AI writing is really good

[MillieDynamite]

What's Dystopian for one group is another group's Utopian dream.

Was it though? Dystopias are just more fun to write. Utopia stories get dull quite fast.

 

[stuffscribbled]

There's a very simple way to deal with generative AI in Universities.

It's called exams, and it's worked well for centuries.

We need to move away from exams as the default mode of assessment. They are hugely disabling and are an artificial emvironment. When was the last time you had to work in exam conditions as your job? Fear of cheating is often used as a scaremongering way of refusing to offer reasonable adjustments or more inclusive ways of assessment. Currently at the HEI where I work the number of academic integrity cases of suspected AI is very small. Maybe it becomes more prevalent, and yeah finding a truly inclusive form of assessment that is robust and can be trusted is a long and difficult journey, but the answer isn't defaulting to exams.

 

[Bramblethorn]

Would not undergrads be ~1 standard deviation above the mean writer? About 40% pf 18-24 year olds are in college. Sure, we can want better, but there isn't much of a pool of better out there.

Hence my comments about needing to invest in education.

 

[Bramblethorn]

I was too lazy to come up with a response on my own, so I had AI write one for me:

If you couldn't be bothered writing it, why would anybody be bothered reading it?

AI is just a tool—it's up to us to treat it like one, not a threat.

Anybody who's ever designed tools, anybody who's ever used tools should be able to see what's wrong with this argument. There's no "just a tool".

A tool is designed for a specific function, and (if the designer's competent) with specific types of users in mind. I have a pair of scissors that are designed for right-handed people who want to cut ordinary cloth; I have another designed for right-handed paramedics who want to cut through heavy clothing and seatbelts. Using one of those for the other one's purpose is likely to end badly, as is giving them to a left-hander.

The more complex and powerful a tool becomes, the more important UX considerations become. A well-designed tool is made to connect user intentions as closely as possible to what the tool does, and catch obvious mistakes. If I try to divide by zero on my calculator I will get an error message, because that says I'm trying to do something it wasn't designed for. Training people to use their tools is important, but it's not a cure-all for bad design and neglected UX. Actual tool manufacturers have an interest both in getting the UX right and in educating users, because if somebody misuses their tool and botches a job/loses a finger, that's bad for business.

With generative AI, the function for which it's designed is primarily to impress venture capitalists and attract funding; everything else is an afterthought. Yes, there are applications for which generative AI is a useful and reasonably reliable tool - but if it were only marketed for those applications, the backers would go broke. Here it's the "teaching users how to use it" that's bad for business, which is why you cannot depend on people learning to use it correctly in the way they might, say, a bandsaw or a programming language.

 

[gunhilltrain]

The sequel was Futureworld.

I completely missed that one. I see that robots are now running the control center too, which seems like a bad idea.

 

[gunhilltrain]

The world was always meant to end in dystopia.

It does assuming that every civilization eventually comes to an end. That doesn't stop people from trying it again.

 

[gunhilltrain]

Was it though? Dystopias are just more fun to write. Utopia stories get dull quite fast.

It does look a bit dull. At least you can get pizza.

Utopia Parkway

 

[intim8]

The issue is that it's about 200% easier to manage classes by having students do things digitally, instead of on paper. In this case, the path of least resistance for the professors is also the one that encourages AI use by students.

So have the college IT people set up a virtual machine server. Give every student an account, and only accept papers typed in on the VM software. Put a keylogger on it, and check *that* for cadence patterns that look like AI, or just pasting in large chunks.

I'm not an IT guy, so I probably got some details wrong, but the basic approach seems like it would work.

 

[OffNSFW]

In the "Rip Me To Shreds" thread, someone had Copilot rewrite a snippet I'd posted for criticism. While you could argue about whether or not the prose was better, what's undeniable is that the AI added details that didn't make sense, and removed details so that much of the rest didn't make sense.

If it can't keep 600-ish (I think) words straight, I doubt it's going to write a cohesive story of even a few thousand words.

