Do AI humanizers work? What the research says about fooling AI detectors

Do AI humanizers work? Sometimes. Against weaker detectors they often do, and one 2026 study found a commercial humanizer slipped past two leading detectors 96% of the time or more. Yet other independent tests found the strongest detectors still caught 90% or more of humanized text. Whether a humanizer "works" depends on the tool, the detector and the text, and using one to hide AI use where it isn't allowed is still misconduct.

What does an AI humanizer actually do?

A humanizer takes AI-generated text and rewrites it so that detectors are less likely to flag it. The methods vary. Some tools keep each sentence's structure and simply swap words for synonyms. Others, usually built on large language models, reword whole passages and sometimes add or drop sentences.

The idea is older than the products. In 2023, Kalpesh Krishna and colleagues built a paraphrasing model called DIPPER and showed at NeurIPS that it cut one detector's success rate, DetectGPT's, from 70.3% to 4.6% at a 1% false positive rate. It also evaded GPTZero, watermarking and OpenAI's own classifier of the time. Researchers at Pangram Labs later wrote that this finding set off an explosion of humanizer tools online.

Their 2025 audit of 19 humanizer and paraphrasing tools found that many existing detectors failed on humanized text. It also graded the tools on quality. The best kept the original's tone and vocabulary level. The worst added nonsense phrases, produced sentences that couldn't be parsed and often distorted the meaning, which a careful reader may notice even when a detector doesn't.

Some tricks are cruder. Swapping letters for look-alikes from other alphabets, or inserting invisible characters, can confuse software without changing what a reader sees. The RAID benchmark includes both as standard attacks. Our invisible character remover strips them out so any detector sees plain text.

Do AI humanizers work against detectors? What four studies found

Results vary widely. Here is what independent and published tests found when they put humanized text in front of detectors:

Read together, these say that humanizers defeat most detectors much of the time, that the best detectors resist many humanizers, and that at least one humanizer has beaten even those in a published test.

Experienced human readers are harder to fool than you might expect. In the ACL 2025 study, the expert panel noticed that the same clues survived humanization. One example: the humanized articles gave people titles such as "Dr." or "Prof." far more often than human-written articles do. Humanizing changes the surface of a text, and some of the habits underneath stay put.

Why do the results differ so much?

Several things change between studies, and each one moves the numbers.

• The humanizer. A chatbot simply asked to sound more natural is a different opponent from a commercial tool built to beat detectors, and humanizers vary a lot in quality.
• The detector version. Vendors retrain often. The Notre Dame team tested Pangram 3.2, for example, and Pangram has since released a newer model.
• The text. Short academic abstracts give a detector far less to work with than a 4,000-word paper.
• What counts as caught. Some studies use a vendor's label, others a score threshold, and some count a "possibly AI" label as a miss.

The Chicago authors put it plainly: they expect detection to be a continuing race between detectors, models and humanizers, so any single result is a snapshot.

Do detector companies say they can catch humanizers?

Several do. Turnitin says it detects text from "bypasser" tools in English submissions, without naming the tools (we look at that claim in does Turnitin detect AI?). Pangram trains a model specifically on humanized text, described in the paper by its own staff mentioned above.

Treat these as vendor claims until an independent test covers the humanizer you have in mind. The studies in the table show Pangram catching most humanized text in three tests and almost none in a fourth, and Turnitin catching half in the one test that included it.

Is using an AI humanizer cheating?

If the rules say no AI, or require you to disclose it, then yes. A humanizer exists to hide where text came from. Running AI output through one doesn't make the ideas or the work yours. It only makes their origin harder to see, which is the opposite of the disclosure most AI policies ask for.

The research also points to an unfair side effect that has nothing to do with cheaters. The Notre Dame study found that light AI polishing, the kind some institutions allow, was flagged far more often than fully humanized text. People who use AI honestly can face more scrutiny than people who hide it. That is a strong argument for institutions to rely on drafts and conversations rather than detector scores alone.

If you were wrongly flagged, don't reach for a humanizer to "fix" your own writing. It replaces your words with a model's and wipes out the evidence that would clear you. Our guide to being wrongly flagged by an AI detector covers what to do instead.

Checking a suspect text with our detector

Like any detector, ours can be wrong in both directions, so treat a result as a reason to look closer, not as proof. What the report gives you is something concrete to look at: a sentence map, style explanations and any hidden characters. It is free, needs no account and runs in your browser, so the text is never uploaded. Our methodology page explains how it is tested.

Sources

1. Krishna et al., Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense (NeurIPS 2023), checked October 2026
2. Masrour, Emi and Spero, DAMAGE: Detecting Adversarially Modified AI Generated Text (Pangram Labs, 2025), checked October 2026
3. Dugan et al., RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors (ACL 2024), checked October 2026
4. Jabarian and Imas, Artificial Writing and Automated Detection, BFI Working Paper 2025-116, checked October 2026
5. Russell, Karpinska and Iyyer, People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text (ACL 2025), checked October 2026
6. Van Vlasselaer, Van Droogenbroeck and Spruyt, Who wrote this? Evaluating the reliability of AI detection tools in higher education (International Journal for Educational Integrity, 2026), checked October 2026
7. Karr, Khvatskii, Hua and Chawla, Why AI Detection Fails for Academic Integrity (ACM AI Summit 2026), checked October 2026
8. Turnitin Guides: Turnitin's AI writing detection capabilities FAQs, checked October 2026