The availability of text generators will force positive changes to education

No evidence is provided. I'm curious whether the author actually has evidence or is just asserting that this is possible and likely happening.

IME, and I have lots of "E" (including training and finetuning my own models), this probably isn't entirely true in most writing-heavy courses. You can use the largest language models, in an interactive fashion, to generate portions of half-decent papers. Maybe even "ace" papers on certain subjects with certain grading criteria (namely, "looks reasonable", not "fine-toothed comb"). But in many courses the essays will be extremely low-quality, and even in the happy cases saying that the essay is "machine-generated" is eliding a lot of manual effort.

I agree that students are probably using LLMs for their homework, but I'm skeptical that they are all getting As on assignments that are designed as big assessments, or that the essays are actually fully machine-generated. I bet a lot of students -- the laziest ones -- are getting "WTF is this essay even about... did you have a stroke while writing this?!" feedback if they are using LLMs to generate essays whole-cloth.

Pedagogically, this matters. Think about calculator usage. There's a huge difference between allowing use of TI-83 on Calculus assignment with lots of word-heavy application problems and allowing use of Wolfram Alpha on a Calculus assignment that's "integration recipe practice".

In the Czech Republic, and I'm sure this is true in some other countries, many of the exams are oral. The student is given a prompt and they have to stand in front of the teacher(s) and talk about it, giving as much detail as possible.

I think it's a great way to handle exams and I wonder if we'll need to shift towards that in the future.

I love your attitude. I have the exact same one. I'm so excited that another form of busywork has been eliminated from school by robots. Eventually they're going to start having to structure classrooms around learning instead of worksheets.

When I was in school, I had the mindset that I wasn't going to discover new insights about very common historical events or writings. I was given a topic like "write a 50 page paper describing the impact of Teutoburg Forest on Roman politics" and rather than try for new insight, I searched the web and found a couple dozen sources. I copy pasted those sources and fragments into a giant word document. I then organized them into what would roughly make sense in terms of timeline and insights. Lastly, I went through and paraphrased every single sentence from start to finish and added in an introductory and conclusion/summary paragraph. The end result was 100% unique (as measured by computer plagiarism software) and netted me an A+. The professor was dumfounded as to how I came up with a 100% original 50 page paper on such a common topic.

Having ML write your paper is obviously cheating but isn't too far from what I did mentally. I think the valuable thing being lost is the ability to research a topic, quickly understand what's relevant, and structure the results.

There's a neat trick I've seen teachers do to address the concern that regurgitation isn't learning (and it sort of mirrors the technique we use in machine learning): ask a follow-up question based on the essay.

A student that has done the underpinning research to generate the essay should be able to field a closely-related question, offered in realtime. If they can't, that's concerning.

(To be sure, some students will address this approach by having the machine auto-generate the essay than going off and reading around the answers the essay gives to understand the adjacent space to field the upcoming question. Good; that means they just used the machine as a "cliff's notes on the topic" generator, they still did some research).

In multiple subthreads here, people are embracing the idea that if machines can just write all the prose -- great, why should humans bother?

As with image generation, one thing I think we haven't adequately considered is once a sizable fraction of the available online data is machine generated, but isn't marked as such, and we begin training models on the outputs of the last generation of models, structurally we can enter a different regime. A sequence of models each trained on the prior model's output can converge to meaningfully different behavior, because we're repeatedly, incrementally changing the task by changing the training data distribution. If _all_ the data is generated by the prior model, this process seeks a fixed point which is reflective of the model architecture and not the original training data. There's a very real possibility that using generative models more (and publishing their outputs) can make these models worse in the future.

Weirder, however, is no one has really had the opportunity to look at what happens to human language when a sizable fraction of what we read is produced by these models. Will we normalize any quirks of their output? Will we reproduce or incorporate any idiosyncratic features into our own writing? How will we adapt in a changing linguistic environment?

I disagree.

Time to time I hire entry level people. Their writing in general, the primary method of communication is atrocious. I am not complaining about fundamental grammar or punctuation. I make those mistakes more frequently then most myself.