This is the only post in this thread that I find adds anything to the conversation that hasn't already been repeated a dozen times elsewhere on this site, because it gets at something very important: A.I. is dumb and much of what it generates is bad.

And A.I. can generate bad output a lot faster than humans can generate good output, and all that bad output becomes part of the full data set that A.I. programs are trained on. I'm not sure what has people convinced that A.I. results will necessarily improve in quality, other than blind fatalism and tech industry hype. It's already getting noticeably lower in quality.

The question is, will humans who uncritically adopt A.I. as their latest technological crutch become dumber along with it?

 

[Yarglenurp]

This is the only post in this thread that I find adds anything to the conversation that hasn't already been repeated a dozen times elsewhere on this site, because it gets at something very important: A.I. is dumb and much of what it generates is bad.

And A.I. can generate bad output a lot faster than humans can generate good output, and all that bad output becomes part of the full data set that A.I. programs are trained on. I'm not sure what has people convinced that A.I. results will necessarily improve in quality, other than blind fatalism and tech industry hype. It's already getting noticeably lower in quality.

The question is, will humans who uncritically adopt A.I. as their latest technological crutch become dumber along with it?

No one here is speaking as a trained AI scientist, so my expectation of anyone adding to the conversation is pretty low.

But you’re basically talking about the dead internet theory. Which is valid, but the guys who own the models actually do control the data input on these things, so they’re not going to let that effect run too hard. The guys being paid $500K with a $1 million signing bonus isn’t going to overlook the possibility of bad input data.

 

[Voboy]

So have the college IT people set up a virtual machine server. Give every student an account, and only accept papers typed in on the VM software. Put a keylogger on it, and check *that* for cadence patterns that look like AI, or just pasting in large chunks.

I'm not an IT guy, so I probably got some details wrong, but the basic approach seems like it would work.

Yeah.

I'm not optimistic. The deans I knew at most universities were not renowned for being into major sweeping changes, nor were they eager to hunt down and/or program new tools for something like that.

 

[XerXesXu]

This is the only post in this thread that I find adds anything to the conversation that hasn't already been repeated a dozen times elsewhere on this site, because it gets at something very important: A.I. is dumb and much of what it generates is bad.

And A.I. can generate bad output a lot faster than humans can generate good output, and all that bad output becomes part of the full data set that A.I. programs are trained on. I'm not sure what has people convinced that A.I. results will necessarily improve in quality, other than blind fatalism and tech industry hype. It's already getting noticeably lower in quality.

The question is, will humans who uncritically adopt A.I. as their latest technological crutch become dumber along with it?

What SS left out was that Copilot's rewrite was used as a criterion for how bad the snippets were, not their merit. Snippets by several authors were used, LC68, EB and SS. Copilot rewrote the first two in exactly the same number of words as their snippets and misunderstood none of their usages of words. The few changes made neither added nor subtracted from their snippets. SS' snippet was rewritten with 30% fewer words and several errors. Those errors were in the nature of what you'd expect from an insentient machine reliant on appropriate use of words and semantic relationships between words to 'understand' what was written. SS had a leash slither, snakes slither, so Copilot also had it coil. In a particularly ill formed sentence 'he wasn't sure whether that was the foot -- it was a foot on his neck, he decided -- or the ground.' Copilot tossed a coin and decided it was the ground.

Some will pass the Copilot rewrite test, some will fail it. Passing isn't a mark of merit, only basic competence and an ability to use appropriate words and grammar to describe what's happening. If Copilot understands you, human readers will also understand you.

 

[StillStunned]

Some will pass the Copilot rewrite test, some will fail it. Passing isn't a mark of merit, only basic competence and an ability to use appropriate words and grammar to describe what's happening. If Copilot understands you, human readers will also understand you.

Yet tellingly no-one else who criticised the snippet expressed any difficulty understanding it.

