Can You Recognize a GPT-Written text? (spoiler: look for the #testament)

With the arrival of ChatGPT in November 2022 (it's astonishing it was less than 2 years ago...) one of the questions raised was about how (and if) it was possible to recognize text generated from AI.

While there were many software products on the market able to go to internet and search bit and pieces of your degree thesis or school assignment, at the time it seemed impossible to recognize text generated by a large language model as such.

I tried to generate texts and pass to multiple "plagiarism detection" software available online (for free), also using ChatGPT to recognize text generated by ChatGPT. Nothing worked better than a random choice apparently.

Mario Santagostino was the first to make me notice that, in an email generated by GPT models to thanks someone for something there was a good chance to find the word #testament.

Enter the word #Testament

Let me make an example: if you ask ChatGPT, or Azure GPT for the matter, to:

PROMPT: write a congratulation email for the ABC team for winning a deal with customer XYZ

there is a good chance (close to 100% of my experiments with various models) that you end up with a text that has a sentence like "This victory is a testament to your unwavering focus and the robust strategies..."

If you search for "testament ChatGPT" over the internet you may discover that there are other words very common in generated text... looking back at my logged query "it's crucial" seems to be another preferred.

I did not build a list of "crucial words"

Do you know any word "always" present in your GPT responses?

If so please post it in a reply ... it would be good to have a "people detector"...