In "Brave New World" Mustafa Mond is the only one left who understands what has disappeared from the human race. I wonder what will happen to the students of today and tomorrow whose intellectual curiosity impels them to learn how to read, think and write in a world where everyone else is cheating.
During Covid I was amazed to discover that my college and graduate education (read, read, read, write, talk) had equipped me to explore all sorts of interesting ideas, no matter how complex the writers. Somehow, fifty years later, I have the ability to synthesize intriguing notions on all sorts of topics into my own original, increasingly coherent body of thought.
As folks here know, you can only understand the "Life of the Mind" if you are living it. It's not only the give and take of ideas, it's also marveling in the workings of your own unique intelligence, figuring out ways to feed it, and developing your own creativity.
Unfortunately such realization doesn't bestow the power to halt this mindless destruction of our species' intellectual and creative capacity. For those obsessively delivering these LLMs, the vita contemplativa isn't worth doing the work for even a few undergraduate humanities courses.
These are freshman courses- which even at Harvard are at best one step up from a high school survey course. Before we administer the last rites to undergraduate humanities and social sciences, maybe we should wait and see how Chat does in a senior seminar where close reasoning and the use of evidence are the main things your professors are going to be looking for.
Great piece Maya. I’d love you to join a Summit we are doing in Nashville with leaders across prek to gray education. This topic is so key. I’m at dquazzo@gsv.com if you are interested. We’d love you on a panel with TurnitIn leadership etal. Best, deborah quazzo
I'd consider myself an expert ChatGPT user. I use it hours every day and have used it (or its predecessors) for years now.
First off, great essay, Maya!
Secondly, I like to point out that ChatGPT is not something you can just ask for an essay and get back super high quality work. You have to know your subject matter enough to have a long back-and-forth conversation with it iterating the output.
One fun and effective tip is to ask it to take on the persona of some expert in whatever field.
It's a lot of work to get the best quality out of these things. Hours of iteration to create a good quality essay.
This makes me pretty happy in retrospect that I went to U of Chicago. The grade deflation was miserable in the moment. But in retrospect I'm glad I went through that - forced me to produce really good work and finally start getting my crap in order.
I read the Truman essay GPT generated. It was painful. The analysis on the atom bomb especially was unbelievably shallow and seems like something out of a middle school paper. Throughout the entire paper the overuse of $10 words is insane. And the general prose style feels like the sort of writing I'd do when I was desperate to reach a word count. This paper strikes me as something that would get a C or maybe a C+.
So yeah, I knew Harvard grade inflation was a problem, but holy crap.
The solution, as with all problems caused by AI, is Butlerian Jihad. We the people did not ask these tech freaks to try to make humanity obsolete. We must ban this crap before it gets even further out of control.
This is a great essay, Maya! I teach political science at the college level, and this helps me think about how to best handle generative AI. I really appreciate a student’s perspective on these issues.
Yeah, I'm thinking blue book exams should make a major comeback. Maybe in addition to a mid-term, you have a draft-writing session for the term paper, and then some discussion sections become peer-review sessions so that people can discuss and come up with revisions in a space that is watched by people instead of doing major edits at home.
As an aside, of course grade inflation is worse at the Ivies than at UC Berkeley. It has always been true that private schools are easier to stay in than the prestigious public schools- private schools' motivation is to keep their students in school once they've accepted them, because they're either providing money or making the school look good somehow. The UC's primary motivation is to serve the students of California, and so if one flunks out, that frees up a spot for another qualified person to take their place.
> AI could automate the vast majority of the legal writing grunt work
There was a semi popular reddit thread where an AI enthusiast tried to use ChatGpt to 'help' their girlfriends legal case and her lawyer refused to look at any of it. This is kinda understandable. For a lawyer to sign off on something another person or AI generated, they would still need to read, follow citations and understand precedent to do due diligence. You could maybe have some form of insurance to hedge against mistakes? Maybe there is some rote form filling that can be done more efficiently? Still for AI specifically to help it comes with some guessing and risk. So I think it is still pretty unclear whether/how AI impacts law.
Nitpicky point that doesn't detract from the main thesis though.
This was a great article, I work in a counter-abuse role, and I wanted to mention that the "Google Docs Edit History" proposal is actually much better than it may look on the surface. This is because it's very hard for anyone to generate 'realistic' looking metadata, especially when 'realism' is going to be determined by some other AI system.
The team building the system that looks at the metadata to determine if it's fake has a good idea of what 'normal' edit history looks like because they'll have access to lots and lots of real examples. An individual trying to fake this isn't going to have access to a huge library of 'real' edit histories, so won't be able to fake it.
Exactly - if someone just takes a LLM-written essay and then hand-types it into Google docs, that’ll look like someone who just sat down and wrote an essay from start to finish in fifteen minutes, which is nothing like how a real person writes something.