 

[XerXesXu]

Yet tellingly no-one else who criticised the snippet expressed any difficulty understanding it.

I did.

‘blinking away tears‘

You say he wasn’t crying, his eyes were watering, but you didn’t write ‘eyes watering.’ Why? If you had, the readers would have known he wasn’t crying.

‘he looked up just in time to catch a fist on his chin.’

Ever kept wicket? How many balls did you catch on your chin? You may, from time to time, have been caught on the chin-guard by a ball.

‘There was a tall form standing over him, elongated to almost comical proportions. Or were there two forms? No, one was a stone.’

What are ‘comical proportions’? Was he seeing double? What’s happening? The reader should be able to make sense of it, but they could only understand that ‘There was a tall form standing over him’ and needn’t decrypt the rest.

‘It was all very absurd, and he would have chuckled but there wasn't any air, and he found himself coughing instead.’

Enough air to cough but not laugh? I expect many readers smiled wryly at, ‘It was all very absurd’.

‘he wasn't sure whether that was the foot -- it was a foot on his neck, he decided -- or the ground.’

We’ve done that.

‘Forcing his body to obey, he twisted. The weight scraped from his neck and he rolled over.’

This appears to be his escape. You say No, it was the opponent taking his foot off the protagonist neck. That’s difficult to realise from what you wrote.

I'd guess that many, brought up in an age of dumbing down and silver stars for all, felt unable bring the honesty you'd requested to the table.

 

[alohadave]

Yeah.

I'm not optimistic. The deans I knew at most universities were not renowned for being into major sweeping changes, nor were they eager to hunt down and/or program new tools for something like that.

It's also not as trivial to implement as it's made out to be. Sure, the concept is simple, but lots of simple concepts are a pain in the ass to implement.

This is a management problem, not a technology problem.

 

[StillStunned]

SS had a leash slither, snakes slither, so Copilot also had it coil.

So I can't use the word "slither" unless I'm referring to a snake? And then it will automatically be coiled, even if it's described as slithering?

That's like saying you can clench your fist, and you can clench your jaw, so it's the author's fault if AI says "he clenched his fist until his teeth ground".

‘blinking away tears‘

You say he wasn’t crying, his eyes were watering, but you didn’t write ‘eyes watering.’ Why? If you had, the readers would have known he wasn’t crying.

Man gets punched in the face. Man blinks away tears. I think most readers can make the connection that the punch has made his eyes water, not that he's crying.

‘he looked up just in time to catch a fist on his chin.’

Ever kept wicket? How many balls did you catch on your chin? You may, from time to time, have been caught on the chin-guard by a ball.

Ever heard the expression "to catch a fist"? It means being punched.

‘There was a tall form standing over him, elongated to almost comical proportions. Or were there two forms? No, one was a stone.’

What are ‘comical proportions’? Was he seeing double? What’s happening? The reader should be able to make sense of it, but they could only understand that ‘There was a tall form standing over him’ and needn’t decrypt the rest.

‘It was all very absurd, and he would have chuckled but there wasn't any air, and he found himself coughing instead.’

Enough air to cough but not laugh? I expect many readers smiled wryly at, ‘It was all very absurd’.

‘he wasn't sure whether that was the foot -- it was a foot on his neck, he decided -- or the ground.’

We’ve done that.

‘Forcing his body to obey, he twisted. The weight scraped from his neck and he rolled over.’

This appears to be his escape. You say No, it was the opponent taking his foot off the protagonist neck. That’s difficult to realise from what you wrote.

You seem to find these passages unclear, and that's what I asked about, so fair enough.

I'd guess that many, brought up in an age of dumbing down and silver stars for all, felt unable bring the honesty you'd requested to the table.

Given the rest of the feedback in the thread, on my own snippets and on other people's, I really don't think anyone was being kind.

But this really belongs in the other thread, since it doesn't concern the AI discussion.

 

[Kadyia]

If you couldn't be bothered writing it, why would anybody be bothered reading it?

