Can you bypass AI detection?

How a detector decides

A detector gives your text a score, then compares it with a line. The line is set so that a chosen share of real human writing, often 1 in 100, is flagged by mistake. Anything above the line is called AI. Getting past detection means getting the score under that line without changing what the text says.

Tricks that don't work

The RAID benchmark applies the usual tricks to ChatGPT and GPT-4 text: lookalike letters from other alphabets, zero-width spaces, extra whitespace, odd capitalisation, synonym swaps, misspellings, dropped articles, inserted paragraph breaks, paraphrasing. At its 5% line, MELD caught 100% of the AI texts under every one of the 11 tricks (public_report.txt).

Character tricks also tend to backfire with people: a teacher or editor who finds Cyrillic letters or invisible characters in a document has stronger evidence than any detector score.

What moved the score, and what didn't

Pilot set, report.txt: the same 300 drafts, the same detector (MELD) at its 1% line. The jump comes from rewriting sentences, not from editing words. Best of four is picked by the same detector that grades it, so treat the one-try rows as the honest ones.

What the research says about paraphrasing

Paraphrasing has beaten detectors before. In a 2023 study, a purpose-built paraphrasing model cut DetectGPT's detection rate from 70.3% to 4.6% while holding false alarms at 1% (Krishna et al.). The same paper proposed a defence: a provider keeps a record of what its model generated and checks new text against it.

Detectors trained since then include paraphrased text in their training data, which is part of why the RAID paraphrase trick above no longer worked against MELD. Expect the same cycle to repeat.

Passing one detector isn't passing them all

• Of commercial humanizers' rewrites that MELD let through, GPTZero still flagged 33–36% and Originality 22–33% (about 2,400 rewrites from two public benchmarks).
• Prompting alone can get past a detector on some kinds of writing. Asked to explain things simply and avoid AI habits, GLM 5.3's answers were caught 2% of the time at MELD's shipped line and GPT-5.6 Luna Pro's 7.5%. The same approach on advice-forum posts was caught 29–96% of the time, depending on the model.
• Detectors flag some human writing too, so "never flagged" isn't a promise anyone can keep.
• Detectors change. A rewrite that passes today can be flagged after an update.

When not to

If a school or employer bans AI help for a piece of work, getting past a detector doesn't make it allowed, and the version history of a document can show how it was written. Rewriting is for your own posts, emails and messages, where the problem is that an AI draft reads like AI.