I think ChatGPT may be the death of "bulk" writing. Maybe my view is skewed because I've been in the scientific realm for so long that, for most things, ChatGPT would provide an overall better written product. Maybe it's not such a big deal if people (outside of literary authors) are worse at long-form formal writing?
We already do most of our writing in short-form anyway (like responding to a blog post, which is admittedly longer form). As a result, my writing contains a higher density of ideas and less padding (even if those ideas are often garbage). Maybe the next "this meeting could have been an email" will be "this book could have been a blog post"?
I'm sorry, but you've failed to ask the most important question here:
What were your professors trying to teach you with those assignments? Were they graded on effort or on quality?
There are pedagogically-valid reasons to assign work that will be graded on effort, while never explicitly admitting as much to students.
For example: I'm a math TA, and I spent a couple hours Sunday grading student work. For some of the questions assigned, I (and the professor) didn't care if students got the right answer. We had asked the students to compute certain tables "by hand", because we wanted to force the students to experience every step of the algorithm. When I graded, I didn't check if they got every number in the table right*; if they'd filled in the table with anything other than random digits, they'd learned what we wanted to teach. (Also: it was a 1pt problem.)
In undergrad, I took a number of humanities classes, and many of them had weekly or biweekly short-essay submissions. None of them actually cared what we wrote; they just wanted to make sure that we were keeping up with the reading and thinking about what we read. A student used Chat-GPT to complete these submissions would be little different from a math student who used a calculator to cheat on their non-calculator homework: only cheating themselves.
Obviously, grading on effort is inappropriate for major exams. But were these major exams, or weekly check-ins? And if the TAs who graded Chat-GPT's work were checking whether it was keeping up with the reading, were they wrong to give it an "A"? After all, Chat-GPT is quite good at reading and interpreting human texts.
* If you're Prof. Hatice's Algebra course and reading this, then I make no guarantees to continue grading with this laxity!
Did any of the graders hazard a guess as to whether what they had written was produced by a bot?
You told them that there was some chance it was written by you, some chance it was written by the bot. I would have thought that many of them would mention this in their comments, and I am puzzled not to hear about it.
Esp. the writing course, which was scathing (and accurate) about the shortcomings of the submission.
So, this makes me wonder: did you instruct the graders *not* to mention whether they thought it was written by a bot? Or perhaps instruct them to give the comments that they would give if they believed it were written by a human?
Humans still play chess despite being vastly inferior to computers. In fact using computers for training has enhanced the quality of play through new ideas and training. Similarly LLMs will not replace writing and will often enhance it.
I'll go out on a limb and argue that there will be more great writing and more efficient bad but necessary writing (manuals, summaries, technical docs, etc). Second part is fairly straightforward, no one's career goal as an engineer or data scientist is to write technical documents for a product, however these roles require technical product knowledge. If I can train an LMM to read and document code to a template, then I edit, my work is now more interesting and I can spend saved time improving my product. There is no shortage of productive work to do and no shortage of necessary and time intensive writing.
Writing is a time and focus intensive endeavor. Very few people have the temperament to be writers. Many people have interesting ideas and literal sensibilities. If you allow for LLM augmented writing you expand the population of writers and create more literature. A greater more diverse pool of literature creates more opportunities for genius. Literature will become inherently better just as chess is better, benefiting authors and readers. Chess players are not making less money because computers play better.
For younger grades, especially grades 4-8, I'm worried. There is absolutely nothing stopping them from doing all their take home written work via AI. Those kids are already more tech savvy than their parents, and can find a way to access anything that's available. The time, resources, and agility required for elementary and middle schools to prevent this themselves seems like it's not there and parents broadly do not have the ability either (some will, but as a whole, nope). And these are crucial ages for writing skills that are life skills, not academic ones. I fear we'll have many students that come into that age range proficient in reading and writing and then slowly drift downward over time before it's obvious they've been using AI and need remedial help.
I think the question about higher Ed is much less important than younger grades. For higher Ed, so much depends on whether AI can break into the realm of consistently great analysis rather than the mere good writing it produces today (both possibilities seem plausible to me). If it can do that, we're talking about a major breakthrough and the ability to measure college students grades is small potatoes in comparison. If it can't break through, colleges will tweak assessment methods or raise standards and it'll be fine.
Another path is they become a tool just like spellcheck. I can't spell but since I always type it's not a huge set-back. I could also see these tools becoming the norm for research, starting structure, proof-reading, grammar puncher-uppers. They would level-up the entire class.
If writing skill is affected by AI in the same way that spellcheck has affected spelling, I think we will be worse off. Writing is much more generalizable than spelling and is a conduit for novel ideas and analysis in a way that spelling is not.
