I made a GPTs that bypass GPTZero, in my own <i>limited</i> tests, it has 100% bypass rate.<p>The story:
One of my best friend was accused by professor of using AI to write his essay, which is not true: he use translators for words he doesn't know. I know GPTZero is not accurate, I sent him a few articles and tell him to make appeals to the department.<p>I'm really curious how GPTZero words under the hood. I open their website, paste 2 paragraphs, one generated by Bard, another I wrote myself. Of course, GPTZero knows mines was human and Bard is AI. I notice some wording differences between Bard, ChatGPT and me: AI has a better vocabulary than mine.<p>I started to explore: replacing words. I replace words like "quest" to "looking for", "in addition" to "more", "earth" to "the planet" etc. and voila, it is now 0% AI generated. Surprising how easy it is to work around.<p>Of course I didn't believe myself, hey, they are a company that has money from investors, are they lying? I checked their website and I saw this <a href="https://gptzero.me/writers" rel="nofollow">https://gptzero.me/writers</a>, "so you can write words that AI will never generate". Huh? That's just so weird. It makes me feel that they know internally that GPTZero is not accurate, so they have to launch new products because of investors.<p>According to the discovery I found about wording difference can affect GPTZero's detection, I quickly write a prompt for ChatGPT, in short: "Do not write paragraphs in parallel structure. Do not use SAT words, replace them with simple words or phrases". I did experiments for about 2 hours, and it actually bypassed GPTZero really stable!