Initially, I point out that you not only read what I wrote, but you provided a thoughtful and well laid out argument against the “AI written” points. I also point out that if you “cut and paste” what “I wrote” was written into isgen.ai, it will come back as 100% human written, 0% probability of being AI generated.



This presents three options that I can see:

• I have an absurd and inappropriate sense of humor, and I wrote it.
• AI wrote it and I did some voodoo (swapped words around, moved semicolons) and it turned into an undetectable piece of AI generated text.
• AI detection sucks and you will get false positives and false negatives.

I do know that there are many, many AI detectors out there and you’ll get different numbers and assessments depending on which you choose to use. As sort of a control group, I ran a snippet of the Story of O through their AI detector (and I hope we can agree that the beautiful and oh-so talented Anne Desclos was not using Generative AI in 1954 (though I’d love to think of her as a mysterious time-traveler). It told me that her writing had a 43% probability of being AI generated.



What if Anne had turned her Story of O in as a writing assignment and the AI generator came back at a 63% probability of being AI generated instead of 43%? Would the teacher mark her down a grade? Fail her? Perhaps, this teacher sets his AI generation number at 65% and he merely scolds her, “I know you used AI, but I’ll let you slide this time.” Of course, if she’s a modern day girl, then she will and should get called on her blatant plagiarism of a deliciously decadent author. It’s a little easier to detect actual plagiarism.

Back to my point. We live in the real world. And the detection of Generative AI isn’t quite there yet. Certainly, it’s not-quite-there in terms of AI detection with students who put effort into their work and don’t just cut-and- paste what ChatGpt, Gemini, Grok or whatever the currently popular ice-cream flavor spits out.

I know students and authors exist who get accused of using AI when they haven’t; and that’s rather tragic. It’s a different kind of tragedy, when a student or author gets away with using Generative AI and gets a better grade than some who did it the old fashioned way (laid by the pool and tanned, while handwriting out your essay). We live in the real world and students and others will use AI. If AI is a “tool,” as you point out, it’s also a tool akin to alcohol (in the sense that the Eighteenth Amendment wasn’t successful in prohibiting its use) or firearms (where the Second Amendment runs interference on preventing access).

So, if we can’t prevent its use (I’m not a huge fan of honor codes, because I believe there is a certain percentage of people who are corrupt, and I hate setting up a system where the intelligent and clever corrupt can get one over the hard working honorable). And if the current state of AI detection is ambiguously uncertain, we need to address that situation. Now an article actually came out yesterday in the Guardian which makes the point I wish I made in a much more coherent fashion. It’s “AI cheating is overwhelming the education system – but teachers shouldn’t despair” by John Naughton, Saturday, August 24, 2024.

His tagline is that “with adjustment to the way we teach students to think about writing, we can shift the emphasis from product to process.” That’s what I wanted to say, but wasn’t clever enough to say. He argues, “There is currently no reliable way to be sure if a student has used generative AI in writing an essay.” And I agree.

You make a very strong argument about the importance of user experience in tool design, especially as they become more complex. But I think you’ll agree that regardless of design, tools, like pharmaceuticals, often end up getting used in ways that were not foreseen or intended. With Generative AI, I tend to agree with Allison Gropnik, Distinguished Professor of Psychology, Affiliate Professor of Philosophy and member of the Berkeley Artificial Intelligence Research Group:

“LLM’s are extremely effective cultural technologies, analogous to writing, print or internet search. They summarize and transmit information that humans have already learned but they are not designed to recover the structure of the external world. In contrast, children are extremely effective truth-seeking agents, exploring and experimenting.” (Naughton cites her in his article).

He also cites Josh Brake (who writes the Subtack newsletter styled, the Absent-Minded Professor. Professor Brake has said,

“There are certainly exceptions, but the pedagogical strategies that the availability of LLMs obliterates are those fundamentally based on the distrust of students. It has not killed writing, but it has killed a certain kind of way of motivating writing. If your students didn't already see the value of writing as a process by which you think, then of course they will be curious about farming the labor out to an LLM. And if writing (or any other task for that matter) is truly only about the product then why not? If the means to the end are unimportant, then why not outsource it?”