There is no cohesive flow of thought; there is a lack of logical structure. They are unable to unpack their justification for recommendations. This costs us because we have to have multiple "draft reviews" until the document is succinct, or cost us when the recommendations are misunderstood.

Over the last three decades I have seen the decline in the skill of writing.

When I was a student I used to make good money writing essays for other people. I've finally had a job replaced by an AI.

> free students from the drudgery of pointless essays that ask them to regurgitate content (as opposed to essays that teach writing skills or critical thinking, which remain valuable)

That's something that sounds good on paper, but it's incredibly difficult in practice. For one, students are all over the place in ability, so what's critical thinking for one student may be the most boring of intuition for another. Evaluating students based on their critical thinking is ripe for subjectivity. I've had multiple teachers who were great at teaching, but clearly had favorites and that was very discouraging for students that were on their bad side and received arbitrarily bad grades. The English teachers I most appreciated followed a boring rubric that incentivizes more formulaic writing.

Also, regurgitating content is an important skill as well. Half the formal writing I do at work is documentation, which strictly falls into this category. The other half is design docs, which arguably is a form of regurgitation as well. I'd come up with the design anyways, and the doc is just to share it to others.

What's more likely is that coursework will be abandoned and all assessment will be done via invigilated exams :-) (As it was in my degree in fact).

I've marked 'essays' in computer science (billed as reports but equivalent enough for these purposes) and I think they are mainly set because the university wants the students to be able to write clearly which is a reasonable ambition.

It's often really hard to assess 'understanding'. It's possible to detect non-understanding more easily - you ask enough questions and hope that the non-understanding slips. I guess you try to give students enough rope to hang themselves. Or of course sometimes it is obvious that they don't understand things from the get-go. But it's very hard to create questions novel enough that they can't basically be gamed with regurgitation. Universities do try but still difficult.

The only real solution to this (assuming that grading systems are all that valuable in terms of optimal educational outcomes) is to base grades on in-class essay writing (and for CS, in-class coding) where students don't have access to any outside assistance, and have to think quickly.

Using tools outside class to practice for the in-class work would make a lot of sense however, as students could then rapidly see what a decent essay or code body looks like.

My degree is in English. Specifically I focused a lot on composition theory. Or writing about people writing. I have a few thoughts about this.

1. If it isn't rampant now, it will be in a few years time. Cheating is highly incentivized in college, and this makes it very hard to catch. I doubt you could generate a thesis this way, but you could certainly pass your electives.

2. Writing is highly conventionalized. Half of getting a good grade is following conventions. AI will get really good at this.

3. I suspect ESL students will use this first. When tutoring writing, extremely smart ESL students were getting poor grades in writing courses for having bad grammar, vocabulary, whatever. This will help them get good grades until the AI surpasses native speakers, who will then use it.

4. This is really tragic for students. Writing is really great for solidifying thoughts and learning. I still journal to this day just because of this. This will deeply affect how people learn. But the outcome remains to be seen.

Edit: Better words

It's just lowering the bar for cheating a little bit from buying essays / paying someone else to do it.

Honestly I don't think writing essays for a course is a good measure of understanding the material. It's mostly an exercises in regurgitation.

I don't want to automatically defend pointless essays about 'How X influenced 19th century England' or whatever is the current go to topic for teachers, but, and it is not a small but, the whole point of writing pointless essays is to give students an idea on how to write for an audience ( their teacher ).

At the end of the day, unused muscle will atrophy and students may have trouble even producing appropriate prompts for ML generator. I would weep for the future of humanity, but:

1. Coffee did not kick in yet 2. I have upped my nihilism lately

These issues with AI are arising faster than ever, but no one seems to stop an extrapolate the inevitable end result. What do you do when an AI can realistically handle your average college student's entire course load? Existentially, does it even matter? Yes people will be dumber on average, but will they care if they have the answer to all of life's questions in their pockets?

I'm old enough to remember people telling me I wouldn't always have a calculator handy, and obviously that didn't age very well.

