Do AI detectors work? Students face false cheating accusations

AI detectors do not work. I have spoken with many people who think that the particular writing style of commercial LLMs (ChatGPT, Gemini, Claude) is the result of some intrinsic characteristic of LLMs - either the data or the architecture. The belief is that this particular tone of 'voice' (chirpy sycophant), textual structure (bullet lists and verbosity), and vocab ('delve', et al) serves and and will continue to serve as an easy identifier of generated content.

Unfortunately, this is not the case. You can detect only the most obvious cases of the output from these tools. The distinctive presentation of these tools is a very intentional design choice - partly by the construction of the RLHF process, partly through the incentives given to and selection of human feedback agents, and in the case of Claude, partly through direct steering through SA (sparse autoencoder activation manipulation). This is done for mostly obvious reasons: it's inoffensive, 'seems' to be truth-y and informative (qualities selected for in the RLHF process), and doesn't ask much of the user. The models are also steered to avoid having a clear 'point of view', agenda, point-to-make, and on on, characteristics which tend to identify a human writer. They are steered away from highly persuasive behaviour, although there is evidence that they are extremely effective at writing this way (https://www.anthropic.com/news/measuring-model-persuasivenes...). The same arguments apply to spelling and grammar errors, and so on. These are design choices for public facing, commercial products with no particular audience.

An AI detector may be able to identify that a text has some of these properties in cases where they are exceptionally obvious, but fails in the general case. Worse still, students will begin to naturally write like these tools because they are continually exposed to text produced by them!

You can easily get an LLM to produce text in a variety of styles, some which are dissimilar to normal human writing entirely, such as unique ones which are the amalgamation of many different and discordant styles. You can get the models to produce highly coherent text which is indistinguishable from that of any individual person with any particular agenda and tone of voice that you want. You can get the models to produce text with varying cadence, with incredible cleverness of diction and structure, with intermittent errors and backtracking and _anything else you can imagine. It's not super easy to get the commercial products to do this, but trivial to get an open source model to behave this way. So you can guarantee that there are a million open source solutions for students and working professionals that will pop up to produce 'undetectable' AI output. This battle is lost, and there is no closing pandora's box. My earlier point about students slowly adopting the style of the commercial LLMs really frightens me in particular, because it is a shallow, pointless way of writing which demands little to no interaction with the text, tends to be devoid of questions or rhetorical devices, and in my opinion, makes us worse at thinking.

We need to search for new solutions and new approaches for education.

> We need to search for new solutions and new approaches for education.

Thank you for that and for everything you wrote above it. I completely agree, and you put it much better than I could have.

I teach at a university in Japan. We started struggling with such issues in 2017, soon after Google Translate suddenly got better and nonnative writers became able to use it to produce okay writing in English or another second language. Discussions about how to respond continued among educators—with no consensus being reached—until the release of ChatGPT, which kicked the problem into overdrive. As you say, new approaches to education are absolutely necessary, but finding them and getting stakeholders to agree to them is proving to be very, very difficult.

I recently deployed an AI detector for a large K12 platform (multi-state 20k+ students), and they DO work in the sense of saving teachers time.

You have to understand, you are a smart professional individual who will try to avoid being detected, but 6-12th grade students can be incredibly lazy and procrastinate. You may take the time to add a tone, style and cadence to your prompt but many students do not. They can be so bad you find the "As an AI assistant..." line in their submitted work. We have about 11% of assignments are blatantly using AI, and after manual review of over 3,000 submitted assignments GPTZero is quite capable and had very few (<20) false positives.

Do you want teachers wasting time loading, reviewing and ultimately commenting on clear AI slop? No you do not, they have very little time as is and that time will be better spent helping other students.

Of course, you need a process to deal with false positives, the same way we had one for our plagiarism detector. We had to make decisions many years ago about what percentage of false positives is okay, and what the process looks like when it's wrong.

Put simply, the end goal isn't to catch everyone, it's to catch the worst offenders such that your staff don't get worn down, and your students get a better education.

Doesnt google docs have a feature that shows writing history.

You could ask the student to start wrkting on google docs, and whenever someone gets a false positive, they can prove they wrote it through that.

And Besides 99% of people who use AI to write, dont bother claiming it as a false positive, so giving students the right to contest that claim would not be that much if a problem long term.

Yeah, those are great points, and our students do use Google Docs today, and you are right most students do not even contest it.

We let them resubmit a new paper when they are caught, and they get some one on one time with a tutor to help move them forward. Typically they were stuck or rushing, which is why they dumped a whole AI slop assignment into our LMS.