Brake said a lot more worth reading. These articles and newsletters and talks do a much better job than I do grappling with the issue. But for me, I come back to the fact that Generative AI exists. And teachers have to deal with it, because I don’t see “just say no” working any better with Generative AI than Prohibition did with alcohol.

And I do feel Generative AI has a little more use than just grist for venture capitalists to raise money. While it’s terrible at creative fiction, it’s not bad at spotting missing parentheticals in sections of code. It’s decent at summarizing long transcripts of political speeches from a YouTube transcript so I don’t have to watch the entire thing. It’s not bad at scanning a contract and telling you when you'll get paid.

I am concerned with the ethics of how Generative AI is used. And I am concerned with the ethics of how professors detect and accuse students of AI use. Unfortunately, ethical rules and codes of conduct are often enforced unevenly. And I find there is usually a class of people who are exceptionally competent at getting around the rules. I’d hate to find that AI usage is not enforced uniformly. We already have too many disparities in the educational system.

 

[Voboy]

Brake said a lot more worth reading. These articles and newsletters and talks do a much better job than I do grappling with the issue. But for me, I come back to the fact that Generative AI exists. And teachers have to deal with it, because I don’t see “just say no” working any better with Generative AI than Prohibition did with alcohol.

I suspect we're not too far from a time when, pragmatically, the academic product will get graded regardless of its process: good writing will be rewarded somehow, bad writing will be condemned somehow; some will have entirely been AI-generated, some AI-aided in some sort of hybrid way, others entirely human-generated. I think some teachers and administrators will put a positive spin on all that, others negative. It will help in some academic settings, hurt in others. It will NOT be black-and-white.

I reckon it can fairly be said that there is a talent (maybe "knack" is a better term) for crafting useful prompts, and that the future school space will eventually come to reward that talent. I am aware that the workplace already does, especially in the programming realm (which I know nothing about).

While it’s terrible at creative fiction, it’s not bad at spotting missing parentheticals in sections of code.

Again, I am no programmer; I acknowledge this is probably a benefit of AI, though I've seen Terminator enough times to get a strong sense of unease at the idea of computers programming other computers. Or themselves (lookin' at you, HAL).

I am concerned with the ethics of how Generative AI is used. And I am concerned with the ethics of how professors detect and accuse students of AI use. Unfortunately, ethical rules and codes of conduct are often enforced unevenly. And I find there is usually a class of people who are exceptionally competent at getting around the rules. I’d hate to find that AI usage is not enforced uniformly. We already have too many disparities in the educational system.

If that ship has not already sailed, the moorings are being cast off. I doubt that can be prevented. I am not optimistic that this gets solved, but that just makes it another issue we seem powerless to mitigate. We're in many holes at the moment, and we show no signs of putting down our own shovels.

Again, this isn't black-and-white. And it'll only get greyer, I fear.

 

[onehitwanda]

We need to move away from exams as the default mode of assessment. They are hugely disabling and are an artificial emvironment. When was the last time you had to work in exam conditions as your job? Fear of cheating is often used as a scaremongering way of refusing to offer reasonable adjustments or more inclusive ways of assessment. Currently at the HEI where I work the number of academic integrity cases of suspected AI is very small. Maybe it becomes more prevalent, and yeah finding a truly inclusive form of assessment that is robust and can be trusted is a long and difficult journey, but the answer isn't defaulting to exams.

Counterargument:

1. exams test knowledge in situ, under conditions where cheating is almost impossible due to the presence of invigilators. Particularly in the hard sciences, exams are a de facto method of establishing whether a candidate understands the theory backing their course to a level that warrants the award of a degree.