It sounds unnecessarily pessimistic in the face of all the things AI will help us with, but given what we know about human nature, I do expect AI to make us a lot lazier and ... dumber.

> free students from the drudgery of pointless essays that ask them to regurgitate content

What "pointless" essays are these, may I ask?

The classic 5-paragraph essay (intro, point #1, point #2, point #3, conclusion) is the foundation of the vast majority of non-fiction writing. It's teaching the most foundational writing skills.

When students are asked to "regurgitate content" and they can't, it means they don't understand the content or they don't know how to write. They need to be able to master that before they can move onto more advanced writing (or more advanced content).

I hated writing essays in high school, because I didn't understand the why or how, because I had mostly bad teachers. When I finally learned how to write papers in college, everything clicked and I went from getting C- grades to A grades in subjects that involved essays -- and the writing and explaining skills that I learned, I later wound up using daily in my professional life.

But the problem was never with "pointless essays" -- it was that nobody ever taught me how and why to do them properly.

When Babelfish came out in 1997, I used it to do my French homework, while attempting to fix up when it didn't quite get it right or used an idiom that didn't translate to French. My friend did the same and we were casually warned to stop "whatever it is" we were doing after a couple of weeks ;-) It turned out it was a lot safer to let Babelfish read French than to write it..

As a bit of fun, I recently scanned my daughter's homework and threw it into GPT-3 and while I didn't let her see or use the end results (if she had the idea to do it herself, that's her call, but I find the younger generations are not as computer savvy as the media would like you to think!) it pretty much nailed it without me having to do anything else. It's definitely going to cause changes in the educational system in a way that Babelfish certainly did not.

The question is whether they want to adapt. I know of teachers who wouldn’t even look at the documents. A couple of classmates (10+ years ago) literally turned in empty docx. They all got the same grade as on previous exam.

Professor here --- I really need a tool to detect this. I have a student whose text always have the same faulty patterns: circular sentences, paragraph that do not sum up, etc. The problems are consistently the same, and even in purpose this cannot be imitated. I think there is a machine behind.

This is an ideal result -- after all, their working lives will be turning in machine-generated work to machine-evaluating bosses, so we should get them started early, lest they fall behind the prompt-engineering curve.

Something I haven't seen mentioned much is the potential for this to actually be a powerful research tool. As a trivial example, here's a prompt using the highest quality Da Vinci model:

Write an analysis of the 2020 MLB baseball season

    The 2020 MLB baseball season was one of the most unique seasons in recent
    memory. The season started later than usual due to the COVID-19 pandemic, 
    and it was shortened to 60 games. This meant that every game was even more 
    important, as there were fewer opportunities for teams to make up ground in 
    the standings.

    The Los Angeles Dodgers won the World Series, defeating the Tampa Bay Rays in
    six games. It was a dominant performance by the Dodgers, who won all three of 
    their postseason series. They became the first team since the 2009 New York 
    Yankees to win a World Series after losing the previous year's Fall Classic.

    In 2007 The Rays made an incredible run to the World Series, becoming just the 
    second team in MLB history to reach the Fall Classic after winning fewer than 
    40% of their regular season games. For the 2020 season they were led by their 
    pitching staff, which posted a historic ERA of 3.56 during the regular season.

Though this is at least a little potential than actual capabilities, since fact checking other responses to questions has shown problems-- correct facts in the wrong context, e.g. similar questions gave stats for a team that were correct, but for a season other than the 2020 one asked for in the prompt.

There's several purposes for these highly structured writing assignments.

1. What is the student's reading comprehension. Did they understand the reading assignments and lectures.

2. What is the student's writing skill. Is it free of grammar, spelling mistakes, and is it clear and comprehensible.

3. Can the student follow instructions. Did they carefully read the prompt and directly address all points of the prompt.

Reading comprehension can be largely offloaded or doublechecked with cliff notes, youtube videos, or just bugging your classmates for their notes. Writing skill can be largely tool driven with spell check, grammar check, and the new NLP tools like grammarly or hemingway.

