Can AI detectors catch humanized text?

"Humanizers" are tools that take AI-generated text and rewrite it so detectors miss it. Some swap synonyms, some add typos, some run the text through a second model with instructions to sound more casual. Do they work? Partly, and less than they claim.

What a humanizer changes

Most humanizers apply some mix of the same transformations:

• Synonym substitution. Replacing common words with less common ones to make the text less predictable.
• Sentence reshuffling. Splitting long sentences, merging short ones, and moving clauses around.
• Noise. Adding contractions, filler words, small grammar slips or typos.
• Paraphrasing with another model. Asking a second language model to rewrite the text "like a human".

Each of these attacks one signal. Synonym swaps target predictability. Reshuffling targets rhythm. Noise targets the polished surface of model output.

Why traces remain

A detector that relies on a single signal can be fooled by attacking that signal. An ensemble is harder to fool, because a humanizer has to defeat every signal at once without destroying the text.

In practice that trade-off shows. Synonym swaps produce odd word choices that no fluent writer would make. Paraphrasing with another model often reintroduces the very patterns it was meant to remove, because the second model writes like a model too. And the overall structure, the order of ideas and the way an argument is built, usually survives every pass.

That is why we train LookAI on text that has already been through humanizers: two humanizers in a row, paraphrasing tools, and deliberately sloppy rewrites. The model learns what humanized text looks like, not only what raw model output looks like.

What the numbers say

On our held-out English test set, LookAI catches about four out of five humanized texts at the AI threshold, while keeping fewer than 1 in 100 human texts in the AI band. Raw AI text is caught at a similar rate. Results are lower for very short texts and for text that a person has substantially rewritten by hand.

Two caveats matter more than the headline number:

1. The arms race is real. New humanizers appear every month, and some will work better than the ones we trained on. We retrain regularly and version every model, so each result tells you which model_version produced it.
2. Heavy human editing is not cheating the detector. If someone takes an AI draft and rewrites it thoroughly in their own words, the result is substantially their own writing. A detector that flagged it would be making a mistake.

How to use this in practice

• Treat a high score on humanized text as a reason to look closer, not as proof.
• Look at the highlighted passages. Humanized sections often stand out next to the writer's own work.
• Pay attention to confidence. When the signals disagree, LookAI lowers the confidence and returns an inconclusive result rather than guessing.
• Combine the result with context: drafts, version history, a conversation with the writer.

Humanizers make detection harder, not impossible. A detector trained on their output, calibrated carefully and used with human judgement, still catches most of what they produce.