How Accurate is Undetectable AI’s Detector? We Tested It On Itself

Undetectable AI has become popular in academic and blogging circles as a way to circumvent false AI detection. It's a quick and easy way to safeguard yourself against false positives that have already claimed the reputation of more than one student.

However, here's an interesting dilemma: If it's meant to hide AI writing using AI, how can you ensure that its outputs won't be flagged as machine-generated? That's the reason why we must test its output against the best AI detectors available.

2026 update: I’m preserving the original test order and screenshots below, but detector interfaces and scoring models can change. Treat this as a useful snapshot of the workflow, not a permanent guarantee that the same scores will repeat today.

Now, we've tested Undetectable AI against other detectors before. But what about its own detector? Will it pass its own test? Time to find out.

A Short Discussion on Undetectable AI’s Checker

We’ve had lots of Undetectable AI articles before like this one, and this one, and this one, oh and here’s another one. The list goes on. One feature that we’ve yet to highlight is its AI detector, or to be more specific, detectors.

Undetectable AI doesn’t really have its own detector. Instead, it aggregates the AI likelihood scores from eight different detectors, namely:

• GPTZero.
• OpenAI.
• Writer.
• Crossplag.
• Copyleaks.
• Sapling.
• Content at Scale.
• ZeroGPT.

Undetectable AI also frames the detector and humanizer as a loop: check the text, humanize it, then review the result again. I would still treat the detector score as a signal, not a guarantee.

If you’re interested to learn more about the accuracy of individual detectors, read our larger AI detector benchmark and our current guide on checking whether text looks AI-written. Detector behavior changes fast, so the useful part is the pattern across tools, not one score in isolation.

Undetectable AI vs. Its Own Detector

So, enough foreplay, let’s get into the nitty-gritty. For this testing, we’ll ask ChatGPT to write seven unique texts and then input it into Undetectable AI. Once it’s generated the output, we’ll check if Undetectable AI’s detectors flag their own output as AI.

Let’s get started.

Test #1

Original GPT Text

Tweaked Output from Undetectable AI

Undetectable AI Detection Score

Did it pass? Yes!
AI Detector Score: 75% (6 out of 8)

Test #2

Original GPT Text

Tweaked Output from Undetectable AI

Undetectable AI Detection Score

Did it pass? Yes!
AI Detector Score: 100% (8 out of 8)

Test #3

Original GPT Text

Tweaked Output from Undetectable AI

Undetectable AI Detection Score

Did it pass? Yes!
AI Detector Score: 100% (8 out of 8)

Test #4

Original GPT Text

Tweaked Output from Undetectable AI

Undetectable AI Detection Score

Did it pass? Yes!
AI Detector Score: 100% (8 out of 8)

Test #5

Original GPT Text

Tweaked Output from Undetectable AI

Undetectable AI Detection Score

Did it pass? Yes!
AI Detector Score: 100% (8 out of 8)

Test #6

Original GPT Text

Tweaked Output from Undetectable AI

Undetectable AI Detection Score

Did it pass? Yes!
AI Detector Score: 100% (8 out of 8)

Test #7

Original GPT Text

Tweaked Output from Undetectable AI

Undetectable AI Detection Score

Did it pass? Yes!
AI Detector Score: 100% (8 out of 8)

Results Tally

What Have We Learned? 

If we take these scores at face value, Undetectable AI's rewriting was effective in this sample against its own detector view. However, you should always review the output yourself because humanizers can introduce odd phrasing or grammar issues while trying to sound less machine-like.

And also, there’s another elephant in the room that I’ve yet to mention. Scroll back up and look at the eight detectors that Undetectable AI use. Notice something out of place?

If you said "OpenAI," you'd be correct. OpenAI discontinued its public AI detector, and that is exactly why old detector screenshots need context. I would not treat an OpenAI-branded likelihood score inside any third-party interface as the same thing as an official current OpenAI detector.

This is actually just one of the inconsistencies I notice with this detector. Another is that when I input Undetectable AI's output to one of their detectors' official websites, it sometimes gives me a different result from what I see in the Undetectable UI.

Beyond that, the practical conclusion is that Undetectable AI performed well in this point-in-time test, but the safest workflow is still to check the output, revise it manually, and avoid relying on a single detector score. If you're interested in learning more about this tool, you can read more about it in our official review.