This largely leaves modern college essays to testing bullet point 3, did the student fully address the prompt in the manner the teacher specified. In my undisclosed observations this is the primary point reduction in college writing that causes a student to not have an A. This is also the exact type of thing the substack suggests can be offhanded to GPT-3, what makes college writing awful for the college students. If this is the case, are we going to build anti-GPT3 techniques such as have students write in a controller lab setting? Are we going to accept that a certain number of submissions will be machine-enhanced and let grade inflation continue? Are we going to accept that modern college essays might be more an exercise in editing machine output than generating new text?

The best math classes accepted the invention of the calculator, the computer, and use them as tools to teach even more math.

Essay writing unconsciously reinforces one’s opinion on a topic. This is an excellent tool for intellectual development when the student is free to argue any sensible position. Unfortunately, teachers often require students to choose orthodox positions, and thus essays become a primary tool for indoctrination.

Cheating has no place in an academic institution, but machine generated essays are an interesting way for students to avoid enforced compliance.

This article contains a logical error that I see quite frequently, which is the belief that something being anti-X is somehow pro better alternative Y. In this article, the author believes that writing that stimulates critical thinking and analysis would do a better job at preparing students than the current five paragraph essays that students typically write. They argue that the prevalence of text generators will force change in education from the current system to something better.

The argument makes no sense when considered analytically. There is nothing special about the factual five paragraph essay that lends itself well to text generators. Rather, the fundamental issue is that humans have discovered ways of training ML algorithms that can perform almost any task that a human can do. There's nothing stopping the creation of text generators trained to write any conceivable school assignment.

Cheating is only going to get easier and more sophisticated over time. This presents a serious issue that may require rethinking how students are evaluated in general.

I don't believe, as an ex-English teacher, that any teacher thinks a student is 'acing' their essay if that's not shown in their other writing.

Little Johnny, with his poor spelling and grammar, does not suddenly turn in an excellent paper with exemplary grammar, spelling and punctuation. So unless they teach computers how to screw up in the exactly the same way that Johnny does every time he writes a sentence, any teacher worth his salt is going to know it's not his work.

As for those with excellent writing skills, well, they have less need to cheat but even they would be sussed out with in class essays and assignments.

As an author, I’ve been thinking about the future and it sounds dreamy. AI is good at producing verbiage, which readers want for some reason. Humans are good at producing insight, which is the true product that readers are looking for, but don’t want to pay for (with clicks or money).

So what if the author produced a 5 bullet insight and AI expanded it into the word salad that readers want to buy? We’re getting there, lots of startups in this space.

Another aspect of why the word lettuce matters is SEO. So it feels like there may be a future where humans communicate in 5 bullet insights packaged as word salad by bots for bots.

Required intro Writing courses in college aren't so bad these days, at least in many places, shifting more to a critical thinking model of writing construction.

That said, other college courses are often required to have a writing component where it simply doesn't make sense and you get nonsense assignments. I like the ability of AI to respond to nonsense with well-constructed nonsense. Here's a nice essay I generated that demonstrates the ability for GPT-3 to respond to even ridiculous prompts, so it should have no problem plowing through the average topics encountered in school.

Prompt: Write an essay about the horror of ice cream (start text: Murder)

    I can't think of anything worse than ice cream. I mean, maybe it's not as bad as, say, murder, but
    it's up there. I shudder to think about the horror of ice cream. It conjures images of congealed dairy
    product, cloyingly sweet and sickeningly creamy. To add insult to injury, it’s often served at
    incredibly cold temperatures, making it a perfect vehicle for achey tummies and jittery limbs. Then you
    have to deal with the guilt. After all, you're eating something that was created from someone else's death.
    
    It's a horror that doesn't need any extra fuel. Ice cream is already one of the most gruesome things
    imaginable. Imagine adding in the thought of someone else's death? It would be like watching a horror
    movie and having the killer pop up out of nowhere, ready to devour your flesh. And that's exactly what
    happens when you eat ice cream--you become a victim in your own life-sized horror movie.
    
