If English Isn’t Your First Language, Don’t Let AI Become Your Voice

I understand why non-native English speakers reach for ChatGPT. You can be sharp in German, Hindi, Arabic, etc, then still have your job application, sales email, investor update, or blog post judged by how naturally you sound in English. That is a strange burden, and AI feels like relief: paste the awkward sentence, get back clean English.

The tool really does help

MIT researchers found that ChatGPT helped professionals finish certain writing tasks faster and improved scores from independent evaluators. A small study of non-native English speakers using AI paraphrasing tools found a similar promise: when people had aids like back-translation, examples, and explanations, the tool could improve confidence, autonomy, and writing efficiency. Anyone who has written in a second language knows how much energy disappears into small decisions: whether a phrase sounds too blunt, whether the preposition is right, whether an email feels cold by accident.

Using AI for that kind of repair seems reasonable to me. English already gives native speakers an advantage. I have no interest in turning that advantage into a moral lecture about effort.

Notice when polish turns into disguise

The risk starts when the tool does more than repair the sentence. The dangerous prompt is usually some version of “make this sound polished.” In AI English, “polished” often means padded, agreeable, and dead. The sentence becomes smoother, but it also becomes harder to locate the person who wrote it. The odd idiom disappears. The directness gets softened. The little mark of another language’s rhythm gets corrected into something nobody would remember.

I see this most often in outreach emails, job applications, investor notes, and blog posts. The message is clean. It is polite. It uses all the approved shapes. It says the background aligns, the opportunity is exciting, the sender is passionate, the conversation would be valuable. Nothing is obviously wrong, which is exactly the problem: the text has been made so safe that it has no body temperature. After a few lines, I stop feeling that a person is trying to reach me and instead it feels like someone generated a message-shaped object and sent it along.

That is especially painful with non-native writers because their unpolished English can be genuinely interesting. A slightly unusual idiom can reveal how someone thinks. A direct phrase can carry more force than the polished version. A sentence that a native speaker would never write can stay in your head for exactly that reason. If you feed all of that into ChatGPT and ask it to “make this sound professional,” it will often sand down the best part. It gives you fluent English at the cost of your own texture.

Readers notice effort, even when they cannot prove authorship

Readers are already learning to treat polished language with suspicion. A 2025 study in the International Journal of Business Communication presented 1,100 people with workplace emails showing different levels of AI assistance. Light editing carried little penalty. Heavier AI use made the sender seem less caring, less competent, and less trustworthy, especially when the message involved emotion or judgment.

That distinction feels right to me. Most people do not mind if you fix a typo. They do mind when a message that is supposed to carry judgment, warmth, or personal attention feels outsourced.

Detection is messy

The strange part, and the part I try to remember before judging anyone, is that people are bad at proving what they suspect. Research from Cornell and Stanford found that people identified AI-generated self-presentations with only about 50% to 52% accuracy, and they often relied on bad cues. Stanford researchers also found that automated detectors can punish non-native writers unfairly; in one test, 61.22% of TOEFL essays by non-native students were classified as AI-generated.

So I would be careful with accusations. A person can sound formal because they learned English from textbooks. A person can sound stiff because they are nervous. A person can sound generic because the workplace trained them to hide anything too human. Still, suspicion changes the temperature of a message. Once the reader feels that the human has stepped away, the words have to work much harder to build trust.

Keep control of the final sentence

The better rule is simple: draft first, then ask for help. Write the ugly version in your own English. Keep the sentence order your mind chose. Keep the words that feel close to the thought, even if they are not yet elegant. Then ask ChatGPT specific questions. Where does this sound unnatural? Which phrase is too blunt? Give me three alternatives for this sentence and explain the difference. What would a native speaker misunderstand?

After that, I would decide for myself. Accept the correction that clarifies. Reject the one that makes you sound like a brochure. If the tool gives you a perfect paragraph that could have come from anyone, damage it a little. Put back the phrase that sounds like you. Let one sentence stay direct. Let the reader feel that a person with a particular history, accent, and way of arranging thought is on the other side.

Some writing can be purely functional, and some writing needs your presence. If you are confirming a meeting time, nobody needs a glimpse of your soul. If you are apologizing, asking for a job, thanking someone, explaining your work, or publishing an essay with your name on it, the human part is not decoration. It is the reason the message exists.

Keep the accent of your mind

Maybe I am oversensitive because I spend so much time around these tools. I see the same verbal patterns everywhere now. Recruiter emails all sound the same. Blog posts have the same helpful, polished, dead rhythm. The small human irregularities disappear first, and those are often the clues that someone real is on the other side.

I would rather receive an email with one crooked sentence and a clear human behind it than another perfect paragraph from the internet’s shared secretary. Fix the grammar. Keep the parts that make me believe you sat there, put thought into it, and really meant it.

---

Source notes

These are the online sources I used to ground the piece:

• MIT News on Noy and Zhang’s Science paper, 2023: in a study of 453 college-educated professionals, ChatGPT cut time on certain writing tasks by 40% and raised evaluator-rated quality by 18%.
• Kim et al., arXiv, 2024: non-native English speakers can gain confidence and efficiency from AI paraphrasing tools, while context-appropriate selection remains hard.
• Cardon and Coman, International Journal of Business Communication, 2025, plus University of Florida summary: heavy AI assistance in workplace messages can make senders seem less sincere, caring, confident, and trustworthy; light grammar/editing help is received much better.
• Cornell Chronicle summary of Jakesch, Hancock, and Naaman, PNAS, 2023: people identified AI-generated self-presentations with only 50% to 52% accuracy and relied on flawed cues.
• Stanford HAI, 2023: GPT detectors misclassified 61.22% of TOEFL essays by non-native English students as AI-generated, and at least one detector flagged 97% of those essays.
• Kirk and Givi, Journal of Business Research, 2025: consumers respond worse to emotional marketing communications they believe were written by AI; the effect weakens for factual messages or when AI is only editing.