Can an App Check if Non-Native English Sounds Polite?

In most of the studies behind this piece, the emails that struck native readers as abrupt were judged on directness and softening, not on errors, so a grammar checker cannot tell a non-native English writer whether a message sounds polite. The one experiment that varied spelling and grammar errors found that most of the negative impressions the errors created were reduced once readers knew the sender was from another culture (covered below). What marked the abrupt emails out was a request made too directly, with no softening words around it, no greeting or closing, and no sign that the writer knew they were asking for something. Those habits are often shaped by a first language, and they are invisible to anything that checks spelling and verb agreement.

Most of the tools people reach for check grammar, fluency and formality. A smaller group reads tone. None of them knows who you are writing to unless you say, how big the favour is, or what counts as polite in that person’s culture, and that last layer is where the research places most of the damage. The answer to “does my English text sound natural and polite” splits in two. The first half is the conventional phrases that soften a request, and those can be checked and learned. The second half is whether this reader, in this relationship, will take your words as you meant them. You can only estimate that, and a tool can only estimate it with you.

I co-founded Subtext, a tone checker, so I have an interest in how the last three sections read. They are the only place I make claims about it. Everything before them comes from sources I opened myself, in some cases only the abstract or a page describing the study, as the Sources list says.

Why a correct English message can still read as rude

A message can be grammatically perfect and still fail, because politeness in English lives mostly in the words around the request. Economidou-Kogetsidis studied email requests that Greek Cypriot university students sent to faculty at a university in Cyprus that teaches in English, collected over several semesters¹. The emails were marked by “significant directness”, especially in requests for information, “an absence of lexical/phrasal downgraders”, omitted greetings and closings, and forms of address the author judged inappropriate or unacceptable. Her argument is that such emails can be read as impolite because “they appear to give the faculty no choice in complying with the request and fail to acknowledge the imposition involved”¹. The paper’s title is one of the lines from the data, “Please answer me as soon as possible”, and it shows the pattern in seven words. The writer means to be courteous, uses the word please, and still produces a sentence that sets the reader a deadline.

A downgrader is a small word that lowers the pressure. “I was wondering whether”, “if you have a moment”, “would it be at all possible”. Native writers of English tend to reach for these. Writers coming from languages where directness is the polite default may skip them and add a please instead. In a 2013 study of 4,353 Wikipedia requests, each rated by five US annotators, Danescu-Niculescu-Mizil and colleagues found that a please in the middle of a sentence averaged +0.49 on their politeness scale, while a request that opened with “Please” averaged -0.30². They read an opening please as a direct strategy that can seem insincere. The study does not report the writers’ first language, so it shows how US raters hear a leading please, not who writes one.

Native speakers write directly too, with more ways to soften

Directness is common among native writers as well, and what separates the groups is how much softening they add. Biesenbach-Lucas compared email requests to professors from graduate students who were native and non-native speakers of English at an American university³. She analysed 533 emails that TESOL graduate students sent to a single professor over six semesters, 382 from native and 151 from non-native speakers, and every non-native writer came from an Asian background. Both groups wrote more direct requests and hints than the conventionally indirect forms found in older speech act studies, so the drift is not confined to learners. But native speakers “demonstrate greater resources in creating e-polite messages to their professors than nonnative speakers”³. They had more ways to soften, and they used them more. The groups are unequal in size. The study lists the single professor and the narrow sample of non-native speakers (all from Asian language backgrounds, all in one degree programme) as limitations, so it describes one setting and does not settle the question.

What native readers actually hold against you

Native readers judge the sender as well as the sentence, and they judge the register more harshly than the grammar. In a later perception study, Economidou-Kogetsidis showed the same direct, unmodified student emails to Greek Cypriot learners and to native English lecturers⁴. The two groups had “significantly mismatching perceptions of what is appropriate”. In most cases the native lecturers rated the same email as “significantly more abrupt, less polite and as not adequately acknowledging the imposition involved”, and they rated the personality of the sender “significantly less favorably” than the learners did⁴.

Hendriks ran two online experiments in which native English speakers read email requests written by Dutch learners, with the softening words varied. The 110 raters in the first and 158 in the second rated comprehensibility and the sender’s personality⁵. Underusing those softening words “had a negative effect on participants’ evaluation of the personality of the sender of the e-mail”⁵. The message was understood either way. The effect was on how agreeable the sender seemed, it came from one phrase, and it was small, about a third of a point on a scale of five points.