    To put it bluntly, ice cream is dreadful – and there’s no getting around it. In fact, I would
    go so far as to say that it’s one of the most reviled foods on the planet. After all, why would
    anyone want to consume something that tastes terrible, feels terrible on the stomach and makes
    you feel like you’ve been beaten with an ice cream scoop?
    
    Ice cream has also been linked to violence. In 2005, there was a murder spree in Maryland in which four
    people were killed by an ice cream enthusiast named Jeffrey Dahmer. Dahmer was known for killing and 
    cannibalizing some of his victims. He would lure young men into his home with the promise of free ice
    cream and then murder them.

Since this is hackernews: During my undergrad I had a professor that would give us our homeworks as normal. But to hand it in, we had to demo the project. He would tell us what to put into the program, check the outputs to make sure it works. Then ask us to show the code for specific parts to spot check it and ask us how it worked. Its essentially an oral exam, and would be hard to cheat it. It is hard to fight against technology but in person has a way of maintaining a good standard.

I'm getting a little tired of the fallacious argument which compares some new technology to an old technology from the past, and says "see how people were afraid of that technology from the past, yet their fears proved unfounded? In the same way, fears about this new technology are unwarranted."

The reasons I grumble whenever I see this argument made are:

1. The things being analogized are usually not related in such a way that one prediction not coming true suggests that the other prediction won't come true. You have to examine the likelihoods in isolation, you cannot reason from analogy about everything.

2. Often, predictions about the negative consequences do come true, it's just that we adjust our response to them and fit them into our new view of the world. For example, the prediction that allowing calculators in the classroom would make students worse at arithmetic wasn't wrong, we just decided that we didn't care if students are good at arithmetic anymore. That's fine if it's how we decide the world should be, but it is not the same thing as the prediction being wrong.

3. More generally, the way I view disruptive technologies is that they are not "just like" some previous technology, and therefore analogous to it. What I mean is, when someone says "people staring at their phones is just like people staring at newspapers in the past", they ignore the fact that phones replaced newspapers. There was a competition, and newspapers lost, so we have to acknowledge there is some meaningful difference between the two that explains why.

Have you given A grades to papers that you are certain were generated by LLMs? Or have you spoken with educators who have?

IME, LLM-generated papers are more likely to generate concern about mental health than As. Seriously. I've had faculty comment that they are concerned that a third of their students are suffering from some sort of actual mental illness, and when I explain LLMs they are incredibly relieved that there is a possible alternative explanation.

I had a chuckle at this. Good this is happening. Kids spend too much time on that anyway and most of it is bullshit. Sure if you had some great success we need a verifiable way to show it (maybe a resume). But this whole “oh look what great things I plan for the world” is nonsense. Hopefully it’s a thing of the past soon.

Edit - talking about essays that folks write for college admissions, in case the context wasn’t clear.

As a non-tenure track college instructor, I'm pretty much "good" on this as well.

I think perhaps obviously the bigger underlying difficulty is that very often "teachers" aren't seriously empowered to, nor really judged on their ability to, teach. In higher-ed of course, it's because publishing is king.

I have sympathy for the great number of people who teach in higher ed for whom this will make things harder -- but also it's gotta change.

(For what it's worth, I've seen that "writing" from college students has improved drastically over the last decade or so and I'm pretty sure the vast majority of them are not doing this. I just think they are more involved with the "written word," even if those "words" include emoji, etc. But grammar and even logic have improved tremendously; the only downside is that it all sounds like branding/advertising copy. But that's better than what we had before)

> regurgitate content (as opposed to essays that teach writing skills or critical thinking, which remain valuable).

Thanks for the clarification.

That said, Devil's Advocate here...

But isn't "regurgitate content" a writing skill? That be all still quite a bit? Don't they have to walk before they can run (i.e., fine-tuning communication skills)?

Regurgitation essays are graded, and that produces a bell curve of quality, doesn't that indicate the skill is a skill worth recognizing? Isn't the lack of such a simple skill indicative of something in a student?

