I’ve always thought that AI detectors are a hit-and-miss, and real-life cases back that up. Over the last year, more and more cases of students receiving sanctions for false positives have been popping up all over the media.
With students’ futures on the line, how can we both ensure responsible AI use in the classroom while limiting false positive cases?
Undetectable AI rises up to the challenge with their own AI plagiarism checker. In this article, I’ll be reviewing their own AI detector, test it, and tell you what exactly makes it more accurate than the rest.
What is Undetectable AI?
Undetectable AI is a popular AI bypasser platform whose purpose is to humanize machine-generated text. It’s composed of three main features: the AI text humanizer, AI human typer, and the AI plagiarism checker. We’ve reviewed the first two extensively in the past, so make sure you check them out here and here, or in some of our comparison articles like this one.
What is Undetectable AI’s Plagiarism Detector?
Saying Undetectable AI has their own plagiarism detector is a bit disingenuous, but what they offer is equally good, if not better. What Undetectable AI does is aggregates AI likelihood scores from eight popular AI detectors to get a better picture if your text is vulnerable to getting flagged as machine-generated.
We’ve rated some of the AI detectors that Undetectable AI uses highly in our latest batch of accuracy testing. This is the complete list of tools they use:
Just from this list, I’m a bit hesitant on using Undetectable AI as my primary AI detector because they still include OpenAI, whose detector was shut down more than a year ago because of unreliability.
But at the end of the day, what matters most is if Undetectable AI can detect text from LLMs accurately. So, let’s find out.
How Accurate Is It?
True Positive Tests
This first set of tests will measure how accurate Undetectable AI is at detecting machine-generated text as, well, machine-generated. Here’s how well it did:
Test #1
Test successful! Machine detects text as AI.
Test #2
Test successful! Machine detects text as AI.
Test #3
Test failed. Machine detects text as human.
Test #4
Test successful! Machine detects text as AI.
Test #5
Test successful! Machine detects text as AI.
Test #6
Test successful! Machine detects text as AI.
Test #7
Test successful! Machine detects text as AI.
False Positive Tests
Now, let’s try if Undetectable AI can detect human-written text as not AI.
Test #8
Test successful! Machine detects text as human.
Test #9
Test successful! Machine detects text as human.
Test #10
Test successful! Machine detects text as human.
Overall Score
Does Undetectable AI Show Real Results?
As I mentioned earlier, I’m a bit hesitant to use Undetectable AI as my main detector because it still lists OpenAI as one of the detection tools they use. So, that had me wondering: if OpenAI is still there, can we trust that the results from the other seven detectors are the same when I use their own respective platforms themselves?
That’s what I’m going to check now and I’m gonna use the text from test #5. For reference, it was detected as AI using GPTZero, CopyLeaks, Sapling, Content at Scale, and ZeroGPT according to Undetectable AI.
Is that true? Let’s find out.
Test #1: GPTZero
Expected result achieved!
GPTZero detects text as AI-generated.
Test #2: CopyLeaks
Expected result achieved!
CopyLeaks detects text as AI-generated.
Test #3: Sapling
Expected result achieved!
Sapling detects text as AI-generated.
Test #4: Content At Scale
Expected result achieved!
Content at Scale detects text as AI-generated.
Test #5: ZeroGPT
Expected result achieved!
ZeroGPT detects text as AI-generated.
The Gist of It
Despite my reservations, I’m now convinced. Undetectable AI is not only an effective AI bypasser, it’s also a great AI detection tool.
“Unreliable” is a term that gets thrown around a lot in terms of AI detection. In fact, one of the earliest entities to use this term for AI detectors is OpenAI itself. That’s also why they took down their own AI detection tool.
What Undetectable AI does here is mitigate the risk of false positives by aggregating scores from eight (technically, seven) different detectors. This is an ingenious move, and one that’s paying off in heaps.
If you’re interested in reading more about Undetectable AI, I got you. Here are some of my recommended articles from our catalog. Have fun reading!