Translated SOP and AI Detectors: The Real Risk
Harvard Kennedy School bars AI translation of essays, and detectors already misread non-native English. Why a translated SOP is risky; a safer workflow.
Translated SOP and AI Detectors: The Real Risk
Writing your statement of purpose in your first language and machine-translating it is a rules problem before it is a detection problem. Harvard Kennedy School's 2027-28 guidance says: "You should also not use AI to translate your essays into English." And a 2023 study found AI detectors already misclassify plain non-native English. The safer route is an English draft you wrote yourself.
"Will it get flagged?" is the question applicants ask first, and nobody can answer it for your file. What you can know is what you sign, and how translation changes your text. This page covers both, then gives a workflow. If you are drafting now, a statement of purpose review reads the draft for content, not for its English.
The program page and the study quoted below were checked on October 30, 2026.
What schools say about AI and translation
Harvard Kennedy School's senior director of admissions and financial aid published its rules on September 10, 2026, for the 2027-2028 master's application. The relevant lines, quoted in full:
"All materials submitted as part of your application (essays, resume, etc.) must be your own, original work. AI is only permitted for assisting with edits to your materials and/or identifying issues such as spelling or grammatical errors. … You should also not use AI to translate your essays into English."
And the certification you sign by applying:
"By applying to HKS, you certify that the information in your application is accurate, completely and honestly presented, and that your materials have not been written, rewritten, or translated by anyone other than yourself."
Two things follow. At HKS, a machine translation of your essay breaks the stated rule, whatever a detector says. And the certification covers "anyone other than yourself", so a friend or paid translator is outside it too.
HKS is one school. Other programs word this differently or not at all. The do grad schools check for AI in SOPs guide quotes the written rules of 21 graduate schools, and GradPilot's neutral AI policy directory lists more. If a school publishes no AI rule, the certification inside the application is what you sign, so read it there.
Why non-native English already gets flagged
The study behind this concern is Liang, Yuksekgonul, Mao, Wu and Zou, "GPT detectors are biased against non-native English writers" (arXiv 2304.02819, 2023; published in Patterns). The authors ran seven detectors on 91 TOEFL essays and 88 US eighth-grade essays. On the TOEFL essays, the detectors "misclassified over half of the TOEFL essays as 'AI-generated' (average false positive rate: 61.22%)." The essays flagged by all seven had "significantly lower perplexity."
Perplexity, in one sentence: it measures how predictable each next word is to a language model. Plain, common words in common orders score low, and low is what these detectors read as machine-like. The international students and AI detection bias page covers the study in more depth.
What the study does not show
The study did not test machine-translated text. So "translation gets you flagged" is unknown, not a finding. Our reasoning, labelled as reasoning: a translation engine picks the most likely English for each phrase, which is the low-perplexity pattern the study links to false flags. That is a plausible risk, not a measured one.
The study also found that giving the TOEFL essays to a chatbot with the prompt "Enhance the word choices to sound more like that of a native speaker" cut the false flags. We do not recommend that. HKS permits AI only "for assisting with edits to your materials and/or identifying issues such as spelling or grammatical errors", and its certification says your materials have not been "rewritten … by anyone other than yourself". A tool that rewrites every word choice sits uneasily with both, and it trades your words for a tool's.
Three workflows compared
| Workflow | Fits the HKS rule? | Voice | Technical terms | Detection risk |
|---|---|---|---|---|
| A. Write in your language, machine-translate, submit | No: "not use AI to translate" | Lost: the tool chooses every English word | Can drift: a translated term may not be the one your field uses | Unknown; plausibly higher (see above) |
| B. Write in your language, translate, then rewrite each sentence by hand | Doubtful: the text began as a tool's translation | Partly restored | You can fix them during the rewrite | Unknown |
| C. Draft in English yourself; use tools only to flag spelling and grammar; fix them yourself | Yes: matches the permitted use HKS describes | Yours | Yours, as you use them in class and lab | The study's finding on plain non-native English still applies |
Our view: use workflow C, even if your English draft starts rough. Plain, correct English is not a weakness in an SOP. The countercase: if you think in your first language about your research, write notes in that language first, then write the English draft from the notes without translating sentence by sentence. Notes are planning; the English sentences are still yours.