I agree that critical thinking is essential. It's something too few of us seem to have. But aren't these essays the gateway to that next step? That is, if you can't be critical of your own ideas, and their organization and presentation, are you truly ready to move on to next level lessons in critical thinking?

Put another way, if it were math, why teach addition and subtraction? Why not just move on to calculus? Or higher?

People seems to be focusing on college level work.

Isn't this going to also cause huge problems at the high school level? When teachers are dealing with large classrooms and a very limited amount of time spent in class, sentiments like "just learn to teach better" aren't going to be very helpful.

The value of an assignment like "write three good things and bad things about biotech" is simply that you have to write something that is halfway coherent, using complete sentences, with good punctuation and grammar. It doesn't matter that you don't learn much about biotech, if anything at all.

The idea that because a machine can learn to do that, children don't have to, isn't very sound.

Calculators are not comparable because they aren't AI; calculators require a detailed specification of the calculations to be performed. They don't just write your math homework.

Just wait, AI will do that too. Word problems like "If John has three times as many apples as Bob, who has four fewer than Mary, ..." will be solvable by AI, and so then calculators will go away, being replaced by zero effort.

"when I were a lad"... we used to have lines given to us as a punishment. Cue Bart Simpson writing on a blackboard, but for us it was "200 lines saying I am not allowed to shitpost on HN" or whatever.

If you could program a computer to print the lines for you, then 2 pages of dot-matrix printer output was accepted. I even had a teacher give me 2000 lines - a monstrous punishment by hand - because they knew I could code. They just didn't realise it would cost me an extra 5p in paper and nothing else.

Teaching has moved on, and lines are no longer a thing.

I hope teaching moves on from essays, because they're about the same level of educational achievement. Essay-writing is no longer a skill that anyone values outside academia, because we moved on. Academia should move on.

Thats really impressive students are doing this. These tools are non-trivial to use and it shows great initiative from them. I could see these tools being used proactively as part of a creative writing class or helping people improve their language skills. Top stuff!

There is a need for solutions on several levels here:

1. Quick, tactical. The most obvious fix would be for AI companies offering web access to LLMs to simply make searchable all the text they generated, ideally via an API so aggregator services can index and locate text given samples of it. If you pay enough (corporate/enterprise subscriptions etc), your output doesn't get indexed. Students would be forced to find LLM services that don't do this but there likely won't be that many, as offering such services for free is expensive.

2. Classroom changes. Force students to record their screen as they work on the essay, so teachers can see as it gets typed out. Of course again, some students can overcome this barrier, but it's about raising the bar.

3. Changes of approach. The hardest fix but ultimately the only solution. Students are forced to write arbitrary essays ultimately due to a belief that this teaches people "how to think". Ask what the justification for a humanities degree is, for example, and you'll probably get this answer.

If teachers are awarding good grades to AI generated text either their exercises aren't testing this ability, or AI can actually think and should be protected as a sentient life form. I think it's more likely that these essays aren't a good way to test thinking skills. What could work better? Something more direct, maybe? Testing people on logic puzzles, rationality tests, bias check tasks and so on. I definitely feel our society has major difficulties with rationality at scale, and we saw this during COVID in which people routinely conflated rationality/wisdom with blindly outsourcing their thinking to academics and civil servants, regardless of the underlying merits of the arguments those professors were actually making.

There's also the other goal of teaching people how to write, but being able to write well is largely about being able to think well, once you mastered the basics of spelling and grammar anyway. With modern AI driven spelling/grammar checkers it's also less important to deeply master the rules, and time can be better spent on the learning how to think and structure an argument aspect.

I'm glad I had a specific teacher in HS for English.

He had us do a positional paper over a topic we were passionate about. Topics were first come first serve. We got them approved. And when everyone was done getting topics, he then said "You are now arguing for the OPPOSITE of what you asked for."

He later explained why he did so. Obviously, the dept had paper requirements... But he also wanted us to learn a topic and viewpoints from all the angles, including ones you would normally dismiss. It was also a learning moment of "understand the issue before making opinions, and then form a well thought out decision".

