We have spotted quite a few students using generative AI in their essays this summer and applied standard academic misconduct proceedings, though in most cases the work was so bad they would've failed anyway.
Today I learned of one whose use was sufficiently extensive that they will fail their degree.
I am wondering if this is *the first time a student has failed a whole degree for using AI*? Would love to hear about other cases. If you want to tell me in confidence, my Session ID is in my Bio
This is a really good question. They are failing for not meeting the academic standards we clearly lay out and spend a lot of time teaching them about attribution of sources, whether quoted or paraphrased, not passing of the work of others as your own etc.
That it was done with AI rather than cut and paste from Wikipedia makes no difference to that (though the AI does some weird stuff making up references!)
@mikeg
By reading what they produced!
When you read >100,000 words of student produced work in your specialist area every year, you can spot something fishy.
We also interview before making a final decision and when they cannot even answer the most basic questions on what they have written ...
It is no different from detecting any other form of plagiarism.
@awaterma
You have made a lot of assumptions there!!
We don't give credit for rote knowledge. Our marking criteria only mention understanding of the material taught, argumentation, structure, writing and referencing.
The tells are (1) not drawing on the material taught but other sources, (2) making up sources, (3) coherence over 1000s of words etc. (4) and a writing style of a level higher than the student produces in other work.
We always have an oral to check before imposing a fail mark.
@awaterma
And for what it is worth, I am far from 'freaking out' - we have discussed at length and are happy with the idea of using AI generated text as a basis, so long as it is then edited in such a way that it meets our academic standards.
This student failed because they cut and pasted it rather than used it as a source in an appropriate way.
From the context of this thread it ultimately sounds like the students are taking the output as-is with no attempt to engage with the source material. A generated piece of writing should be penalized, regardless of if by a writing service or an AI. The point of education is to take in knowledge so you may benefit from it and not to just let machines/hired writers do everything.
I am puzzled ... who said rote learning was of any value at all in any sane educational system?
As to how do we know? Well, we know within the standards of proof required by our processes for detecting academic misconduct.
You might ask: how do we 'really know' anyone convicted of a crime was guilty? Balance of probabilities? Beyond reasonable doubt? These are the relevant concepts to be applying, depending upon the legal or procedural context.
@tomstoneham @awaterma
I don't use rote learning in my job at all now that you mention it. Academia readied me for how to take critical texts to the theory of what I am to apply as an engineer and how to make sense of the fundamentals to the systems that I use.
I couldn't tell you off the top of my head what is the effect of adding new pieces to a system without looking up the fundamentals again.
@DukeCarge @tomstoneham @awaterma yes! Plus, there's no real value to rote learning anymore anyway. I tell my junior mentees not to bother to intentionally memorize any fact. If you use it often enough to be worth memorizing, your brain is going to do it anyway without trying. What matters is the how and why. Learn those and you can put together any idea.
That's why the discussion of accuracy of "AI detectors" is silly. "Questioned student doesn't understand at level of essay" is very reliable.
@tomstoneham technologist turned graduate law+policy student here, can confirm.
we have weekly short essays to post online for class discussion; the word soup in a few recent posts is 🤦🏻♀️. you really don't need an edtech surveillance tool to see it. can only imagine their full papers...
that said, not sure it's only those who would turn in failing work anyway. i think some students -- like general public -- probably use it like a search engine.
frustrating to see, but not unexpected.
@josh
They were interviewed and gave a written response to the evidence presented. (We do take care in these matters!)
Denied it in general but couldn't say anything about specific points of evidence or explain the material in the essay.
@josh @tomstoneham Not exactly on topic, but this might interest you, https://social-epistemology.com/2023/03/29/24-philosophy-professors-react-to-chatgpts-arrival-part-i-ahmed-bouzid/.
A two part interview of two dozen American(?) philosophy professors about ChatGPT's arrival. Some serious caveats and considerations, but also a hint of maybe reckless techno-optimism. Outsourcing invaluable parts of deep learning "because we can" worries me in particular.
For someone like myself who makes their living in the field of Human Language Technology, two dates from the past decade or so have stood as watershed moments in that field: October 4, 2011 when Apple’…
@tomstoneham
I had great luck turning this problem into a strength: I asked students to generate a bit of their paper using an AI, then critique its output. Details:
https://hachyderm.io/@inthehands/109479808455388578
Among other things, implicitly saying “the AI’s output is pretty much guaranteed to suck” sent a useful message about using it to cheat.
