AI Checkers Forcing Kids to Write Like a Robot to Avoid Being Called a Robot

As a teacher, my answer to publicly available AI, regarding grading, has been to only let work by students I have seen been done myself (or staff, trusted implicitly) count towards passing/failing/diplomas.

This means graded longer tasks such as writing or programming, which in the past were done outside of class and then handed in, are instead done in person (like most exams are, traditionally). To acknowledge the longer process, the time alotted for such a test is much longer (up to 3 hours instead of 1) and I can expect less of work created in 3 hours than work created over 6 weeks.

Students still have the opportunity to hand in work that was done outside of such an in-person test, which I will grade, but that grade isn't on the record is serves only as an indication.

I wish I could trust students, but plagiarism has become so easy now (added to a zeitgeist of personal gain above all) that a critical mass fails to resist.

I think diploma's are very useful. They get their usefulness from the trustworthiness of the party handing out the diploma. Only recording grades for work that has been done in an environment controlled by that party has downsides, but in the current and foreseeable situation I don't see a better way of maintaining a base level of trustworthiness.

I love teaching without grades or diplomas involved so much more. Then it's all about doing what's best for the students.

I submitted several scientific/research papers my wife wrote or co-authored in the late 90s early 2000s and they were all flagged as AI written by multiple tools. It was good for a laugh.

If you put Shakespeare or the Constitution through one of those, they will say 90%+ AI. Any student who gets dinged by one of those things just needs to show this result until the administration caves and bans the stupid checkers.

They may have to go to the local press or a board meeting to get that result though.

Tests should not be a tool for authenticating learning. They were NEVER good tools for this and now its clearer why. Tests should be used for a teacher to help a student figure out where they need to focus.

It should be like getting a blood test from a doctor unless they are testing for drugs why would you try to cheat with someone elses blood?

We don't need to figure out how to sidestep the fundamental flaws with our institutional education we need to find and fix the root cause.

Being "autheticated" as "knowing stuff" (your degree and GPA) doesn't translate into companies being able to tell if you can do a job. It didnt before AI and it doesn't now.

Everyone should stop and think about how dumb it is that children and adults are trying to cheat at learning. You can't cheat learning. So why do people try and succeed? Because we arent actually asking for learning we are asking for something else.

This is an expected consequence of using ML to detect AI. LLMs are more well-read than the average person, so they'll have a bigger vocabulary than the average person, so the ML model will learn to associate a bigger vocabulary with a higher probability of being AI. Which unfairly penalises well-read humans, who also have a bigger vocabulary than average.

Personally I think the whole endeavour is a waste of energy. Unless there's a massive civilisational collapse, humans will always have LLMs as writing aides, so trying to teach them to write without one is teaching them a skill they'll never need to use.

If you want to teach them to reason on the spot, then go back to requiring them to orally put forth and defend an idea.

Oh god, this really does capture my problem. My university demands 40 page reports that are never read by anyone, so my seniors used ChatGPT to generate the needed drivel needed, so now we have the ai detectors, and are told that the % of AI content as flagged by a tool (turnitin) must be below 10%.

Unfortunately, the way I write, (like any well educated Indian) is a rather verbose, formal and rather florid in it's manner, and with no fewer than 3 conjuctives (when I don't want the reader to understand what I just wrote) per sentence. This style is for good or bad understood by turnitin as being AI generated.

So, I have had to unfortunately simplify and write in a more stilted manner, just to be able to submit my assignment.

Writing teachers have a lot to learn from Math teachers, who have had to fight against the crutch of calculators for decades. The future here is obvious. More in-class tests to evaluate writing ability instead take home is what will happen. For take home, require "showing your work" and the research that went into writing. Show the steps.

Yes, you could still use AI for any take home work, but I think what forces students to AI isn't the lack of will to do the work. It's the "I have a 5 page paper due tomorrow and I haven't started" cliff.

I mean, writing education is messed up in so many ways, but if I'm being realistic this is the path forward.

How can we claim any high ground when we are handing off the need to check for plagiarism to a tool that in effect is the same as the student using it to create plagiarized works.

It doesn't even work, its non deterministic, and most importantly - if the teacher is familiar with the student's writing it should be pretty obvious if this is a fake or not based on something they can be observed doing.

There is a certain style to LLM-created writing. It tends to float in the air, lightly tethered to real world examples. Unfortunately, that's also what passes for academic style in some quarters. I mentioned that a few days ago here.[1]

LLMs have grammatically correct blithering nailed. One amusing result is that Derrida and his followers now sound like an LLM with the temperature medium high.

Maybe we should be teaching kids to not write like that.

Mandatory XKCD.[2]

[1] https://news.ycombinator.com/item?id=41419645

[2] https://xkcd.com/451/

Interesting piece.

I really wonder why these tools even show detection percentages when they’re meaninglessly low. A 17% result from a classifier model is totally meaningless. You could show a well-trained animal classifier the first Google Images result for “cat” and it would still say there’s a 17% chance it’s a cow. Showing low-likelihood predictions will do nothing but freak people out and confuse them.

Interesting that my childhood evaluation systems in India are resistant to this new development. 100% scores from in-person exams, no loss of dynamic range through grading (i.e. 97% score > 96% score), zero score to homework. I imagine that the primary constraint in the US is labour cost since you need what you call a proctor and what we call an invigilator, and perhaps the substantial student power (a student and their parents can overrule teachers).

I like the relative mobility in the Indian system since anyone can swot to the test, but clearly the US system is much more enjoyable for a wider variety of students and they have a wider range of skills.

I think I'd like for my children to have a composite of the two. I dislike neverending homework and love tests because spiky activity allows for lots of self-directed exploration and the tests allow for evaluation. Though it's like that this is because I, personally, am quite capable with tests: I operate very well under time pressure and have an exceptional memory.

OK so back then using SAT words was good, and now using SAT words is bad. Got it.

Ubicomps colonial impulse, back for round 2.

AI checkers need to be sued out of existence.