I don't see how that would be the case. Professional writers have editors. Editors create value. A LLM powered "personal editor" would provide some approximate value.
I think we're talking about different things here. For adults, I see powerful and useful writing AI as a clear good. It's the impact on the formation of writing skills in children that I think would be bad.
Could be different things. I guess just don't see a problem if children are learning how to write in tandem with such an editor-like tool. I think it would have a rising-tides-lifts-all-boats effect. I do agree if children are just copy and pasting LLM outputs that's a problem. I think teachers will be able to solve for that though.
I am aware of that. However, students would have to upload the relevant texts first to do that and most of them aren't that clever. They just pump in generic questions and provide generic answers which can then be critiqued for their lack of specificity.
These programs also require students to pay attention to the examples/evidence provided, and many don't do that either. For example, one student was "writing" about juvenile justice reform in the United States, and she ended up discussing an Indian law about juvenile justice at length. If she'd only clicked on the article she "cited," she would have seen that it had nothing to do with reform in the United States.
In "Brave New World" Mustafa Mond is the only one left who understands what has disappeared from the human race. I wonder what will happen to the students of today and tomorrow whose intellectual curiosity impels them to learn how to read, think and write in a world where everyone else is cheating.
During Covid I was amazed to discover that my college and graduate education (read, read, read, write, talk) had equipped me to explore all sorts of interesting ideas, no matter how complex the writers. Somehow, fifty years later, I have the ability to synthesize intriguing notions on all sorts of topics into my own original, increasingly coherent body of thought.
As folks here know, you can only understand the "Life of the Mind" if you are living it. It's not only the give and take of ideas, it's also marveling in the workings of your own unique intelligence, figuring out ways to feed it, and developing your own creativity.
Unfortunately such realization doesn't bestow the power to halt this mindless destruction of our species' intellectual and creative capacity. For those obsessively delivering these LLMs, the vita contemplativa isn't worth doing the work for even a few undergraduate humanities courses.
These are freshman courses- which even at Harvard are at best one step up from a high school survey course. Before we administer the last rites to undergraduate humanities and social sciences, maybe we should wait and see how Chat does in a senior seminar where close reasoning and the use of evidence are the main things your professors are going to be looking for.
Great piece Maya. I’d love you to join a Summit we are doing in Nashville with leaders across prek to gray education. This topic is so key. I’m at dquazzo@gsv.com if you are interested. We’d love you on a panel with TurnitIn leadership etal. Best, deborah quazzo
I'd consider myself an expert ChatGPT user. I use it hours every day and have used it (or its predecessors) for years now.
First off, great essay, Maya!
Secondly, I like to point out that ChatGPT is not something you can just ask for an essay and get back super high quality work. You have to know your subject matter enough to have a long back-and-forth conversation with it iterating the output.
One fun and effective tip is to ask it to take on the persona of some expert in whatever field.
It's a lot of work to get the best quality out of these things. Hours of iteration to create a good quality essay.
This makes me pretty happy in retrospect that I went to U of Chicago. The grade deflation was miserable in the moment. But in retrospect I'm glad I went through that - forced me to produce really good work and finally start getting my crap in order.
I read the Truman essay GPT generated. It was painful. The analysis on the atom bomb especially was unbelievably shallow and seems like something out of a middle school paper. Throughout the entire paper the overuse of $10 words is insane. And the general prose style feels like the sort of writing I'd do when I was desperate to reach a word count. This paper strikes me as something that would get a C or maybe a C+.
So yeah, I knew Harvard grade inflation was a problem, but holy crap.
The solution, as with all problems caused by AI, is Butlerian Jihad. We the people did not ask these tech freaks to try to make humanity obsolete. We must ban this crap before it gets even further out of control.
This is a great essay, Maya! I teach political science at the college level, and this helps me think about how to best handle generative AI. I really appreciate a student’s perspective on these issues.
Thanks Professor Arellano!
Yeah, I'm thinking blue book exams should make a major comeback. Maybe in addition to a mid-term, you have a draft-writing session for the term paper, and then some discussion sections become peer-review sessions so that people can discuss and come up with revisions in a space that is watched by people instead of doing major edits at home.
As an aside, of course grade inflation is worse at the Ivies than at UC Berkeley. It has always been true that private schools are easier to stay in than the prestigious public schools- private schools' motivation is to keep their students in school once they've accepted them, because they're either providing money or making the school look good somehow. The UC's primary motivation is to serve the students of California, and so if one flunks out, that frees up a spot for another qualified person to take their place.