2. During the nature of my day to day work, I frequently have to react to emerging situations where there is no time to consult an oracle - it is down to me and my team's knowledge and familiarity with both the domain and the underlying systems. These are not "exam conditions" but they are similar.

3. There are other situations - emergency medicine, aviation, lifesaving at sea etc. where the difference in outcome between someone who has deep understanding of a situation and someone who does not can result in death.

My position: exams are a functional, proven and reasonable method to test knowledge for domains where lack of knowledge, incompetence or cheating can have severe consequences.

I am aware that the workplace already does, especially in the programming realm (which I know nothing about).

Which is again great if it's being used by an expert who can spot the edge-cases and failure modes of the generated code, and hilarious (eventually, once the fires are out) if it's used by a cretin who can't.

I'm sure AI has its uses. I work with people who use it. To a degree it's no different to finding an algorithm implementation in Knuth and reusing it. The subtle difference is that Knuth doesn't hallucinate.

 

[Bramblethorn]

This presents three options that I can see:

• I have an absurd and inappropriate sense of humor, and I wrote it.
• AI wrote it and I did some voodoo (swapped words around, moved semicolons) and it turned into an undetectable piece of AI generated text.
• AI detection sucks and you will get false positives and false negatives.

Software-based AI detection does suck, it is known. (Which obviously doesn't exclude the possibility of one of the others also being true.)

I know students and authors exist who get accused of using AI when they haven’t; and that’s rather tragic. It’s a different kind of tragedy, when a student or author gets away with using Generative AI and gets a better grade than some who did it the old fashioned way (laid by the pool and tanned, while handwriting out your essay). We live in the real world and students and others will use AI. If AI is a “tool,” as you point out, it’s also a tool akin to alcohol (in the sense that the Eighteenth Amendment wasn’t successful in prohibiting its use) or firearms (where the Second Amendment runs interference on preventing access).

So, if we can’t prevent its use

Hold on a moment. What you've been arguing up to this point is that "software-based AI detection sucks" (it certainly does) and that depending on its verdict creates unfair outcomes (agreed).

But going from that to "we can't prevent its use" (and eliding the possibility of significantly limiting its use) is a considerable leap. You're tacitly assuming that software detection is the only way this could be done, and that's an assumption that requires justification.

His tagline is that “with adjustment to the way we teach students to think about writing, we can shift the emphasis from product to process.” That’s what I wanted to say, but wasn’t clever enough to say. He argues, “There is currently no reliable way to be sure if a student has used generative AI in writing an essay.” And I agree.

Perhaps more accurate to say there's no reliable automated way to be sure if a student has used generative AI.

Generative AI is new, but the problem of students submitting work they didn't write is not. Rich kids have been paying others to write their essays for a very long time, and there are ways to catch it. A professor who has small enough class sizes to be familiar with individual students can gauge whether the knowledge exhibited in an essay seems consistent with the student's knowledge; a professor who suspects cheating, and has enough contact time to explore that possibility, can talk to the student and probe their understanding of the work they're representing as their own. I had a professor who used to do this for everybody on every assignment.

I do similar things when running job interviews. A candidate can lie their head off about past work experience, and I'm not going to hire a PI to check up on every claim in their CV. But I can ask them questions in the interview: okay, you've used C and Python? Suppose you're starting a new project and deciding which one to use, which factors would you consider?" (Adapt as appropriate to relevant field.)

Having been on both sides of that approach, it's extremely effective IME. But it's labour-intensive, which is where my earlier comments about resourcing come in.

You make a very strong argument about the importance of user experience in tool design, especially as they become more complex. But I think you’ll agree that regardless of design, tools, like pharmaceuticals, often end up getting used in ways that were not foreseen or intended.

Pharmaceuticals are a good analogy here.

At one end of the spectrum, we have formally approved and respectable off-label prescription: the drug isn't being used for its original purpose, but it's gone through studies to establish that it's generally human-safe (or at least acceptable levels of risk) and that it's beneficial for this new intended use.