A 2023 study by the same lead author found the penalty narrower still. It gave 120 Dutch and 131 British readers one of four versions of a request email from a colleague⁶. The curt version (“Can you send me…”) made the sender seem bossier. But a Dutch writer’s curt email was also rated more authoritative and drew more willingness to comply, while an English writer’s curt email was rated less competent. The authors say leaving out softening words “may have a limited but important impact”.

Readers forgive grammar more readily than tone

Readers forgive grammar errors more readily than a curt tone, and Vignovic and Thompson tested whether knowing the sender is from another culture buys any forgiveness⁷. Readers judged an email sender’s conscientiousness, intelligence, agreeableness and trustworthiness, among other traits, after reading messages with spelling and grammar errors, or messages that broke etiquette norms, which the authors defined as short messages lacking a conversational tone. Grammar errors produced negative impressions, and most of those impressions were reduced when readers learned the sender was from a different culture. Etiquette violations were different. The negative attributions “were not significantly mitigated by knowledge that the e-mail sender was from a foreign culture”⁷. A reader will largely excuse your grammar when they know English is your second language. They will not excuse a curt tone in the same way. The paper frames this as attribution, with grammar errors put down to the situation and a curt tone put down to the person.

Phrase problems and situation problems

Thomas drew the line that explains why some of this can be checked and some cannot⁸. She called the inability to understand “what is meant by what is said” pragmatic failure, and split it in two. Pragmalinguistic failure is a matter of “highly conventionalized usage which can be taught quite straightforwardly as ‘part of the grammar’”. Sociopragmatic failure “is much more difficult to deal with, since it involves the student’s system of beliefs as much as his/her knowledge of the language”⁸.

In practice the first kind is the missing “I was wondering if” or the “please answer me as soon as possible”. It is a phrase problem, and a tool or a phrasebook can catch it. The second kind is deciding that a request of one line to a senior colleague needs no softening at all, because in your first language it would not. No phrase fixes that, because the words are fine and the situation was misread.

Learners also notice these two kinds of error differently depending on where they live. Bardovi-Harlig and Dörnyei asked English learners to judge utterances that were either grammatically wrong or contextually inappropriate⁹. Learners studying English in a country where it is spoken rated the pragmatic errors as more severe than the grammatical ones. Learners studying English in their home country rated the grammatical errors as more severe⁹. A learner who studied English in their home country therefore tends to rank grammar errors as the worse kind, while the studies above found native readers reacting more to directness and softening.

What instruction and higher proficiency help with

Instruction and higher proficiency both help with the conventional phrases, and neither closes the whole gap. Usó-Juan collected 110 naturally occurring email requests that Spanish learners of English as a foreign language sent to three faculty members, before and after a period of strategy instruction, and found the requests more appropriate afterwards¹⁰. The abstract reports a before and after comparison, with no control group or delayed test mentioned, so it shows the phrases can be taught and not that the gain lasts. Qin, Jia and Ren analysed 306 request emails from 32 teachers who speak English and 121 learners whose first language was Chinese, French or Spanish¹¹. The learners’ first language had a significant effect on their writing. Learners with higher proficiency used more indirect request strategies, but they had not developed the pragmatic competence to adjust their usage across written contexts. A more advanced learner uses more indirect requests and still does not yet adapt to the situation, which resembles Thomas’s second category⁸.

Why you cannot hear it yourself

Rereading your own message may not catch this, for two reasons. The first is the general one, covered in does this text sound rude. Kruger and colleagues ran five experiments on tone over email and their results suggest that people tend to believe they communicate over email more effectively than they do. Their fourth and fifth experiments suggest the cause is that writers “hear” a statement as they meant it, which makes it hard to see that the audience may not¹². A 2025 study of everyday text messages (347 receivers, 171 senders) and work or education emails (361 receivers, 61 senders) found no sign of misread positive or negative tone¹³. It rated positive or negative tone, not politeness, so it leaves this section’s question open. The second reason is an inference from the studies above, not a finding, and is specific to writing in a second language. The silent voice you hear is reading with your first language’s politeness rules. A request that is perfectly courteous in Greek or Dutch sounds courteous to you in English too, because the judgement is being made by the same ear.

