AI detectors and humanizers: how accurate are they?

The short answer

AI detectors are not reliable enough to prove anything on their own. They can wrongly flag human writing, especially from non-native English speakers, and AI text can often slip past them after light editing. Even the companies that make detectors say a score should be a reason to have a conversation, not a verdict.

"Humanizers" are tools that reword AI text so detectors miss it. They're inconsistent, they can make writing worse, and using them to pass off AI work as your own can break school or workplace rules.

How AI detectors work

Most detectors look for how predictable the writing is. AI models tend to choose likely next words, so their text often has low "perplexity" (it's easy to predict). Human writing tends to vary more.

The problem: plenty of human writing is predictable too. Simple, clear sentences, formulaic essays and writing by people still learning English can all look "AI-like" to a detector. And AI text that's been edited or reworded can look human.

What the research says about accuracy

• OpenAI shut down its own detector. OpenAI released an AI text classifier in January 2023 and withdrew it on July 20, 2023, "due to its low rate of accuracy." In its own tests it correctly caught only 26% of AI-written text and wrongly labeled human text as AI 9% of the time (OpenAI).
• Even commercial detectors come with warnings. Turnitin says its AI detection "may not always be accurate" and can misidentify human-written text. Its testing found more false positives when scores fall between 0 and 19%, so it shows those scores only as an asterisk, and it says the report should not be the sole basis for action against a student (Turnitin guide).
• Short text is hardest. OpenAI noted its classifier was "very unreliable" on texts under 1,000 characters (OpenAI).

Even a small error rate matters at scale. If a school checks thousands of essays, a low percentage of false flags can still mean many students wrongly accused.

Who gets wrongly flagged

A Stanford-led study tested seven popular detectors on essays written by humans. The detectors were nearly perfect on U.S. eighth-graders' essays, but flagged more than half of TOEFL essays written by non-native English speakers as AI (average false-positive rate 61%). At least one detector flagged 97.8% of those human essays (Liang et al., 2023, arXiv; published in Patterns).

The same study found that when researchers used ChatGPT to simplify the word choices in U.S. eighth-graders' essays, the average false-positive rate jumped from about 5% to about 57%. And when they used ChatGPT to make the non-native essays' vocabulary richer, false flags dropped sharply (arXiv). In other words, detectors can penalize plain writing, not AI use.

What "humanizers" do

Humanizers (and paraphrasing tools) reword AI-generated text to make it less predictable, so detectors are less likely to flag it.

• They often work, but not always. A 2025 study of 19 humanizer tools found that many detectors missed humanized text, while a detector specifically trained against these edits still caught much of it (DAMAGE, ACL Anthology 2025). A benchmark testing six evasion methods against 13 detectors found no method won on evasion, writing quality and cost at the same time (TH-Bench, arXiv).
• Quality can suffer. Rewording can introduce awkward phrasing, change meaning or drop facts.
• It's an arms race. Detectors retrain against humanizers, and humanizers adapt. Neither side "wins" for long.

Our honest view: if a task requires your own work, a humanizer doesn't make AI work your own. It just hides it. If AI use is allowed, you don't need one; say how you used AI.

If you've been flagged

Being accused when you did the work yourself is stressful. Stay calm and gather evidence:

1. Ask what tool and score were used, and whether it's the only evidence.
2. Show your process: version history in Google Docs or Word, notes, outlines, drafts, browser history of your research.
3. Offer to talk through your work. Explaining your reasoning in person is strong evidence.
4. Point to the guidance. Turnitin itself says the report shouldn't be the sole basis for adverse action (Turnitin guide).
5. Know the policy, and ask about an appeal process if needed.

Going forward, writing in a tool that keeps version history is a simple habit that protects you.

For teachers and managers

• Treat a detector score as a reason to ask questions, not as proof.
• Be especially careful with non-native English writers.
• Make rules about AI use clear in advance, including what's allowed.
• Use assignments that show process: drafts, reflections, in-class writing or short conversations.