My university gives “guidance” on the use of generative AI in student work. It’s not really guidance, because it simply doesn’t care — you can allow it or prohibit it. They even give us boilerplate that we can use in our syllabuses! If we want to prohibit it, we can say
In this class, the ability to [skill or competency] is essential for [field of study/professional application]. Because this course emphasizes [skill for development or specific learning outcome], using Generative AI tools [including those available to you through the University of Minnesota,] are not permitted.
If we allow it, we can say
In this course, students will [statement of learning outcomes, competencies, or disciplinary goals]. Given that Generative AI may aid in [developing or exploring course, discipline, professional, or institutional goals/competency], students may use these tools in the following ways:
The example allowing AI goes on much longer than the prohibitive example.
I will be prohibiting it in all my classes. So far, I’ve been pretty gentle in my corrections — when someone turns in a paper with a substantial, obvious AI, I tend to just flag it, explain that this is a poorly written exploration of the thesis, please rewrite it. Do I need to get meaner? Maybe. All the evidence says students aren’t learning when they have the crutch of AI. As Rebecca Watson explains, ChatGPT is bad for your brain.
I was doing a lot of online exams, thanks to COVID, but since the threat of disease has abated (it’s not gone yet!), I’ve gone back to doing all exams in class, where students can’t use online sources. My classes tend to be rather quantitative, with questions that demand short or numerical answers, so generative AI is mostly not a concern. If students started answering with AI hallucinations, it would be! I’m thinking of adding an additional component, though, an extra hour-long in-class session where students have to address an essay question at length, without AI of course. They’ll hate it and dread it, but I think it would be good for them. Even STEM students need to know how to integrate information and synthesize it into a coherent summary.
Another point I like in Rebecca’s video is that she talks about how she had to learn to love learning in her undergrad career. That’s also essential! Taking the time to challenge yourself and explore topics outside your narrow major. Another gripe with my university is that they are promoting this Degree in Three program, where you undertake an accelerated program to finish up your bachelor’s degree in three years, which emphasizes racing through the educational experience to get that precious diploma. I hate it. For one, it’s always been possible to finish the undergrad program in three years, we don’t put obstacles in front of students to get an extra year of tuition out of them, and we’ve always had ambitious students who overload themselves with 20 credits (instead of the typical 15) every semester. It makes for a killer schedule and can suck much of the joy out of learning. It’s also unrealistic for the majority of our students — every year we get students enrolled in biology and chemistry programs that lack basic algebra skills, because the grade schools are doing a poor job of preparing them. We have solid remedial programs at the same time we tell them they can zoom right through the curriculum? No, those are contradictory.
I think I’m going to be the ol’ stick-in-the-mud who tells students I’ll fail them for using ChatGPT, and also tells them they should plan on finishing a four year program in four years.
Why would people want to abandon the joy of writing? The prosaic listing of data can get boring fast, but the creative parts are another matter.
The Guardian
‘AI doesn’t know what an orgasm sounds like’: audiobook actors grapple with the rise of robot narrators.
https://www.theguardian.com/books/2025/jul/03/audiobook-voice-actors-ai-robot-narrators
another good piece from Democracy Now on AI, so much worse than I even thought:
“Empire of AI”: Karen Hao on How AI Is Threatening Democracy & Creating a New Colonial World
https://youtu.be/1NzW3o8zFEc?si=50t34MQHQ0Du8pyE
@ 1
Because it’s skibidi and totally ohio. Why should they waste their precious time that could be better spent playing Call of Duty or perusing brain-rot on Tik-Tok when a chatbot can do all your work for you? You don’t need to know anything, that’s for nerds without rizz.
Jebus, I’m turning into a fucking old fart. Maybe it’s time I finally take that bath with a toaster before civilization finally collapses.
I’ve been experimenting with ChatGPT and there are some neat things you can do but I have discovered that it gets a lot wrong.
I am using it to study poker and it has been helpful but I caught it saying some obviously wrong things like “ you are in the cutoff so you are last to act post flop”. Wrong! Button acts last post flop.
Then I had it analyze some hands I downloaded from PokerStars and it did provide some helpful insight, but also made up hands I didn’t play. Including a hand it said I lost but the hand as described by ChatGPT I would have won.
So even though I was able to improve my understanding of some key concepts I have to be wary of the feedback because so much of it is false.
I recently had a student ask me: “I working on problem … and I asked chat how to tackle it, it suggests … is that correct?”
My answer: “If you have to ask, you shouldn’t be using chat-GPT in the first place! What does the literature say about … ?”
