Working AI detectors

[MacKov4]

🏷️ Discussion Type

Question

💬 Feature/Topic Area

Study Projects & Resources

Hi everyone,

I'm currently writing an explanatory note for my final thesis.
We're allowed to use 20% artificial intelligence, but it's difficult to write the text myself for this project, so I always use AI. I searched the internet for random AI detectors, but they work so strangely and inaccurately that I don't understand the validity of this function. I'd like to ask GitHub users if there are any working detectors that can simply check part of your text with real and valid results?

[Pronoydebnath]

From what I’ve seen, there really is no AI detector today that can give “real and valid” results with complete certainty.

Most detectors work by analyzing statistical writing patterns, not by actually proving whether a human or AI wrote the text. Because of that, false positives and false negatives are very common — especially with academic writing, technical writing, or non-native English writers. ([Springer]1)

Even recent academic research and university discussions conclude that AI detectors should be treated only as indicators, not proof. ([jalt.open-publishing.org]2)

If you still want to test your writing, the tools people most commonly mention are:

• GPTZero
• Turnitin AI Detection
• Originality.ai
• Pangram

But none of them are consistently reliable, and the same text can receive completely different results across different detectors. ([Skyline]3)

Honestly, if your university allows up to 20% AI usage, the safer approach is usually:

• use AI only for brainstorming, structure, grammar, or rewording,
• keep your own drafts and notes,
• and make sure the final explanation reflects your own understanding and writing style.

That’s generally more defensible than relying on detector scores, since even fully human-written text can get flagged incorrectly. ([tryleap.ai]4)

[Me-Null]

Yes, AI detectors exist, but you need to understand that the AI detection of Turnitin is a closed-source and proprietary technology. Nobody knows how it really works. You can only guess. If you submit a paper to GPTZero or any other online, you will get a specific score. When you submit it to Turnitin, you will get a different score, and sometimes it will highlight different paragraphs as well.

Also, Turnitin does not allow for anybody outside of an academic institution to just check text for AI detection just like that. The only way to find out is to just submit, and then you will be told, or, if you are a teacher, you can submit to Turnitin that way. Basically, what I'm saying is that Turnitin AI detection is not exposed to the public. You cannot even buy it, and whoever sells it is lying. It's a scam.

I don't need to give my credentials here, but I can tell you, from my experience, just one example of the many I've seen with my own eyes: GPTZero will score 3% AI, and Turnitin will score 58% immediately, triggering a misconduct meeting with the student. This happened time and time again. If you think you want to use online paraphrasers, all of them use AI in the background to do the paraphrasing, which means you will get flagged.

Turnitin is not very precise, but if you copy and paste a 2,000-word essay from AI, Turnitin will flag it with great confidence, like 80 to 90%, as AI-generated. The only method for you is to get the AI to help you paraphrase manually, one paragraph at a time. Really, there is no workaround on this one. The AI will write the essay for you, and then you will take the essay, one paragraph at a time, with your own words and with your own paraphrasing skills, and re-write it. That is the only way to get away with this.