Can AI text be made undetectable?

There are two different things people mean by "detecting" AI text, and they work in completely different ways.

Style-based detectors are the commercial tools schools, publishers and recruiters paste text into. They do not know who wrote anything. They are classifiers trained to estimate whether a passage resembles the statistical patterns of machine-written text: how predictable each next word is, how evenly sentence lengths are spread, which phrases recur. The output is a probability, often shown as a percentage, and it is an estimate about style.

Watermarks are different. A provider can nudge its model's word choices during generation in a pattern only its own detector can read. Whether a given model does this, and whether anyone outside the provider can check it, varies by company and changes over time; our guide to AI text watermarks tracks which providers have announced what. A watermark is a statistical signal spread across many word choices, so it needs a reasonable length of text to read and weakens as more of the words are changed.

Neither method reads the visible formatting of a text in any decisive way, which matters for what a formatting tool can and cannot do.

The strongest evidence against treating a detector score as proof comes from people with every reason to want detection to work. Wikipedia's editors, who spend a lot of effort cleaning up undisclosed AI-written articles, tell contributors in their guidance on signs of AI writing not to rely solely on detection tools: the tools "perform better than random chance" but "have non-trivial error rates", and can be thrown off by "text modifications (e.g. paraphrasing, markup, and spacing changes)" and by models the detector was not trained on. The same page says a high detector percentage is not, on its own, valid grounds for deleting an article.

Errors run both ways:

• False positives. Careful, formal, evenly structured writing looks statistically similar to model output. That puts people at risk who write in a drilled, regular style, including many non-native English speakers and anyone who learned to write from templates. Our guide to signs of AI-written text covers why the same tells appear in human writing.
• False negatives. Text that has been edited, mixed with human writing, or produced by a model the detector has not seen can score as human.

People are not much better. The same Wikipedia page cites a 2025 study finding human ability to tell LLM text from human text no better than chance, and a preprint in which a group of experienced annotators did well while infrequent AI users did only slightly better than chance.

The AI Formatting Humanizer is a formatting tool. It replaces em dashes with commas, straightens curly quotes, removes invisible characters and emoji, converts pseudo-font letters back to plain ones, and tidies spacing. It never changes a word.

That removes the tells a human reader notices at a glance, and the hidden characters that can cause trouble in forms, code and search. It does not change the things style-based detectors weigh most: the words chosen, their order, and the shape of the sentences. Because Wikipedia's guidance mentions markup and spacing changes as something that can affect detectors, a score may move after any edit, including this one. That is a sign of how fragile the scores are, not a way to beat them, and the tool makes no claim either way.

De-slop goes further: it rewrites the stock phrasing and structural habits of model output. That changes the text more, and scores usually move more as a result. It is still an editor, not an evasion tool, and its own FAQ says no promise is made about detectors.

The kinds of change that most alter how a text reads, to a person or a classifier, are the ones no tool can make for you: adding your own examples and numbers, taking a position, cutting what you would not have written, and rewriting in your own words. At that point the text is substantially yours.

Services that promise undetectable output are making a claim nobody can check in advance:

• There are many detectors, each with its own model and threshold, and they are retrained. Passing one today says little about another, or about the same one next month.
• Detectors return probabilities. A text can score low on one run and higher after a small change to a single paragraph.
• Watermarks, where they exist, are read by the provider's own tool, which the person promising "undetectable" usually cannot run.
• Paraphrasing tools that push hard against detectors tend to damage the writing: odd synonyms, broken idioms, lost meaning. The text becomes harder to flag and worse to read.

The honest position is the symmetrical one. A detector cannot prove that you used AI, and a rewriting tool cannot prove that you did not.

If you are allowed to use AI, follow the disclosure rules of wherever the text is going: a school's academic integrity policy, a publisher's guidelines, an employer's rules. Use the tools to clean and edit, then make the substance your own. Disclosure ends the question in a way no score can.

If you are not allowed to use AI, no cleanup changes that. Formatting tools do not make text yours.

If you are accused on the strength of a detector score, the useful evidence is process, not another score. Version history in Google Docs or Word, earlier drafts, notes, research tabs and the ability to discuss the work all show how it was written. Point to published guidance, such as the Wikipedia page above, that detector results carry real error rates and should not be used alone.

If you review other people's writing, treat a detector score as a reason to look more closely, not a verdict. Look for specifics only the author could know, and ask about the work directly.