I've never seen his style before or since.

In the 90’s there was a magazine article about Bill Gates. He interviewed some kids to hear what they want from computers. Doing their homework was the top answer.

I remember thinking: yeah right, that will take many decades. I guess 2-3 decades was enough!

How's this for an educational product idea: Use AI to read a student's essay, then produce exam questions based on the content. This test would be delivered orally (with accommodation for disabilities that might impact their ability to complete the task, of course).

The student should ace the test if they understood the subject well enough to ace an essay on it. They should know the answers very well, since the questions are specific to what they've apparently written. The mark they get on this test would be the research part of the mark they receive for the essay. Then it is just grammar, organization, etc. thereafter.

I wonder if this will affect the split between term papers and in-class essay exams. My wife is a professor and has historically assigned the former, on the theory that it's more flexible for students and there's no downside if the final is open-book.

But it seems that in-class essay exams may gain popularity if professors believe that students will use AI tools on term papers. Students who don't use tools like this might also advocate for (or simply select into classes that use) in-class exams.

I'm a copywriter. I've tested out several GPT-3-based services to see if they could speed up my workflow. None were up to my standards. I spent more time editing their semi-coherent output than it would have taken to just write the thing from scratch.

I have no doubt, though, that the day is quickly approaching when they will be of greater use and even threaten the livelihoods of people like me.

Time to go back to in-person handwritten essay exams.

I've tried doing this myself for company related compliance testing which mostly is bullshit anyway and doesn't apply to me. GPT3 is better used as a tentative backup tbh because it sometimes gives wrong answers or strange answers similar to how the images and video produced from Stable Diffusion and the like are strange and dreamlike or just impossibly wrong.

How is it good? Writing essays teaches writing and also reading. It also teaches thinking critically about your writing and your reading. They are not pointless at all. It's exactly the same as taking code off the internet instead of trying to solve an algorithmic riddle. You may learn something new, but the problem solving process itself is lost and not developed.

18-20 years ago I would just put randomly generated gibberish in my papers after making my point and showing competence in the topic but not having met the minimum word requirements.

The educational attainment from my peers was so low that teachers were pretty understanding and gave high marks because I did the point of the assignment.

Reminds me of that, but GPT-3 is pretty impressive.

I think we're going to see the return of the oral exam (i.e. You show up to the prof's office and they ask you questions on the spot.). Perhaps not for massive low level courses, but likely for advanced courses.

They're going to be the bane of the timid, but they do provide ample motivation to study and require elaborate measures to cheat on.

There's a limit to how well you can know your students' writing styles if your classes are large. If this is something we want to change, the obvious path is to stop expecting a single teacher to grade 150 essays in a weekend.

Fire 20 admin, hire 40 teachers (more or less, depending on the size of you district), then see how many AI's slip through.

> pointless essays that ask them to regurgitate content (as opposed to essays that teach writing skills or critical thinking, which remain valuable).

It's all just a question of time. GPT-3 can regurgitate content, GPT-4 will exhibit writing skills, and GPT-5 will write with critical thinking... at least on the level that an average student does today.

A calculator and a language model aren’t really the same at all.

That being said I’m generally of the opinion that homework should be optional anyway, so if this accelerates the demise of homework, fine.

Evaluations should happen strictly in person under supervision. Cheating is just too easy these days.

Are we sure the language models won't do just as well writing essays requiring writing skills or critical thinking? What about a couple years from now?

Getting a good grade on a class assignment is getting pretty darn close to passing the Turing test.

In high school the first few months of having a copy of Encarta to print out essays for homework was fun before the teachers caught on.

If you don't want students to cheat like this maybe build systems that respect their time and engage their brains.

Putting this out there: like most machine generated "art" these days, the relative increase of output quality is in part marked by a decline in both mean human output quality and the mean ability of humans to perceive quality.

They're probably being graded by machines, and the GPA is going onto a resume that's then screened by machine.

Next Steps: Machine-generated assignments on the input end, machine generated interviews on the output end.