Attached: 3 images OK, trying an experiment with my Programming Languages class! • Have an AI generate some of your writing assignment. • Critique its output. Call BS on its BS. Assignment details in screenshots below. I’ll let you know how it goes. (Here are the links from the screenshots:) Raw AI Text: https://gist.github.com/pcantrell/7b68ce7c5b2e329543e2dadd6853be21 Comments on AI Text: https://gist.github.com/pcantrell/d51bc2d4257027a6b4c64c9010d42c32 (Better) Human Text https://gist.github.com/pcantrell/f363734336e6063f61e451e2658b50a6 #ai #chatgpt #education #writing #highered #swift #proglang
@inthehands
We thought about doing that but didn't get time.
Personally I am happy for AI to be used as a tool (we don't ban spellcheckers and autocorrect or even proofreaders) so long as the student is taking the final decision about what goes in and what does not. Editorial responsibilit, as it were.
That is how law firms like Allen & Overy use it.
Of course, in many disciplines it is currently a pretty rubbish tool!
LLMs will 100% be used. “Bans” will mean only some kids get caught and some kids get away with it, and others don’t get the benefit of using it. Probably better to adapt — how can LLMs make writing better. Where to go from that 0th order draft.
If you really want something written BY the student & fair to other students, why not have ‘handwritten’ (typed on a school machine), while students are supervised?
@tomstoneham I've not heard of anybody failing a degree for cheating via LLM, but I've certainly seen a huge increase in the use of LLMs to cheat. I see some of the other commenters are wondering how you could tell ... well in one case I encountered, a student's essay literally included the quote "in my experience as an artificially-intelligent chatbot..."
It seems as though many cheaters don't bother to proofread 😆
@andrei_chiffa
By reading it. Most student cheating is obvious to a specialist who knows what they have (and have not) been taught and has read millions of words of student attempts to write about it.
But there is a more fundamental point I need to write up in detail:
University teaching is basically a form of intensively supervised reinforcement learning on a carefully curated, small data set aiming to produce a specific capability.
It is obvious the AI didn't go to class *here*
So it's not as much generative AI detection as it is detection of the fact the student did not attend or even familiarise themselves with the class, correct?
@andrei_chiffa
It is easy to overlook how well trained on a large dataset academics are. We each read around at least a million words of student work on our specialist area every year. Some of us have been doing that for decades.
Our pattern recognition for 'not produced by a student' is pretty good 😆
@tomstoneham @andrei_chiffa
hu thats intresting ...
when I studied ( political science ) of course we had courses with preselected textes etc ... but in the work we where expected to go further then that. So that at the end what we talked and read about in class was only a tiny amount of the work I would present
Its intresting that none of you talks about the increasing workload studends have to fullfill. Espacially with inflation where most students i know work 2 jobs or more.
how does this effect the decision to use tools like chat gpt?
@generic @andrei_chiffa
Your two points are related. We don't penalise going *beyond* the set texts but we do require that the essay demonstrates understanding of what was actually taught.
We don't *require* going beyond the set texts because that rewards those who don't have jobs or caring responsibilities.
@gibbondemon
That would be really sad. In person exams are known to be biassed towards socially privileged and neurotypical students.
Making that regressive shift would mean the white, male, neurotypical bias of AI had won even when it wasn't being used!
@tomstoneham I’m really curious about what are acceptable and unacceptable uses of generative AI for students.
My most productive professional use is to improve the readability of analysis I’ve produced. I also use it to help spot patterns in data (and trivial stuff like reformatting data structures).
Would I get failed for these as a student? Not looking for an argument, just to be clear: I’m really interested in how academia’s facing the challenges introduced by this technology.
@tom Plenty of acceptable uses and we discussed this at department meeting before any marking.
Uses of AI which also constitute good old fashioned plagiarism are the problem here.
Neil Mackenzie
Tom Stoneham
Doctor Memory
Kate Nyhan is changing servers
jnyrose
Michael Gorton
Dr N. Brodie
Andrew Waterman
Alex Froio
Chuck
zceline
Josh
Lupposofi
Rebecca
🇨🇦 D'Arcy Norman
Paul Cantrell
Blake C. Stacey
legraLeGra
Dr. Fyzziks
Andrei Kucharavy
O.O
Andrew Knighton
Thomas Buck 