Signals for teachers marking work in English in the Balearic Islands - September 2026
Read this first, because it changes how you use everything below. Your students write English as a second or third language. That is precisely the group automated detectors get wrong: in the best-known study, seven detectors flagged honest essays by non-native writers as AI 61% of the time on average, against almost zero for native writers. Do not run your students' work through an AI detector. The signals on this page work the other way round — they rely on something no tool has and you do: you know how each of your students writes.
Two opposite risks sit in the same room. At one end, a B1 student whose honest writing is short, repetitive and formulaic — the exact statistical profile detectors call artificial. At the other, students in the Balearics who really do write near-native English: a parent from the UK or Germany, years in an academy, a European or CLIL section, gaming and watching everything in English. «This is too good to be theirs» is a bad signal on its own here more than almost anywhere. The evidence is at the bottom of the page ↓.
Prepositions and set phrases: depend of, consist on, in the other hand, according to my opinion, in this sense. Missing dummy subjects: Is important to say, Exist many problems. Agreement: people is, there is many. Verb patterns: I enjoy to read, explain me, say me. Adverb placement: I like very much this film, Always I go. Direct calques: make a party, take a decision, do a photo, have reason, I have 15 years. None of these are things a model produces, and none of them are things a student stops producing overnight.
Word and Grammarly will fix recomend and enviroment. Nothing flags actually used for currently, assist for attend, realize for carry out, sensible for sensitive, success for event, career for degree course, professor for teacher, discuss for argue, sympathetic for nice, molest for bother, large for long, compromise for commitment. So: clean spelling proves nothing, clean grammar proves little, but a text with no false friends in it at all is a text nobody here wrote unaided. Take one page and look only for these.
Handwritten exam scripts and in-class timed writing are the cleanest baseline, followed by Cambridge, Trinity or EOI practice papers done under supervision. Compare vocabulary range and average sentence length; both should be reasonably stable from one piece to the next. Compare only English with English — a student's Catalan or Spanish level tells you nothing useful here.
Phrasal verbs chosen well and not repeated (brush off, single out, water down), collocations that no coursebook drills (a stark contrast, a growing body of evidence, a mounting sense of), and hedging that sounds like a person rather than a rule (arguably, it is tempting to conclude, this is less clear-cut). Set this against the rest of the text: if the argument itself is thin and the phrasing is that fluent, the two do not belong together.
Search the document for -ize, color, center, favorite, analyze, program, traveled, fall for autumn, and date formats like September 4, 2026. A student who has learned British spelling all year does not switch to a flawless, uniform American standard for one essay. Two caveats: some of your students genuinely absorb American English from games, series and YouTube — in them it will be inconsistent, mixed with British forms in the same text, which is the human pattern; and -ize is also valid British usage, so it is the whole cluster that matters, never one item.
Latinate register from cognates where an English speaker would use a plain word: utilise, realise for carry out, the majority of, posteriorly, the aforementioned, in the actuality. Very long sentences held together by which. Anaphoric the same («the company and the workers of the same»). No contractions anywhere — machine translation does not produce don't or it's, while students overuse them and misplace them. And leftovers: an accented character, a place name still in Catalan or Spanish, Mallorca and Majorca alternating within one page.
Browser autocorrect, Word and Grammarly are universal and using them is perfectly honest, and they clean up exactly the visible surface errors: recomend, enviroment, wich, responsability. That is why the signals higher up are about idiom and interference instead. Never use this one on its own. One thing autocorrect does not fix, and which is worth a glance: students here habitually lowercase nationalities, languages, months and weekdays — english, on monday, in january — because that is the rule in Spanish and Catalan. No model ever does.
Select the text and check the word count per paragraph, or just count lines by eye. It is not the length that is suspicious, it is that every paragraph is nearly identical to every other.