These systems aren't perfect, but in general they do a reasonably good job. Further, those prescriptions are managed by doctors and pharmacists who are supposed to be looking out for common side-effects and known contraindications ("don't drink alcohol or eat grapefruit while you're taking these") and to talk to patients to ensure they understand what they're taking. Again, an imperfect system but far better than no system.

At the other end of the spectrum, we have black-market Ritalin being used as a study drug and flunitrazepam being used to roofie people, and mail-order "antibiotics" which might actually be antibiotics or might be sugar pills.

If generative AI was being used in a similar way to approved/off-label prescription, with somebody who isn't the vendor helping the potential user evaluate whether it's right for them, and formal criteria for testing their efficacy and safety, I'd be a whole lot happier about it. But as things stand, we're more at the "unregulated mail-order" end of the spectrum.

“There are certainly exceptions, but the pedagogical strategies that the availability of LLMs obliterates are those fundamentally based on the distrust of students. It has not killed writing, but it has killed a certain kind of way of motivating writing. If your students didn't already see the value of writing as a process by which you think, then of course they will be curious about farming the labor out to an LLM. And if writing (or any other task for that matter) is truly only about the product then why not? If the means to the end are unimportant, then why not outsource it?”

And that, IMHO, gets back to resourcing: yes, we should be focussing far more on the process not just the product, but that requires a high-intensity, individualised approach which is at odds with treating education as a for-profit industry.

And I do feel Generative AI has a little more use than just grist for venture capitalists to raise money. While it’s terrible at creative fiction, it’s not bad at spotting missing parentheticals in sections of code. It’s decent at summarizing long transcripts of political speeches from a YouTube transcript so I don’t have to watch the entire thing. It’s not bad at scanning a contract and telling you when you'll get paid.

How good is "not bad"? I've seen AI-generated meeting summaries, and while they might be mostly accurate, the remaining errors can be a huge problem.

This is a question that people selling AI tech should be able to answer. Get a sample of 100 contracts selected according to some appropriate criteria, ask the bot to extract key information, compare to a human expert's findings, publish the results, provide that info to prospective users and let them decide for themselves whether something that gets "when will I get paid?" wrong X% of the time meets their needs.

Is there any AI product on the market that does publish these kinds of accuracy metrics?

I am concerned with the ethics of how Generative AI is used. And I am concerned with the ethics of how professors detect and accuse students of AI use. Unfortunately, ethical rules and codes of conduct are often enforced unevenly. And I find there is usually a class of people who are exceptionally competent at getting around the rules. I’d hate to find that AI usage is not enforced uniformly. We already have too many disparities in the educational system.

No argument here, other than that "competent at getting around the rules" often just translates to "has a rich dad who donated to our building fund" etc.

 

[M_K_Babalon]

Was it though? Dystopias are just more fun to write. Utopia stories get dull quite fast.

The way things are going...

Of every movie I've seen, the utopia is dystopian for somebody.

 

[M_K_Babalon]

I was too lazy to come up with a response on my own, so I had AI write one for me:

AI is just the latest step in a long line of tools—like computers, calculators, and slide rules before it. Okay, maybe I pushed that analogy a bit far, but you get the point. AI is here, it's good, and you won’t be able to ban it. Plus, trying to catch students using it can backfire on those who are actually doing their work.

Teachers should focus on critical thinking, creativity, and ethical use. Teach students how to use AI (because it's not going anywhere), but also design tests that require real understanding. Think handwritten essays, blue books, and so on. AI is just a tool—it's up to us to treat it like one, not a threat. The goal is to create assignments and tests that still challenge students, even with AI in the mix.

Auto Tune is also a tool, and we've seen where that went.

 

[Kadyia]

Auto Tune is also a tool, and we've seen where that went.

Absolutely. But Auto Tune is still easily detectable at least by professionals if not by the untrained. But remember, the TLDR of the original post, e.g., "Not only is GPT really good, too good to detect, but also so easy nearly all college students are using it. It is driving English professors into despair."

