You might have seen the ‘delve’ issue pop up somewhere during a discussion about AI. AI likes that word. AI likes it a lot. The content AI generates does it’s best to sound like how people naturally write, but for some reason it’s far more likely to use the word ‘delve’ than ordinary people are. However, knowing that and using it as a basis to accuse someone of using AI is not the way to solve this problem – ordinary people do still use the word ‘delve’. It’s not like ChatGPT invented the word. It’s not like this is some hallmark of Olde Englishe accidentally incorporated into the LLMs and accidentally kept. It’s a modern-day word, one that isn’t slang or dirty or unprofessional.
And here is the problem – attempting to sort organic uses from inorganic uses of this word devolves into witch hunts. Once the paper, article, etc. is written, there’s no way to sort LLM from ‘human but kind of awkward’ 100% of the time. Little tips and tricks to find AI writing are worthless because the AI is ultimately mimicking people, and someone out there has been using ‘delve’. In an effort to sus out students using AI, the detection programs of old are disincentivizing the average student from using a thesaurus and learning new words at best, and actively targeting autistic and English-Second-Language students for talking and writing ‘wrong’ at worst, leaving them to learn yet another set of rules for their writing to blend in.
“Why not make an AI to do the detecting work?”, you may ask – and you’re right, people are trying, and they’re coming up against a wall. The LLMs on the market today were made with permission to lie in order to please the researchers who made them. ChatGPT is incredibly agreeable, and if it can’t find tons of data pointing to an author, it’s prone to saying ‘yes, I wrote that’ even when it obviously didn’t. Ask it who wrote Frankenstein, it would probably be able to respond accurately most of the time. Ask it who wrote ‘Research Paper on How Different Kinds of Water Affects Plants’, it has no idea, but it won’t tell you that, it’ll take credit.
Finding the witch by trying to drown them was always destined to catch innocent people. Some school governing bodies seem to have realized this already and swapped away from ‘detectors’ into programs that allow them to record the student’s actions while typing. If a completed essay is copy-pasted into the program, it’s failed. If the student is visibly typing, editing, back-spacing, etc., it can be graded. This helps – it’s possible for a child to put on a sophisticated enough act to fool even this method without writing their own work – but the reality is that Pandora’s Box has been opened, and the tech is only getting better at it’s tricks. Over email, over the phone and perhaps even video call, it may soon be impossible to tell whether you’re talking to a real person or not. The AI itself certainly won’t tell you.
Sources: https://mitsloanedtech.mit.edu/ai/teach/ai-detectors-dont-work/