This is why the perception gap in the second Economidou-Kogetsidis study is so wide⁴. The learners rated their own kind of email as appropriate because, by the standards they had grown up with, it was. The lecturers rated it as abrupt because, by theirs, it was. Each group read the email accurately by its own standards.

So asking whether your English sounds polite means asking how it sounds to a reader whose expectations you cannot check, which is what a second reader is for. The reader’s side of the same gap, working out what a message you received was meant to say, is in what does this text mean.

What grammar and fluency tools check, in their own words

The large writing tools mostly promise correctness and fluency, and some promise formality. Each description below quotes the tool’s own page, so nothing here is a feature I inferred.

LanguageTool’s homepage describes a checker for “grammatical errors and spelling mistakes across languages”, says it “masters more than 30 languages and dialects”, and offers a paraphraser that rephrases “to be more formal, fluent, simple or concise”¹⁴. Its product menu says the checker “helps you find the right tone”; nothing on the page says it detects politeness.

DeepL Write, by its own blog, supports English, German, French, Spanish, Italian, Portuguese, Chinese (Simplified), Japanese and Korean, and, for French and Spanish at least in that post, lets you pick a style (Business, Academic, Simple, Casual) and a tone (Friendly, Diplomatic, Confident, Enthusiastic)¹⁵. Both are settings you apply, and the tool does not detect whether your draft was diplomatic.

Wordtune promises to “Switch between casual and formal tones with the click of a button”, says it “can only write in English”, and claims its translation feature “will make you sound like a native-English speaker with just a click”¹⁶. The first two are checkable. The third is marketing.

QuillBot’s paraphraser, per its own blog, offers two free modes (Standard and Fluency) and nine preset modes for paying users, works “in over 25 languages”, and is pitched at people writing “in a first, second, or third language”¹⁷. Again, the modes are instructions you give it.

All four can make a sentence more fluent and more formal. Formal and polite are not the same thing, and the research above is about the difference. A formal sentence that still gives the reader no choice is still the kind of email the 2011 paper took its title from.

What the tone tools check

Two tools claim to read tone before you send and a third rewrites toward it, and their claims are narrower than the question you are asking. A fourth, QuillBot, scores tone after a paraphrase.

Grammarly’s tone page shows a set of labels for the overall tone of a text, such as Formal, Confident, Accusatory and Worried, with no statement about languages other than English, about limits, or about who the reader is¹⁸; what the detector reads is covered in AI tone detection. How that compares with a reader built for messages is in Subtext against Grammarly.

Tonely, on its App Store listing, is “the world’s first On device Ethical AI-powered iMessage extension and keyboard designed to help you understand how your words may come across before you press send”, runs entirely on the device, lists English as its only language, and is free to download with a subscription of $9.99 a month or $99.99 a year¹⁹. Processing on the device is a privacy advantage, and English only is a limit for anyone who also messages in their first language.

ConfiText’s homepage says it “rewrites your messages so they come across clear, confident, and exactly how you meant them”²⁰. It names no languages and makes no claim about reading how a message will land before rewriting it.

QuillBot’s help page, last updated on 10 June 2026, describes Premium tone settings that “help you see how the tone of your text changes after paraphrasing”²¹. A panel charts the original and the paraphrase on five scales: Casual to Formal, Unfriendly to Friendly, Wordy to Concise, Complex to Simple and Unclear to Clear. Those score the text itself, and the page says nothing about politeness or how a reader will judge it.

Can a program rate politeness at all?

A program can rate the politeness of English requests close to a person’s accuracy within one site, and it is still behind across sites and across languages. Danescu-Niculescu-Mizil and colleagues built a politeness classifier from a corpus of annotated requests, using features that stand in for indirection, deference and impersonalisation, and reported that it “achieves close to human performance and is effective across domains”². The base is narrower than the quote. The 2013 paper had five US annotators rate each of 10,957 English requests (4,353 from Wikipedia, 6,604 from Stack Exchange) and tested only on the most and least polite quarter of each set. Within one site the classifier scored 83.79% on Wikipedia and 78.19% on Stack Exchange, against human accuracy of 86.72% and 80.89%.