We can argue whether "GPT is really good," but the the hair-pulling stress I read in the Author's Hangout about people having stories' flagged as AI and the difficulty faculty are having identifying Generative AI means that AI presents a different problem than Auto Tune.

Both are tools and both are used by tools. But if I use Auto Tune, you and just about everyone else is going to know it. So yes, Generative AI is a tool and Auto Tune is a tool. What's your solution?

I'll be honest. A month ago, I'd never bothered to even play with Generative AI, but I got dragged into it by my work. So I know a bit more now than I did before. I don't think it's going to put actual writers out of business anytime soon. It is grammatical and it doesn't make spelling errors. But as far as I can tell, it doesn't have a deep understanding of human nature. It tends to be repetitive. And it really seems to like adverbs. Lots of them and multi-clause sentences. But so do a lot of writers that I admire.

There is also likely a spectrum of AI use. One person might use it to brainstorm ideas. Another might use it to make sure they don't shift tenses. I couldn't say I've never used AI. When I was writing a novel set in 1920s Patagonia, South America, I downloaded a lot of old books about Patagonia whose copyrights had expired written in the 1870 to 1920 period. Then I uploaded them into Google Notebook LM and would ask it to go through those books and do things like give me a list of flora and fauna in Patagonia.

I wanted information about labor unrest and the Patagonian rebellion and it was very helpful in collating that information from thousands of pages of material. I'd use that information to give my writing an authentic feel for the place and time period. I also use a program to identify things like passive voice and how many adverbs I use and dialogue tags, etc. I don't know, but I think there is some level of AI going on there. While you can't accept everything that it says as gospel, it does help me tighten up my writing and identify areas I want to focus on as I rewrite.

And I actually used some sort of search function on Microsoft to decide exactly what year I wanted to start my novel and when it would end. And while I might complain that I'm writing my own words, isn't what I'm doing essentially like Auto Tune? I'm using it to fine tune my grammar, to research ideas, to collate facts. So I guess I'm like a singer who wants to say I only use Auto Tune to smooth out the rough edges. But is it fair for me to say I don't use AI, just because I write my own text?

 

[Kadyia]

How good is "not bad"? I've seen AI-generated meeting summaries, and while they might be mostly accurate, the remaining errors can be a huge problem.

I obviously am no expert. But from my perspective, it's like having a dumb, but earnest intern who really, really wants to make you happy. If you think you can use AI to do your job for you, you're going to get slammed. When I say it really, really wants to make you happy, I mean if it doesn't know the answer to your question, Generative AI will literally make shit up.

And here's the rub. It's really good at making up shit. I ran into a situation where I fed it the wrong document at my work. And then I read it's summary and it sounded good. But if you have ever worked with an earnest intern who has no idea what they're doing, you'll know trust but verify is the rule of the road. I read somewhere about an attorney or attorneys who turned in a motion generated by AI without verifying it and they got in a load of trouble.

Still, in the document intensive job I do, I've reluctantly had to embrace it, because it does make me more productive. But you have to do the work and you have to know what you're doing, because you can't rely on AI at this point (in my opinion) to get it right every time. The more specific you are, the more helpful it will be. There is a bit of an art form to asking the right question.

I'm not talking about writing fiction mind you. I'll concede that it's hard to identify whether something was written with generative AI. But it's not hard to identify a quality, well-written story and generative AI still fails that test in my opinion.

 

[HordHolm]

I read somewhere about an attorney or attorneys who turned in a motion generated by AI without verifying it and they got in a load of trouble.

Yep, they used AI to generate a court submission and it hallucinated some references, which the judge then checked and found were fantasy, not in the record of federal case law. He then hauled the attorneys over the coals and they admitted they used ChatGPT because they were, basically, being lazy. He fined them several thousand dollars. LegalEagle on youtube has a good (and amusing) summary of the events.