I really did not intend to write this article. I did not intend to test a popular AI tool. I did not want to criticize, analyze or test anything. I work for an office that produces about 5,000 words per day in articles, and we have created a new writing tool. I figured we would wire in the Undetectable AI API, which would help mask our AI-generated text well enough for production.
I’m aware of AI detection techniques and methods to make AI-generated text look more human. The best technique of all is to write it out myself – have the AI make an outline and cover the base ideas, then write them in my own words. It helps a lot to be opinionated, cranky, detail-oriented and experienced.
Setting up the Undetectable AI API is a pain. When you sign up on the website, it creates a 3-day trial, and the API will not run with a trial account. I found that if I canceled the trial, I could create a paid account. We generated three articles before we tested the results.
The results were awful – wordy AI slop that tested worse after the API than without it. We tweaked the API settings, and the quality went down, not up.
I emailed Undetectable about my experience, and they refunded half the credits we had already spent. Then like a good scientist, I resolve to test very carefully to see if the Undetectable AI API would be a tool we could use in our stack.
What is Undetectable AI
Undetectable AI was founded in 2023 by Christian Perry. He brags that the entire staff is overseas except for one other person in the US. He bootstrapped the company using proceeds from a company he helped found and then sold. Undetectable has grown to over $10 million in revenue in two years, with about 70 employees.
Undetectable AI offers a money-back guarantee: “If anything we produce is flagged as not human, we will refund the cost of humanization.” It’s clearly on the top of their pricing page. That sounds iron-clad. They also promise they are “rated #1 on Forbes.”
Why I Am Testing
My colleague wired in the Undetectable AI SDK yesterday, thinking we would pay their prices for decent masking of our AI-generated text. During testing, he ran three articles through it. Then he came to me and said, “This result is testing higher for AI-generated content than our original text.” What’s going on here?
We went over the API settings and tweaked a few of them, but the results only got worse, not better. So I resolved today to make a small test case and use the browser UI – not the API – to run Undetectable through its paces. I still fully expected that a professional company with a solid guarantee would perform well.
The Test
For my test sample, I pulled a paragraph from an article we generated yesterday about FireStick and IPTV. I knew going in that it was AI-written – I wrote it. Here it is:
All FireStick models support IPTV app sideloading. The differences appear in performance and stability. The Fire TV Stick 4K Max (2nd Gen) and Fire TV Cube handle multiple streams and larger playlists without significant issues. The standard Fire TV Stick and older models struggle with playlist files exceeding 5,000 channels.
The Fire TV Stick Lite experiences noticeable buffering with high-bitrate streams above 10 Mbps. Users report frequent crashes when running resource-intensive IPTV players on this entry model. The Fire TV Stick 4K performs adequately for most IPTV applications. It handles standard streaming without major complaints.
I ran it through GPTZero. It scored 55% AI. That’s my baseline.
First test – Just basic Undetectable AI- I entered the test above and got the text below as a response. Nothing huge. A bit odd that there are about twice as many words. Some strange phrasing “are total rockstars” and “take it all in stride”.
Here is a picture showing a GPTZero test on it:
In my opintion, the text is worse than the original The added expressions do not add to the meaning, which is exactly what people hate about AI generated text.
This result does repeat what we saw with the API.
I then found a more detailed screen UI to humanize my text with some controls on it. This allows me to set Stealth mode and many other parameters. And I was very encouraged when it shows 99% human at the bottom.
So to double check – I pasted the result text into Undetectable AI’s own checker:
The good news – it is green. The bad news: it is 46% GPT, not 99% as the other screen indicated.
And when I test with my tool, 69%
So, clearly, Undetectable AI does not meet my expectations as a dependable, valid tool for improving my AI-generated text. The quality of the text is worse. And it scores worse on AI detection scores. The only test that showed improvement was their own test.
Canceling a Subscription
Certainly, I am disappointed. And now my task is to cancel my subscription. Checking my user account status, I’m surprised and disappointed that there isn’t a quick cancel button. Instead, you get this screen:
So I sent this email.
Getting a Refund
No problem canceling. But they offered me a partial refund – not the full refund:
I pointed out their guarantee. And got this in reply:
So there you have it- No easy refund at all. A nice note: I canceled your account, so you can’t do anything without paying us. Now, do you need any further help? I will be glad to assist. As if there was some service that they offered for free?
The Honest Review Problem
So, in writing this article, I see another problem. If you google up – is Undetectable AI a Scam, you get this result:
Honest reviews from all their competitors. You can tell from the company name and the headline that the review is: “Yes, they work but our product will work better.” This is silly.
So What’s Next for my Project?
Having AI write our text is a given. It has revolutionized my department’s output. Our output is now 10 times greater and of far higher quality than a year ago. Our research is deeper, we have achieved more of our goals, and it is far easier. The icing on the cake is to ensure that the text is readable to real humans, who tend to read it jumps and scans, and also that it has depth and meaning beyond the usual AI slop. With the newest tools, we know this can be achieved with some text style tuning and a bit of human touch-up.
As for thinking some website that advertises a “refund the cost” option on its payments page – well, not all humans can be trusted, can they? Lying on your payments page is dark. Check out Better Business Bureau.