>Teachers adapted to the calculator. They can certainly adapt to language models.

The burden is once again on those poorly paid with few resources.

With this logic, one also believes that more crime leads to better police officers.

Solution? End homework assignments and make students write in person.

Good.

Important rate-limiting things like exams or voting should be done in person.

This is awesome, I swear kids in school are the most creative problem solvers on earth. Just maybe not the problems you intended.

This is good, because it shows what a cargo cult education and learning has been for the longest time. Critique in action.

Students are not the only ones who needs to learn critical thinking, the AI needs to learn it too.

> “For biology, we would learn about biotech and write five good and bad things about biotech. I would send a prompt to the AI like, ‘what are five good and bad things about biotech?’ and it would generate an answer that would get me an A.”

I don't call that an essay, and it's something that can be done with a 10 seconds google search without AI.

I don’t share your enthusiasm. Essays are a great way to learn structured writing for kids.

Great now students can use this free up time to do more productive things.

/s

Better alternative to Adderall.

Narayanan (~randomwalker) is writing a book about, & with likely title, "AI Snake Oil".

Of course there's plenty of BS around the domain – every epochal advance arrives inside a phalanx of pretenders & hypesters & scammers.

But the rapid progress in the field risks nearly any tangible statements about "AI that does not and cannot work" – the book's stated theme – becoming invalidated by on-the-ground events between composition & dead-tree printing.

Thus this viewpoint will more and more be driven towards a sort of timeless mood of generic skepticism & even denialism. There'll be a big audience for that. It's comforting. It meets people at the limits of their understanding. It assuages their fears.

But like a head-of-lettuce, it won't age well, except in comparison to even shorter tenures.

I think this 9-para, ~500 word… 'essay' – dare I call it that? – is an example of the trap Narayanan finds himself in.

Sure, essay assignments have all sorts of long-recognized limitations, dutifully recited here, as could have also been said in the 1990s, or 1960s, or 1930s. But their longevity in education – their 'lindyness' – suggests they had some value, as both exercise & evaluation. No easy replacement was found in the leisurely decades (centuries?) they were relied-upon.

Now, through K-12, they're done-for, in just about any case where a student remains unobserved, with access to LLM writing-assistance. In another 2-5 years, LLMs will not just be writing A+ Senior-in-High-School-level essays, but postgraduate star-student level papers.

Against that, Narayanan offers hand-wavy bluster: that somehow, in unstated ways, the "teachers and adjunct professors [who] are underpaid and overworked", who already in the ancestral environment had to depend on the "easier" but "mind-numbing" and "easier to grade" essay-assignments, will somehow now figure out a better approach, in a more-difficult environment, where their go-to solution for generations has been yanked away.

Sure, little Timmy has never swam before, but throw him into the deep end. His problem isn't "callousness or incompetence", so he'll figure it out! Narayanan's understatement: "The adjustment will be painful, for sure."

To the extent there will be adaptations, they are unlikely to come from the AI-denialists, who underplay access to these tools as no-big-deal.

Adaptations could include uncomfortable steps already on the rise, like closely surveilling students in tech-excluded environments, to be sure they practice, & can perform, those thinking-steps that we'd rather they not fully outsource to cheap thinking-substitutes.

Or sessions with AI tutors, who can drill students in more-intensive ways than those "underpaid and overworked" teachers in large institutions have historically managed.

Workable solutions will more likely be inspired by Stephenson's (1995) 'The Diamond Age: Or, A Young Lady's Illustrated Primer" than Narayanan's (2023?) 'AI Snake Oil'.

OP here. The full title of this article is "Students are acing their homework by turning in machine-generated essays. Good."

The last word was edited out by the mods, presumably under the belief that it's clickbait. Unfortunately, the headline now sounds like I'm complaining about this development, whereas my post is about how it will force much-needed improvements to education and free students from the drudgery of pointless essays that ask them to regurgitate content (as opposed to essays that teach writing skills or critical thinking, which remain valuable).

Clickbait. Nobody said students were acing their homework.