Now he is probably asking “chat” to summarise the various papers I linked him, because he doesn’t want to read them himself, and then he’ll ask me: “is this summary correct?”
AIs don’t know anything, of course. They are just statistical engines. Some AIs are useful. I rely on an AI-driven technology sometimes called “type ahead” or “predictive typing”. I find it useful when I send text messages from my phone because the combination of tiny buttons, clumsy finger and aging eyesight make it difficult to tap precisely. If I tap “t” and it suggests “the” as one of three words, and that’s what I’m about to tap out, then I’m ahead.
I agree that using gen AI to replace research, thinking, synthesis, and composition is probably not a valid use of it.
By the way, not all results from an LLM are hallucinations. Some are, of course, and you need to be aware of that possibility before you put some junk information into a report, essay or article. Fact check your results. This is a factor of LLMs that may improve over time given the research in knowledge graphs (structured knowledge), RAGS and GragRAGs, and so forth.
Rebecca is absolutely right that one of the big draw backs to LLMs is energy consumption which makes them expensive.
As an experiment, I did some gen AI research recently. My son told me that the president of El Salvador, Nayib Bukele, has very high poll numbers in El Salvador. So I searched and indeed that’s the case. In fact, there’s even a Wikipedia article on the subject (here). At around 90% that is very high, and a glaring red flag to my mind. Doing straight searches on these numbers just turned up more articles on how high his numbers are.
So I dug deeper using DuckDuckGo’s Assist gen AI. They say the secret to using gen AI is the prompt…ie how you ask the question. I’m not real experienced with it, but I continued to rephrase my questions toward are this numbers credible, and after several turns the replies to my prompts about that started returning results about why those number might be skewed. Mostly it’s fear. The population is afraid to say anything negative about him for fear of retaliation, and they don’t trust promises of anonymity. Also, the media outlets doing the polls may skew the polls either through the questions they ask or by manipulating the results because they fear retaliation. This would not be something I would use directly for publication, but it might lead to more concrete and verifiable information to question Bukele’s popularity in polls.
“Lewis Black Rags on the State of Education and the Rise of ChatGPT University”
.https://youtube.com/watch?v=_9K05t0M7Eo
unbelievingdwindler @ 5
“I’ve been experimenting with ChatGPT…”
Says Lewis Black: You are supposed to experiment with drugs! 😁
I refused to use it on my last diploma as I felt I wouldn’t learn a damn thing if I used it. And I was right.
Occasionally I’ll use it to rephrase a sentence. Even then I’ll rephrase that rephrase to get something more satisfying.
Oh, how glad I am I don’t teach writing anymore! I taught junior high English for years and I had them write A LOT.
My posted slogan pounded into students’ heads: Writing is Thinking. Which is true.
@4 Akira MacKenzie wrote: Jebus, I’m turning into a fucking old fart. Maybe it’s time I finally take that bath with a toaster before civilization finally collapses.
I reply: PLEASE DON’T DO THAT. As troublesome a life as you may have, I have read, and appreciated, enough of your comments to know that you are a positive presence in this death spiral society. Your ‘checking out’ will just hasten the collapse.
As they say in the movies: ‘Put the toaster down, and back away from the bathtub’
‘Artificial Intelligence’ is Artificial but it is NOT intelligence.
I really respect that PZ is requiring his students to think, instead of using AI to be a ‘cliff’s notes’ on steroids so they can coast through a college career, ending up as ignorant at graduation as they were when they first entered college.
I just got a promotional AI review of a paper I posted on Academia. It was generally flattering but did provide a long list of references I could have cited and some suggested additions. The thing that AI ignored was that it was the text of a 15 minute conference presentation which was of necessity stripped down and very brief ad focused. The only useful thing is that the review uncovered a couple of papers I hadn’t come across which might prove informative. The promo came with an offer of another AI review I’m tempted to throw my 400+ page thesis at it to see if I can give it indigestion.
I work with a bunch of geeks and some of them are actually a sort of AI engineer…I don’t know all their specific areas of expertise. Someone posted this on a work channel. I’m not sure the claim is based on credible evidence. Trump’s “Big Beautiful Bill” likely created with AI – what does this imply?
I’ve played around with ChatGPT too. It’s fun to mess around with, but I don’t trust it. Especially after I tried asking it how many instances of the letter “r” there are in the word “Strawberry.” It told me two. I had to do a lot of coaxing until it agreed it was three.
@12
I’m fine. I just have a very, very morbid and dark sense of humor.
PZ — I thought 3 years was standard for most Bachelors degrees, with an extra year for Honours. Is that accelerated program including Honours in 3 years?
