How I Fooled the Top AI Detectors Within 15 Minutes

Oh I am but a lazy student…trying to use AI to complete my assignments.

I’ve been a ‘lazy’ student all my life. I still remember my math teacher in high school always saying to me: “Chris, you just want to find the shortcuts to everything”. That didn’t change going into university.

‍

If I were still a student, you bet I would be using ChatGPT, Claude, or the top model. And I wouldn’t truly be lazy if I let my efforts go to waste by letting the teacher or professor catch on.

‍

So I pretended to be that Chris 10+ years ago but with all the tools at my disposal today. Everyone’s got access to GPT4o and Claude so my next step was to avoid detection.

‍

Here’s my AI-detection-proof, foolproof, evil mastermind plan.

1. Use ChatGPT and Claude to generate my submission, simultaneously (for efficiency)
2. Find out what I am going against (aka Google search for AI detectors)
3. Run my AI-generated submission through those tools to see if I’ll get caught
4. If I am caught: go back to Step 1, rinse and repeat

‍

First, the most commonly used AI detectors

1. Undetectable AI
2. ZeroGPT
3. Copyleaks
4. GPTZero

‍

Or at least they had the best SEO or were paying Google Ads the most…

‍

‍

‍

Experiment #1

I found a high school assignment about AI online and generated my submissions. Then I fed the output without any changes to the four AI detectors above.

‍

Results:

1. Undetectable AI: 50% AI generated
2. ZeroGPT: 100% AI generated
3. Copyleaks: 100% AI generated
4. GPTZero: 100% AI generated

‍

Snap, no way I am going to submit this assignment as is! It seems like I am going to have to put more work into this assignment. Uhh.

‍

I had one card up my sleeve though. While searching for AI detectors, ads for ‘AI text humanizers’ came up. Here is one example amongst many.

‍

‍

‍

‍

Experiment #2

For the sake of the mental health of educators reading this, I won’t go into exact details on how this AI humanizer works but it was easy. It was also free. Next, I fed the newly generated text into the four AI detectors again.

‍

Results:

1. Undetectable AI: 100% Human!
2. ZeroGPT: 100% Human!
3. Copyleaks: 100% Human!
4. GPTZero: 96% Human!

‍

‍

‍

‍

I am actually kind of mad that I didn’t have the chance to execute my evil mastermind plan. I expected to at least put in some work but I didn’t realize that it was so easy.

‍

Oh and the nail in the coffin for AI detectors? Some AI humanizers have a premium version where it is even more proficient, potentially giving a leg up to students who have the wallet to pay for it.

‍

AI Detection does not work - don’t let confirmation bias get you. It took me more time to write this than to beat the AI detector. It also took me more time to find the ‘Cancel Subscription’ button and cancel my subscription to Copyleaks.

‍

If you are worried about academic integrity, you should change how you assess students. Move away from simple, summative assignments to formative learning opportunities. If you are worried about workload, there are tools such as FeedbackFruits, TimelyGrader, and others that will help you deploy formative learning.

‍