They write tourist-brochure English about the Balearics and slide into errors a local student would never make: attributing paella or gazpacho to Mallorcan tradition, mixing up what happens on which island, placing festivals in the wrong town or the wrong month, describing Catalan as "a dialect of Spanish" or Mallorquín as a separate language, inventing beaches, monasteries and dishes that sound right. They are also weak on anything institutional: what a batle does, what the Consell is responsible for, how the school system is actually organised. One concrete, checkable local detail is very hard to fake — and the generated version fills that space with generalities or invents with total confidence.
Search the exact title on Google Scholar or in your library catalogue. If it does not appear, or appears with a different author or year, it is very probably a hallucination.
Rule of thumb: what you know about the student + evidence of process + a conversation. That combination reduces false positives far more than any single signal. And the warning that opens this page bears repeating: do not put a second-language learner's English through an AI detector. See the evidence ↓.
This section gives context and verifiable sources to anyone who wants to go beyond the summary — none of it is needed to use the signals above. Everything here links to its original source.
In 2023, Liang, Yuksekgonul, Mao, Wu and Zou published a study in the peer-reviewed journal Patterns (Cell Press) testing 7 GPT detectors against 91 TOEFL essays written by non-native speakers of English.
The result: an average of 61% false positives, with up to 97.8% of the texts flagged by at least one detector, and almost no false positives on essays by native speakers. The conclusion is blunt: the same honest text is far more likely to be labelled "AI" simply because of how someone writes when English is not their first language.
This is not a caveat for an English department in the Balearics — it describes the entire class. It is the reason this page is built on comparison and conversation rather than on tools.
Sources: The Markup · Business & Human Rights Resource Centre.Turnitin launched its AI detector claiming a 1% false positive rate. Vanderbilt University did the arithmetic on its own numbers: in 2022 alone it had received 75,000 submissions — at that rate, up to ~750 students a year could be wrongly flagged as AI users.
On top of that there was little transparency about how the algorithm worked, and less than 24 hours' notice before the feature was switched on for every client institution. Vanderbilt disabled it. It is the most cited documented case of an educational institution reconsidering its trust in these tools, and part of a wider trend of universities restricting their use since.
Primary source: official statement, Vanderbilt University (Brightspace, August 2023).There are public "humanising" services that rewrite a model's output to make it less detectable, working precisely by altering the statistical patterns — sentence-length variation, simulated imperfections — that detectors rely on. One academic study documented accuracy dropping from 100%→57% and even 30%→16% after this kind of paraphrasing was applied.
On top of that, the accuracy figures detector vendors advertise in 2026 contradict each other depending on which comparison site you read, and much of that material is marketing or affiliate content rather than reproducible independent research. Treat any ranking of detection tools without a clear academic source behind it with a great deal of caution.
Source: arXiv 2303.13408 — the effect of paraphrasing on AI text detection.Every signal on this page is a reason to ask a question, never a finding. In these islands in particular, the profile that most resembles generated text is not a cheat: a student with an English-speaking or German-speaking parent, one who has spent time abroad, one in a European or CLIL section, one who reads and plays and watches everything in English. Their writing is fluent, idiomatic and free of interference because that is genuinely how they write.
The distinction that holds up is not "is this too good?" but "is this a break from this student?". If a near-native student has written like that all year, there is nothing to explain. If the fluency arrived in one submission, there is.
In the last few weeks alone: Anthropic released Claude Fable 5.1 and Mythos 5.1 (1 September 2026), Moonshot AI published Kimi K3 openly (July 2026), Zhipu AI released GLM-5.3 (August 2026), DeepSeek shipped V4-Pro (August 2026), and OpenAI presented GPT-5.6 — with an intermediate tier known as "Terra" (July 2026).
We found no evidence that these versions ship with a built-in "less detectable writing" mode; that capability comes mainly from the external humanising tools mentioned above. What is constant is that the naturalness of their writing improves with every release, so style signals on their own lose reliability over time. What you know about your students, the process behind a piece of work, and a direct conversation do not age the same way: they do not depend on how the model writes.
Sources: Anthropic · CNBC (Kimi K3) · Hugging Face (GLM-5.3) · Yahoo Tech (DeepSeek V4-Pro) · OpenAI (GPT-5.6 / Terra).