So, once again, I am attempting to help a student reduce an AI detection score to below the mid-30% range, according to Turnitin. Strangely enough, Grammarly agrees with that score, but flags different passages. Their agreement is around 10%. Does that mean the paper is 60% AI-generated?
Like my experience with the Turnitin plagiarism checker, Grammarly also changes its mind about what to flag. Between checks, a clean paragraph with no changes shows up flagged, and then a previously flagged paragraph with no changes shows up clean. I begin to think what AI detectors flag is mostly random.
Change, check, change, check, change again and leap for joy when the flag disappears—until the next time, when it may be flagged again.
The realisation dawns that my efforts are pointless. The machine is learning from my changes, taking possession of whatever text I submit, then flagging it because it now belongs to the machine.
We cannot win. AI is designed to replace us: Every effort made to circumvent or sidestep that intention allows the machine to learn how to better sideline us. The human mind—everything cognitive, data-based, and documented—is being hijacked, one step at a time. And every time we participate, we give it more power.
So, after a PhD in psychology and a lifetime of contemplating the evolution of human consciousness in pursuit of the Truth, the truth is that I would be well-advised to attend a trade school.
I learned something very interesting recently while doing what (still) currently earns my keep: editing for academics, postgraduate students, and nonfiction writers.
In my vocation, AI has become a bone of contention since its invention and infiltration.
Besides the horrors of Turnitin, which checks for plagiarism, serious (and apparently less serious) writing is subjected to AI-detection checks because it is just so good at summarising vast swathes of literature if you guide it about what you intend to construct.
Some time back, I edited a paper that had an AI detection score of 93% for the first 15,000 characters in the original version of a client’s submission—there was barely anything there that had not been generated by AI. The client intended to submit to a peer-reviewed journal, so I adopted a more formal academic tone. The client was disappointed with the outcome—too stiff, she said—but doing so halved the AI-generation score.
More recently, I fed the Beast multiple times when I had the opportunity to put the ZeroGPT through its paces with a text just under 45,000 characters that needed to be “doctored” rather than just edited.
The original scores were 18.1%, 11.8%, and 10.8%. Not bad, all things considered, but the client had used Turnitin and received a score of 62%.
I executed a general edit and tested.
Oops! I reduced the first two sections’ scores by half, but the score had gone up substantially (8%!) in the last section with the inclusion of the necessary articles and making sure the nouns and verbs are congruent. Expressing text in perfect English is now, apparently, AI’s prerogative.
I worked with the highlighted text in the third section some more, considering the word choices, and returning them to the original, less precise, word choices.
And I tested again. Still 4% above the original score, but at least the final two paragraphs were now free from AI contamination. However, like working with Turnitin for plagiarism, passages that were not flagged initially emerged.
So, I worked on those passages and tested again, and again, and again.
In the end, I reduced the overall score from 13,6% to 10,4%. On Turnitin, according to the client, the score dropped from 62% to 20%. AI detectors clearly do not work from the same page.
So, if you are looking to fool the AI detectors:
First, adopt a more formal, academic tone. That’s at least half the battle won.
Second, know that perfect English has become AI’s prerogative. So, let an error or two slip through.
Third, know that once AI has touched a written piece, it possesses it. It learns from repeated submissions, absorbing whatever is fed into it and feeding it back while claiming ownership. So, the more you test your efforts, the higher the score.
Fourth, the scores for the same passages on different AI detectors vary wildly, with Turnitin being the most terrifying.
And then know that, like the spike protein in the blood, once embedded, AI-generated text is impossible to get rid of.