Trained on Wikipedia and tested on Stack Exchange it scored 67.53% against 80.89% for people, and in the reverse direction 75.43% against 86.72%. When Srinivasan and Choi built TyDiP, a politeness dataset with 500 annotated examples in each of nine typologically diverse languages, multilingual models carried their English training over to the other languages reasonably well in the authors’ judgement, “yet fall short of estimated human accuracy significantly”²². The same paper also asked whether each English politeness strategy keeps its effect when mapped into other languages, and whether formality and politeness line up. Neither does so every time.

Can a tone tool judge a non-native writer’s English?

A tone reading of your English is an estimate, because neither classifier study cited above reports whether its requests came from writers using a second language. The tools above that read tone are doing a job of this kind, whatever their internals. The published evidence says the models tested so far are good at English and measurably behind people in other languages, and that the strategies which make an English sentence polite do not transfer one for one. Even where a model is strongest, it is modelling a generic reader.

The nearest test is adjacent evidence, not a tone test. Liang and colleagues ran seven detectors of AI text on 91 TOEFL essays written by humans and 88 essays by US eighth graders in 2023²³. The detectors were nearly perfect on the eighth graders’ essays and misclassified more than half of the TOEFL essays as written by AI, an average false positive rate of 61.3%. It says nothing about tone tools. Rereading the message as its recipient would, whenever a tone score on your own English worries you, is a sensible precaution regardless.

What no tool can check

No tool knows the one thing the research says matters most, which is who is reading, beyond what you tell it. Every perception study above got its result by holding the message constant and changing the reader. The same email was fine to a learner and abrupt to a lecturer⁴. The same grammar error was judged less harshly when the reader knew the sender’s background and not otherwise⁷. A tool sees the message. It does not see the lecturer, or how heavy the favour is to ask.

Whether your relationship with this reader allows a bare request, whether a message of two lines to this person reads as efficient or as cold, and whether the imposition is small enough to skip the apology are all beliefs about the situation, which makes them Thomas’s second category again⁸. A tool can tell you the sentence is direct. It cannot tell you whether direct was the right call, and the field does not have that certainty to give. Rewriting tools share that limit, because most hand back new text without saying how the reader will take it, which the review of AI message rewriters found across the main kinds of app.

What a tool can do is read the draft as someone other than you. It can notice that a request has no softening, that a message has no greeting, that a line sets a deadline you did not mean to set. That catches the first category, the conventional phrases, which Thomas said is “fairly easy to overcome”⁸. For someone writing in a second language that covers most of the easy fixes, and it is the part your own ear is worst at.

How can you check your English yourself before you send it?

Start with the shape of the request. Does it give the reader a way to say no? The 2011 finding was that emails can read as impolite when they “give the faculty no choice”¹. “Could you send it when you have a chance” does. “Send it by Friday” does not, and an opening “Please” is not a reliable fix either. Requests that began with “Please” averaged -0.30 on the politeness scale of the US raters in the Wikipedia study².

Use a softener that changes how the request reads. In Hendriks’s experiments, swapping “can” for “could” changed nothing, and “I was wondering if” was the one modifier tied to a more agreeable sender, by a small margin. Adding “possibly” made no difference⁵.

Acknowledge the ask. One clause that shows you know you are asking, “I know you’re busy”, does the work the lecturers in the perception study felt was missing⁴.

Keep the greeting and closing in a first message. Their absence was one of the features the 2011 study found in the emails that read as rude¹. Nobody expects them in an ongoing chat, but a first email to someone senior is read for them.

Watch your deadlines. “As soon as possible” was the phrase that gave the paper its title¹. If you mean no rush, say so.

Then read it as if you were annoyed. The reason is in the article on why you cannot tell if your own text sounds rude. In Kruger and colleagues’ fourth experiment, 54 students read their statements aloud, half in the tone they meant and half in the opposite one, and only that second half stopped overestimating how well a partner would decode them¹². It was one experiment, on sarcasm, so treat it as a precaution.

What Subtext checks in a draft you are about to send

Subtext, the app I co-founded, names issues such as pressuring or overly demanding wording, which overlaps with part of the first category in Thomas’s split, the conventional slips. I found nothing showing that it checks for missing softening words or greetings. It works in 17+ languages. I read the app’s backend prompts for what it returns; the on-screen marking is as the app shows it, and what it does with your data is as its privacy policy states it. What the app returns for each draft is described in full in what the Subtext app does.

