 80 words to scan and states that texts of 300 words or more give more reliable scores: a flagged paragraph is a weaker signal than a flagged essay, in any language.

Then read the score against its language: a 90% in a language with a published error rate is a different object from a 90% in a language the vendor merely lists, and neither identifies an author. Scribbr, selling its own detector, states that no AI detector can provide complete accuracy; QuillBot is blunter — never rely on AI detection alone for decisions affecting someone's career or academic standing.

Detectors do work outside English, and their own numbers show quality falling as a language moves away from their training data. Every figure available today was published by a company selling the tool it measured — the first thing to know before reading one.

• English models beat the same vendor's multilingual ones, by that vendor's figures.
• Translated text has no published error rate anywhere in this search result.
• A listed language is not a measured one; check for a table, not a count.