As for AI in education, I can see situations where it can be a helpful adjunct, but I don’t think students seem to realise that if they can pass a degree using AI instead of understanding the material and putting in the work to write it up, then why would anyone bother to employ them?
There ya go. garydargan has just posted a useful AI function: digging up references you might have missed in your reading of the literature. Of course, it all depends on the quality of the AI’s search capabilities and the gaps in the investigator’s knowledge base, and even then a human needs to look at the papers to see if they add anything, but this is exactly the kind of mass-data crunchwork that should be given to automatons (of demonstrable utility).
@17 Akira MacKenzie wrote: I’m fine. I just have a very, very morbid and dark sense of humor.
I reply: that’s good to hear. A dark sense of humor is probably very appropriate in this society. In keeping with that thought: I fear that humanity is going to be toast soon enough,
Stay sane, stay safe, stay humane, everyone.
I found this rather interesting and informative.
Victorian Ladies Don’t Poop: The London History Show — https://www.youtube.com/watch?v=SJ_DsLtE-mM
by Jenny Draper
https://calmatters.org/education/higher-education/2025/06/ai-detector/
This is an article in Cal Matters, a nonprofit news organization. It is a somewhat longread about a product called Turnitin, which claims to detect AI compositions. Multiple campuses of multiple institutions pay big bucks for their professors to be able to use it.
tl:dr It makes mistakes!
Perhaps if young people could be gently reminded that LLMs (1) exist primarily to enrich the shareholding class, (2) reinforce racial and other biases inherent in their training corpora and (3) emit vast quantities of carbon dioxide in doing so so, they might start to look less “cool” (yes I am over 40) and more skibidi.
correction: “in doing so”
robro @7,
“So I dug deeper using DuckDuckGo’s Assist gen AI”
I woke up one day last week, booted up my laptop, and there it was, this DDG Assist thingie, enabled by default, unasked, unwanted, obviously must have snuck in with some recent update. This is what I hate most about AI, that they throw it at you everywhere without asking, and the worst thing is, it doesn’t work. It makes shit up, the more “powerful” it gets the more it hallucinates. Absolute garbage, and I instantly abandon any piece of software that does that.
“I’m thinking of adding an additional component, though, an extra hour-long in-class session where students have to address an essay question at length, without AI of course. They’ll hate it and dread it, but I think it would be good for them. Even STEM students need to know how to integrate information and synthesize it into a coherent summary.”
This is not a novel approach. We had hour long essay exams in biochemistry classes and molecular biology classes when in home computers were not really a thing (the wordstar vs wordperfect decade), I’m willing to bet PZ had the same in many of his classes. Semi-recently I used in class essays for exams in my advanced microbiology class. I moved to take home essays because I wanted students to have more time to develop their thoughts more deeply and consult the literature if they choose to. I am now reconsidering the nature of these in the face of AI, but when I test AI with my questions, I usually get superficial responses that amount to a high school student level of superficiality. I am in the camp at this point that if you ask AI to write a paragraph on a question that is basically a multiple choice question (regurgitate facts) it will do it, but if you ask higher level (Blooms taxonomy) questions, AI fails to adequately provide an appropriate response.
It’s usually 4 years in the US. Europe is different. I honestly think it’s because our grade schools don’t adequately prepare students for college. I was in a specifically academic track in high school which demanded 3 years of a foreign language, trig, chemistry, and biology, and I came out of that prepared to breeze through the first year of college — getting a degree in 3 years would have been reasonable.
Most high schools don’t accomplish that. You can graduate with minimal exposure to algebra, which means any chemistry you can take is a joke, and we’re American…who needs another language?
Gary Marcus on why LLMs will never lead to AGI.: tl/dw – they know nothing about the world itself, only about language/images – and that only statistically. AGI will require symbolic representations of the wold, and rules for manipulating them. [Among other tihngs, I would add.]
Not sure why anyone is surprised that the group that used AI from the start got lazier and retained less information… isn’t that kind of a self-selecting group? If you are already lazy and intellectually incurious, why wouldn’t you be drawn to a tool that is pretty much designed for lazy people?
Fast forward 10-15 years when all these AI “writers” are pissing and moaning about how they can’t get ahead in their careers, etc. Maybe it’s because you don’t know anything and all your communication sounds like the thin gruel that is AI writing?
Has it occurred to anyone that AI writing, if it is continually scraping the internet, will get progressively worse as a writer, since most people can barely string two grammatically correct sentences together as it is? And forget spelling, a ridiculous number of people clearly have no idea of what a homophone is or that it makes a difference as to which one you use. (Phase/faze and discreet/discrete are particular pet peeves.)