For a message you are about to send, Subtext names up to three issues in your draft, most serious first, marks the words at fault, reads the main emotion, and offers up to three versions, each carrying a safe-to-send score, shown as a badge, for how likely it is to land as you meant. Each issue has a short explanation that quotes your own words where it points at a phrase, and pressuring, overly demanding, condescension and a robotic tone are among them. Each version fixes what was flagged while keeping your meaning and your register, and carries a sentence of its own naming what makes it land or what could be misread, whether its score is high or low. Subtext is paid, with a few free analyses to start.

Does Subtext ask who you are writing to?

Subtext can ask about the reader. The prompts tell it to read the message “the way a native speaker of its language would hear it”. For languages where politeness is grammatical, Japanese, Korean, Turkish and German among them, it asks about the relationship, or writes at the more polite level when it has to draft without an answer. English does not mark politeness in its grammar, so that rule does not apply to English drafts. In any language, when a draft pushes back on someone (a refusal, a boundary, a confrontation) and nothing says who it is for, Subtext asks who the recipient is before it drafts, while everyday messages get a draft straight away. How this reader, in this relationship, will take your words is still up to you, which is Thomas’s second category. No study cited here tested Subtext, or any tool above, on English written as a second language, so what I describe is what its prompts say it does.

What Subtext does with a received message

A message you received gets a summary in one line of what the other person wants from you and two to five action points, plus a reply if you ask for one, with no label on the sender’s tone, so the tone reading runs only on your own draft.

It knows your reader only as far as you tell it, and it does not know your field’s conventions or whether your manager likes short messages. It can flag a request that pressures or demands, one of the conventional slips, as in work messages that are firm without being cold. Use it as a second reader, and still reread the message as the recipient would.

Subtext sends the text you submit to the providers its privacy policy names, so it leaves your device. Nothing you submit is used to train an AI model, and the policy says none of those providers may train on it either²⁴. A conversation is deleted from Subtext’s servers five days after you last used it, or 90 days if you pin it. On the provider side, the policy says Anthropic deletes inputs and outputs within 30 days and may keep a flagged request for up to 2 years and the related safety scores for up to 7 years, Deepgram (voice transcription), with Subtext’s opt-out set, keeps audio only as long as transcription takes, OpenAI keeps read-aloud requests for up to 30 days, and Subtext has no zero-retention arrangement with any of them²⁴. A check of thirteen writing assistants’ policies, which also details the providers’ retention, found several tools above train on pasted text by default, including Grammarly and QuillBot on individual accounts and DeepL Write’s free tier. Try Subtext in your browserTry Subtext in your browser

Sources

Numbered in the order they appear above. Every DOI was resolved against Crossref and every page opened on 4 and 5 October 2026. The Subtext privacy policy was opened again on 10 October 2026, and entries 12 and 13 were added and checked on that date.

