AntiGPT

Writing in English as a second language with AI assistance

By Vitalik Hakim · September 16, 2026 · 5 min read

This one's personal for a lot of our users, so we'll be practical about it.

The unfair part

Detectors measure predictability, and writing by non-native English speakers is often more predictable in vocabulary and structure, so it gets flagged at higher rates, even when it's completely original. On top of that, when you ask an assistant to 'improve' your English, it tends to sand off exactly the small irregularities that read as human. You can end up with correct, fluent, and flagged.

None of this is your fault, and it's worth saying plainly: your English is not the problem. The tools are measuring a proxy, and the proxy is biased against careful second-language writing.

A workflow that keeps your voice

Write your first draft naturally, even if the grammar is rough. Ask an AI assistant to correct it: correct, not rewrite. Then humanize the corrected text in Basic mode with Standard readability, which makes the smallest changes: it restores natural rhythm and cuts the stock 'assistant' phrases without restructuring your argument. Read the result against your original meaning sentence by sentence, because you're the best judge of whether a change drifted from what you meant.

Keep every draft. If your work is ever questioned, a version history that shows your English improving through your own edits is the strongest evidence there is.

Short, high-stakes writing

Cover letters, emails, and applications are where a generic register hurts most and where one specific, personal sentence helps most. Humanize them, but make sure the one memorable detail (a real project, a concrete reason) is yours and survives the edit. Correct English plus your specifics is a stronger application than fluent English that could belong to anyone.