AI generative writing is the 21st century version of both Orwell’s newspeak (from 1984) of Bradbury’s firemen in Fahrenheit 451. I always though his coda/afterword was one of the most brilliant pieces of writing of the 20th century.
https://katherinesmithth.weebly.com/uploads/9/7/1/7/9717809/coda_from_ray_bradbury_for_fahrenheit_451.pdf
(The other reason I liked Fahrenheit 451 so much is because it is also a clear ode to walking in the outside world. Another thing that has fallen out of fashion.)
https://richardswsmith.wordpress.com/2024/10/08/a-very-short-story-hacked-from-an-afterword-to-fahrenheit-451/
AI generative writing, whatever the LLM that is used, basically runs on a large thesaurus and incidence of use. So the more a particular word is used (correctly or incorrectly!) the more likely AI is to use it (correctly or incorrectly) Neologisms and novel uses of unusual words are barely a blip, which ultimately leads to an ever-smaller vocabulary and constriction of thoughts and ideas. Hence, newspeak. You have to name things before you can discuss them. (See Betty Friedan’s “the problem with no name”.)
https://pmc.ncbi.nlm.nih.gov/articles/PMC2920960/
Reading AI literary slop is like looking at Maxfield Parrish paintings or Roger & Martin Dean illustrations without color.
I’d like to point out that there are no shortage of teachers and professors who use AI to grade papers. Let’s not lay 100% of the laziness and incuriosity problem at the feet of the students.
silvrhalide:
“If you are already lazy and intellectually incurious, why wouldn’t you be drawn to a tool that is pretty much designed for lazy people?”
All tools are definitionally designed for lazy people; after all, pick and shovel will did a hole just as well as an excavator, pen and quill will write just as well as a word processor. Paper spreadsheets (I worked with those!) instead of electronic ones.
Photography instead of paint brushes. Et cetera.
(Me, I like them because I am intellectually curious, not the other way around)
“Has it occurred to anyone that AI writing, if it is continually scraping the internet, will get progressively worse as a writer, since most people can barely string two grammatically correct sentences together as it is?”
Only over and over since the very beginning.
“Reading AI literary slop is like looking at Maxfield Parrish paintings or Roger & Martin Dean illustrations without color.”
It’s not exactly a mature technology, is it?
I see this all the time; people assume that what we have now won’t improve.
—
Oldie but a goodie: https://www.pcworld.com/article/532605/worst_tech_predictions.html
As a grad student in math, one summer I had to teach a remedial basic algebra course. I dreaded it when it was first assigned to me. But it turned out to be a joy. All of the students were older, returning to education after some years working. And all were very motivated. That was the joy part. Except for one poor lady that just could not get it, and seemed resigned to failing the course over and over. Why she kept trying when it seemed clear even to her that she wasn’t going to make it, I have no idea. In her case, I was very sad as nothing I could do seemed to get through.
@25
THIS
But it’s getting harder and harder to find search engines or software without it. Even if you aren’t searching for something, half the online ads seem to be for AI crap like Grammarly (which isn’t as grammarly as it likes to think it is).
@31 Maybe because she’s not a quitter? Or is simply determined to learn it, no matter how long it takes? I did fine with differential calculus but really hit the wall with integral calculus. Powered through anyway, Three times.
Eff. What the hell happened to my first link?
“gender of rap”
Maybe he’ll get a dictionary for Christmas.
Not all tool users are lazy. Some use them to achieve greater results than would otherwise be possible.
Bonus round
In the news: ‘I’m being paid to fix issues caused by AI’ — https://www.bbc.com/news/articles/cyvm1dyp9v2o
—
AI is making me extra money, says Sarah Skidd, a product marketing manager who writes for tech and start-up companies.
In May Ms Skidd was approached by a content agency to urgently rework website copy that had been produced via generative AI for a hospitality client.
What was supposed to save money had, instead, caused a host of problems.
“It was the kind of copy that you typically see in AI copy – just very basic; it wasn’t interesting,” says Ms Skidd.
“It was supposed to sell and intrigue but instead it was very vanilla.”
Ms Skidd spent about 20 hours rewriting the copy, charging $100 (£74) an hour. Rather than making small changes, she “had to redo the whole thing”.
Ms Skidd, who lives in Arizona, is not worried that businesses are switching to AI, like ChatGPT, rather than using copywriters like herself.
“Maybe I’m being naive, but I think if you are very good, you won’t have trouble.”