Economidou-Kogetsidis, M. (2011). “Please answer me as soon as possible”: Pragmatic failure in non-native speakers’ e-mail requests to faculty. Journal of Pragmatics, 43(13), 3193 to 3215.
Danescu-Niculescu-Mizil, C., Sudhof, M., Jurafsky, D., Leskovec, J., Potts, C. (2013). A computational approach to politeness with application to social factors. Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics. . Corpus of 10,957 requests rated by five US annotators each. Funded by NSF, ARO, DARPA and other grants and company support.
Biesenbach-Lucas, S. (2007). Students writing emails to faculty: An examination of e-politeness among native and non-native speakers of English. Language Learning & Technology, 11(2), 59 to 81. . 533 emails (382 native, 151 non-native) sent to one professor over six semesters; no funding stated.
Economidou-Kogetsidis, M. (2016). Variation in evaluations of the (im)politeness of emails from L2 learners and perceptions of the personality of their senders. Journal of Pragmatics, 106, 1 to 19.
Hendriks, B. (2010). An experimental study of native speaker perceptions of non-native request modification in e-mails in English. Intercultural Pragmatics, 7(2), 221 to 255. . Two online experiments with 110 and 158 native English raters; no funding stated.
Hendriks, B., van Meurs, F., Kakisina, B. (2023). The effects of L1 and L2 writers’ varying politeness modification in English emails on L1 and L2 readers. Journal of Pragmatics, 204, 33 to 49. . 120 Dutch and 131 British participants; no funding stated.
Vignovic, J. A., Thompson, L. F. (2010). Computer-mediated cross-cultural collaboration: Attributing communication errors to the person versus the situation. Journal of Applied Psychology, 95(2), 265 to 276. . Sample size is not stated in the abstract, the only part I could open.
Thomas, J. (1983). Cross-cultural pragmatic failure. Applied Linguistics, 4(2), 91 to 112.
Bardovi-Harlig, K., Dörnyei, Z. (1998). Do language learners recognize pragmatic violations? Pragmatic versus grammatical awareness in instructed L2 learning. TESOL Quarterly, 32(2), 233 to 262. . The finding is quoted as stated on the page for its replication, Niezgoda and Röver (2001), in Pragmatics in Language Teaching, Cambridge University Press: https://www.cambridge.org/core/books/pragmatics-in-language-teaching/pragmatic-and-grammatical-awareness-a-function-of-the-learning-environment/D2CE5D9EF68666879682B8BD3162C55D
Usó-Juan, E. (2022). Exploring the role of strategy instruction on learners’ ability to write authentic email requests to faculty. Language Teaching Research, 26(2), 213 to 237. . 110 emails before and after instruction; funded by Spain’s Ministerio de Ciencia e Innovacion, Universitat Jaume I and Generalitat Valenciana per Crossref; only the abstract was available to me.
Qin, W., Jia, R., Ren, W. (2024). Pragmatic competence in an email writing task: Influences of situation, L1 background, and L2 proficiency. Written Communication, 41(4), 726 to 755. . 306 emails from 32 English-speaking teachers and 121 learners; funded by the Shanghai Institute of Educational Science (C2023155) per Crossref; only the abstract was available to me.
Kruger, J., Epley, N., Parker, J., Ng, Z.-W. (2005). Egocentrism over e-mail: Can we communicate as well as we think? Journal of Personality and Social Psychology, 89(6), 925 to 936. . Five experiments; the fourth, on reading statements aloud, had 54 University of Illinois students; funded by a University of Illinois Board of Trustees grant and NSF grant SES-0241544. DOI resolved against Crossref on 10 October 2026; abstract and fourth experiment read in a full-text copy of the paper, because the publisher page did not load.
Pollmann, M. M. H., Roos, C. A. (2025). “I get u”. People correctly interpret the tone of text messages and emails. Computers in Human Behavior Reports, 18, 100689. . Study 1: text messages, 347 receivers and 171 senders. Study 2: work or education emails, 361 receivers and 61 senders. Rates positive or negative tone only. Crossref lists no funder. DOI resolved against Crossref on 10 October 2026; abstract read via OpenAlex, because the publisher page did not open.
LanguageTool. Free AI Grammar Checker. . Checked 4 October 2026.
DeepL. DeepL Write: your AI-powered writing assistant speaks French and Spanish. https://www.deepl.com/blog/deepl-write-welcomes-french-spanish. Checked 4 October 2026.
Wordtune. Express yourself with confidence. . Checked 4 October 2026.
QuillBot. QuillBot’s Paraphraser: The Best AI Paraphrasing Tool. . Checked 4 October 2026.
Grammarly. Writing Tone Detector and Tone Suggestions. . Checked 4 October 2026.
Tonely AI Ltd. Tonely AI, App Store listing. . Checked 4 October 2026.
ConfiText. AI Texting Assistant. . Checked 4 October 2026.
QuillBot. How can I check and use Tone settings in Quillbot? . Last updated 10 June 2026; checked 5 October 2026.
Srinivasan, A., Choi, E. (2022). TyDiP: A dataset for politeness classification in nine typologically diverse languages. Findings of the Association for Computational Linguistics: EMNLP 2022.
Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), 100779. . 91 TOEFL essays and 88 US eighth-grade essays; funded by NSF, NIH, the Silicon Valley Foundation and the Chan Zuckerberg Initiative.
Subtext, Terms, privacy and your account. Privacy policy last updated 26 July 2026. Checked 10 October 2026.