Constructed example: one paragraph three ways
Three constructed examples, written by us, not real applicant statements. Context (invented): Li Wen, a materials science student, is describing her thesis.
Version 1, an over-smoothed version. We wrote this to show the features to watch for. It is not the output of any specific tool.
Throughout my undergraduate studies, I have developed a profound passion for materials science. My thesis research on battery technology allowed me to gain invaluable experience and deepened my understanding of this dynamic and rapidly evolving field.
Version 2, her own English draft.
In my thesis I make cathode material for sodium-ion battery. I tested 12 samples with different calcination temperature. The best sample keep 91% capacity after 200 cycles. I want to understand why lower temperature give better stability.
Version 3, her draft after she fixed the flagged grammar herself.
In my thesis I made cathode materials for sodium-ion batteries. I tested 12 samples with different calcination temperatures. The best sample kept 91% of its capacity after 200 cycles. I want to understand why the lower temperature gave better stability.
What changed. Version 1 has no material, no number and no question; any student in any field could have sent it. Version 2 has everything a reader needs, with grammar errors. Version 3 fixes only the grammar: past tense (made, kept, gave), plurals (materials, batteries, temperatures), and two small additions ("of its", "the"). Every fact, term and number stayed. That is the edit HKS describes as permitted: "identifying issues such as spelling or grammatical errors." The 91% figure is invented for the example.
The simple English SOP guide has more sentence-level rules for this kind of edit, and the technical SOP vs AI detectors guide covers why specific technical sentences are the ones to keep.
If you already translated your draft
- Read the rule you will sign. If your program bans AI translation, the translated text cannot be submitted as it is.
- Rewrite the key paragraphs in English from memory, without the translation open. Start with the paragraph about your research.
- Restore your terms. Check every technical word against your course notes or papers you have read. Use the term your field uses, not the one the tool chose.
- Put the numbers back. Translation does not remove facts, but a smoothed sentence can replace them with adjectives. Each claim needs its sample size, result or date.
- Read it aloud. If a sentence sounds like nobody you know, rewrite it as you would say it to your professor.
Someone else offering to "polish" or translate it for you raises the same certification issue; the SOP writing services and attestation guide covers that.
Questions about translated SOPs
Is it allowed to use a translation app for my SOP?
It depends on the program. Harvard Kennedy School says "You should also not use AI to translate your essays into English." If your program publishes nothing specific, the application's own certification is the rule.
Will a translated SOP be detected as AI?
Unknown. The 2023 study did not test translation. It did show that detectors misclassify plain non-native English, so even an English-first draft can be flagged.
Can I write notes in my language first?
Notes in your own language are planning, not a translated essay. Write the English sentences yourself from the notes.
Can I use a tool to fix grammar?
At HKS, yes: AI may be used for "identifying issues such as spelling or grammatical errors." Check your program's wording.
Applying to US colleges instead? The college essay too formal guide covers translation in an undergraduate essay. The same certification logic applies to letters: see the rules on translating a recommendation letter.
Get a free review of your statement of purpose
Get a free Quick Review of your statement of purpose against GradPilot's general MS statement of purpose rubric. It reads your purpose, academic goals, program fit, evidence of preparation and voice, so you can see whether your own English draft says enough. More guides are in the graduate school essays hub.
Sources
Retrieved October 30, 2026.
- Harvard Kennedy School, Guidelines for using artificial intelligence responsibly in the HKS application process (September 10, 2026; 2027-2028 master's application): own-work rule, permitted editing use, AI translation bar, certification text.
- Liang, Yuksekgonul, Mao, Wu and Zou, GPT detectors are biased against non-native English writers (arXiv 2304.02819, 2023; PDF): 91 TOEFL essays, seven detectors, 61.22% average false positive rate, lower perplexity in unanimously flagged essays, word-choice enhancement result.
Get feedback on your statement of purpose
Paste your draft. The free Quick Review shows what to fix against an admissions rubric before you submit.