For now, she’s hearing of writers whose main role now is to fix copy churned-out by AI.
“Someone connected with me and said that was 90% of their work right now. So, it’s not only me making money off such missteps, there’s other writers out there.”
Well, that’s the key question. Gary Marcus (see link @28) says LLMs are a mature technology, and are not going to improve substantially, because of a fundamental conceptual flaw. While Geoff Hilton thinks they are already conscious! (I suspect he – even he, who was key to designing the basic technology) – is being fooled by the fact that commercial LLMs are designed to keep the user coming back for more, so if they get a response to a particular kind of output, they will produce more such outputs. It’s repeatedly said that even the designers don’t understand how LLMs do certain things. But (assuming this is true, and it may not be, saying so could be a marketing ploy), the result may be that even the designers are being misled by the human tendency to anthropomorphise machines, assigning intention when it is absent.
KG, my instance of Copilot was recently improved to 32k lexical tokens for session context, up from 8k a few weeks ago.
From Wikipedia: “Although decoder-only GPT-1 was introduced in 2018, it was GPT-2 in 2019 that caught widespread attention because OpenAI claimed to have initially deemed it too powerful to release publicly, out of fear of malicious use.[13] GPT-3 in 2020 went a step further and as of 2025 is available only via API with no offering of downloading the model to execute locally. But it was the 2022 consumer-facing chatbot ChatGPT that received extensive media coverage and public attention.[14] The 2023 GPT-4 was praised for its increased accuracy and as a “holy grail” for its multimodal capabilities.[15]”
So, still only a few years old, overall, and ChatGPT is three years old.
One can certainly notice improvements.
Also, remember our earlier discussion regarding modularising and iterative verification?
You wrote: “Gary Marcus (see link @28) says LLMs are a mature technology, and are not going to improve substantially, because of a fundamental conceptual flaw.”
I think that’s conflating AI and LLMs. They are one type, the best at certain tasks, but… well, take a mature algorithmic tool like Mathematica, which is 38 years old. And it does good math.
“Just under six months ago (176 days ago, to be precise) we released Version 14.1. Today I’m pleased to announce that we’re releasing Version 14.2, delivering the latest from our R&D pipeline.
This is an exciting time for our technology, both in terms of what we’re now able to implement, and in terms of how our technology is now being used in the world at large. A notable feature of these times is the increasing use of Wolfram Language not only by humans, but also by AIs. And it’s very nice to see that all the effort we’ve put into consistent language design, implementation and documentation over the years is now paying dividends in making Wolfram Language uniquely valuable as a tool for AIs—complementing their own intrinsic capabilities.”
Sorry, one more point.
“It’s repeatedly said that even the designers don’t understand how LLMs do certain things.”
I just put it to Copilot:
That’s what I woulda said.
FWIW, https://garymarcus.substack.com/p/generative-ais-crippling-and-widespread
(I did not watch the video, but I sought out the actual source material)
Featured in Vox magazine: https://www.vox.com/the-gray-area/418793/chatgpt-claude-ai-higher-education-cheating
–extract–
James Walsh is a features writer for New York magazine’s Intelligencer and the author of the most unsettling piece I’ve read about the impact of AI on higher education.
Walsh spent months talking to students and professors who are living through this moment, and what he found isn’t just a story about cheating. It’s a story about ambivalence and disillusionment and despair. A story about what happens when technology moves faster than our institutions can adapt.
[…]
Q: Let’s talk about how students are cheating today. How are they using these tools? What’s the process look like?
A: It depends on the type of student, the type of class, the type of school you’re going to. Whether or not a student can get away with that is a different question, but there are plenty of students who are taking their prompt from their professor, copying and pasting it into ChatGPT and saying, “I need a four to five-page essay,” and copying and pasting that essay without ever reading it.
One of the funniest examples I came across is a number of professors are using this so-called Trojan horse method where they’re dropping non-sequiturs into their prompts. They mention broccoli or Dua Lipa, or they say something about Finland in the essay prompts just to see if people are copying and pasting the prompts into ChatGPT. If they are, ChatGPT or whatever LLM they’re using will say something random about broccoli or Dua Lipa.
AI is a tool. If you become good with it, you will write better, clearer texts. If you try to use it to avoid work, you will produce bad texts with too much verbiage and errors that are very obvious.
John: could broccoli be making Dua Lipa sexier? And maybe the association with Dua Lipa could make broccoli sexier…?
Dictionaries used to do that, too. Well, sorta.
https://en.wiktionary.org/wiki/nihilartikel