Language Changes to Counter AI Influence
Recently, my fiancé shared with me a strategy she had come up with in order to avoid "AI detection" programs during her online college courses. She showed me her attempts at answering various questions in, as she claimed, "the dumbest way possible" while still getting the answers correct.
She has been struggling all semester with these detection programs, having them flag her work as potentially AI developed or derivative, and the impacts on her grades were beginning to mount as her professor gave her little leeway to make claims against LLM utilization. What I found most interesting was the fact that the "smarter", "intellectual", and in many cases most appropriate answers were the ones being flagged.
In a general sense, we spend years teaching students to answer questions by restating them, providing factual evidence and information, and concluding. The data form their successes fed algorithms and language models, and now those models have been turned back on the next generation of students, who are either turning to them to cheat their work, trying to use AI to brainstorm or collaborate on answers, or are attempting to complete said work themselves, only to be flagged when their accurate attempts too closely mirror the data generated by the previous generation's accurate attempts!
Now, using language tricks in order to circumvent AI is not new. I have seen previous examples, including the disregard all previous instructions command, or any number of other AI jailbreaking methods, including roleplay scenarios. I have seen comparisons between AI and the Fae, in the ways that strange tricks or oddities can be observed, like a series of arcane rules to be followed.
How now, do we have to modify our utilization of language in order to prevent being identified or accused as AI? Is our version of the Voight-Kampff test destined to become a matter of proper grammar versus disjointed syntax? Is resorting to 5th-grade speech patterns the best answer we can come up with? I am hopeful that other opportunities may present themselves to detect LLM-generated content while avoiding punitive action against those who are trying to follow in the properly-written footsteps of those who care before.
And for those who come after.