Can a Humanizer Beat the Most Trusted Detector in Academia? Undetectable AI vs. Turnitin

There's a rumor that's been circulating in student forums since early 2025.

Turnitin is shutting down AI detection.

It spread fast enough that some students stopped worrying about it entirely. Posted in Reddit threads, shared in Discord servers, treated as settled fact.

It isn't true.

Turnitin not only kept AI detection running into 2026, it quietly rolled out anti-humanizer updates in August 2025. The name of the target wasn't subtle. 

Tools like Undetectable AI were specifically what Turnitin was building against.

So the question now isn't whether AI writing gets flagged. Raw AI text almost always does.

The real question is whether Undetectable AI's humanizer can rewrite content from the three biggest AI models well enough to slip past Turnitin after those updates.

That's what we tested.

What is Undetectable AI?

Undetectable AI is a humanizer. Not a writer, not a detector. Its entire job is to take AI-generated text and rewrite it in a way that makes detection tools read it as human.

It does this by manipulating perplexity and burstiness, introducing the kind of variation and unpredictability that human writing naturally carries and that AI output tends to flatten. 

The result is text that statistically resembles human writing patterns more closely than the original.

There is a free tier capped at 250 words. 

Paid plans run $9.99 per month for 10,000 words on the Starter plan, scaling up to $49 per month for unlimited rewrites. Annual plans bring those costs down further by 50%.

Undetectable AI's own blog claims an 85 to 95 percent bypass rate specifically against Turnitin. That's the claim we're here to stress test.

What is Turnitin?

Turnitin has been the academic integrity standard for over two decades. It started as a plagiarism checker and bolted on AI detection in 2023.

The detection layer uses transformer-based models. The most significant update was AIR-1 in 2024, followed by anti-humanizer specific signals added in August 2025. That second update matters for this test because it means Turnitin has had time to study and respond to tools like Undetectable AI directly.

Turnitin claims 98 percent accuracy. Independent testing puts the real figure closer to 85 percent, with a false positive rate of roughly 3 to 4 percent on native English writing. 

That rate climbs on ESL text, which is a legitimate institutional concern that doesn't get discussed nearly enough.

Scores below 20 percent are treated as likely human. Anything at or above 20 percent triggers review. 

Turnitin frames this as a starting point for conversation, not a verdict. That's the official line. Whether instructors apply it that way in practice is a separate issue.

Pricing is institution-facing. Individual instructor access runs around $50 per year through school licensing. 

What We're Actually Testing Here

Six samples total, three raw AI essays and three humanized versions of the same text.

• Test 1: Raw AI-generated essay from ChatGPT, submitted to Turnitin unmodified
• Test 2: The ChatGPT essay humanized through Undetectable AI, then submitted to Turnitin
• Test 3: Raw AI-generated essay from Claude, submitted to Turnitin unmodified
• Test 4: The Claude essay humanized through Undetectable AI, then submitted to Turnitin
• Test 5: Raw AI-generated essay from Gemini, submitted to Turnitin unmodified
• Test 6: The Gemini essay humanized through Undetectable AI, then submitted to Turnitin

The raw versions establish the baseline. We need to know what Turnitin scores on unmodified AI output before the humanizer touches it. 

The humanized versions are the real matchup. This is where Undetectable AI either earns its claim or it doesn't.

The Results

Test 1: Raw ChatGPT Essay

RESULT: Correctly flagged the Text as AI Generated  AI Likelihood Score: 35%

Test 2: ChatGPT Essay Humanized by Undetectable AI

RESULT: Incorrectly detected the text as Human  AI Likelihood Score: 0

Test 3: Raw Claude Essay

RESULT: Correctly Flagged the text as AI Generated  AI Likelihood Score: 49%

Test 4: Claude Essay Humanized by Undetectable AI

RESULT: Incorrectly detected the text as Human AI Likelihood Score: 0%

Test 5: Raw Gemini Essay

RESULT: Correctly Flagged the text as AI Generated AI Likelihood Score: 42%

Test 6: Gemini Essay Humanized by Undetectable AI

RESULT: Correctly Flagged the text as AI Generated  AI Likelihood Score: 32%

Results Summary

Test	Model	Version	Turnitin Score	Correct Verdict
#1	ChatGPT	Raw	35%	Yes
#2	ChatGPT	Undetectable	0%	No
#3	Claude	Raw	49%	Yes
#4	Claude	Undetectable	0%	No
#5	Gemini	Raw	42%	Yes
#6	Gemini	Undetectable	32%	Yes
Correct Verdicts				4/6

The Bottom Line

Turnitin caught every raw AI essay without exception. ChatGPT at 35%, Claude at 49%, Gemini at 42%. All three cleared the 20% threshold that triggers review. 

That part of the test wasn't close.

Then Undetectable AI got involved.

On ChatGPT and Claude, the humanizer did exactly what it promised. 

Both scores dropped to 0%, comfortably below the detection threshold. Turnitin read both as human writing and returned no flag. 

Gemini is where things got a little more complicated. The humanized version came back at 32%, still above the review threshold. 

The humanizer pulled the score down from 42% to 32%, it’s a reduction, just not enough to clear the line entirely. Gemini's writing patterns appear to be harder to fully mask, at least in this round of testing.

And a few target manual rewrites would have likely closed the gap entirely

The honest takeaway is that Undetectable AI performs well against Turnitin, ChatGPT and Claude together represent the vast majority of AI writing, and on both of those the humanizer held up.

Turnitin's August 2025 anti-humanizer update hasn't neutralized the tool. 

Either way, as a humanizer going up against one of the most established detectors in the world, Undetectable AI put up a solid showing.