> AI could automate the vast majority of the legal writing grunt work
There was a semi popular reddit thread where an AI enthusiast tried to use ChatGpt to 'help' their girlfriends legal case and her lawyer refused to look at any of it. This is kinda understandable. For a lawyer to sign off on something another person or AI generated, they would still need to read, follow citations and understand precedent to do due diligence. You could maybe have some form of insurance to hedge against mistakes? Maybe there is some rote form filling that can be done more efficiently? Still for AI specifically to help it comes with some guessing and risk. So I think it is still pretty unclear whether/how AI impacts law.
Nitpicky point that doesn't detract from the main thesis though.
This was a great article, I work in a counter-abuse role, and I wanted to mention that the "Google Docs Edit History" proposal is actually much better than it may look on the surface. This is because it's very hard for anyone to generate 'realistic' looking metadata, especially when 'realism' is going to be determined by some other AI system.
The team building the system that looks at the metadata to determine if it's fake has a good idea of what 'normal' edit history looks like because they'll have access to lots and lots of real examples. An individual trying to fake this isn't going to have access to a huge library of 'real' edit histories, so won't be able to fake it.
Exactly - if someone just takes a LLM-written essay and then hand-types it into Google docs, that’ll look like someone who just sat down and wrote an essay from start to finish in fifteen minutes, which is nothing like how a real person writes something.
Tim is full of great ideas!
I think ChatGPT may be the death of "bulk" writing. Maybe my view is skewed because I've been in the scientific realm for so long that, for most things, ChatGPT would provide an overall better written product. Maybe it's not such a big deal if people (outside of literary authors) are worse at long-form formal writing?
We already do most of our writing in short-form anyway (like responding to a blog post, which is admittedly longer form). As a result, my writing contains a higher density of ideas and less padding (even if those ideas are often garbage). Maybe the next "this meeting could have been an email" will be "this book could have been a blog post"?
I'm sorry, but you've failed to ask the most important question here:
What were your professors trying to teach you with those assignments? Were they graded on effort or on quality?
There are pedagogically-valid reasons to assign work that will be graded on effort, while never explicitly admitting as much to students.
For example: I'm a math TA, and I spent a couple hours Sunday grading student work. For some of the questions assigned, I (and the professor) didn't care if students got the right answer. We had asked the students to compute certain tables "by hand", because we wanted to force the students to experience every step of the algorithm. When I graded, I didn't check if they got every number in the table right*; if they'd filled in the table with anything other than random digits, they'd learned what we wanted to teach. (Also: it was a 1pt problem.)
In undergrad, I took a number of humanities classes, and many of them had weekly or biweekly short-essay submissions. None of them actually cared what we wrote; they just wanted to make sure that we were keeping up with the reading and thinking about what we read. A student used Chat-GPT to complete these submissions would be little different from a math student who used a calculator to cheat on their non-calculator homework: only cheating themselves.
Obviously, grading on effort is inappropriate for major exams. But were these major exams, or weekly check-ins? And if the TAs who graded Chat-GPT's work were checking whether it was keeping up with the reading, were they wrong to give it an "A"? After all, Chat-GPT is quite good at reading and interpreting human texts.
* If you're Prof. Hatice's Algebra course and reading this, then I make no guarantees to continue grading with this laxity!
Thanks, Maya.
Did any of the graders hazard a guess as to whether what they had written was produced by a bot?
You told them that there was some chance it was written by you, some chance it was written by the bot. I would have thought that many of them would mention this in their comments, and I am puzzled not to hear about it.
Esp. the writing course, which was scathing (and accurate) about the shortcomings of the submission.
So, this makes me wonder: did you instruct the graders *not* to mention whether they thought it was written by a bot? Or perhaps instruct them to give the comments that they would give if they believed it were written by a human?
Yeah some of them thought it was written by a bot
Hmmm...
I wish that you had included that info in your original post, since it changes the moral a bit. There's a difference between:
1) Bot fools Harvard profs, gets a lot of A's!
and
2) X out of 8 Harvard Profs detected the bot, but gave it good grades anyhow.
Humans still play chess despite being vastly inferior to computers. In fact using computers for training has enhanced the quality of play through new ideas and training. Similarly LLMs will not replace writing and will often enhance it.
Well, most people don't play chess professionally; a lot of people essentially write for a living
I'll go out on a limb and argue that there will be more great writing and more efficient bad but necessary writing (manuals, summaries, technical docs, etc). Second part is fairly straightforward, no one's career goal as an engineer or data scientist is to write technical documents for a product, however these roles require technical product knowledge. If I can train an LMM to read and document code to a template, then I edit, my work is now more interesting and I can spend saved time improving my product. There is no shortage of productive work to do and no shortage of necessary and time intensive writing.
Writing is a time and focus intensive endeavor. Very few people have the temperament to be writers. Many people have interesting ideas and literal sensibilities. If you allow for LLM augmented writing you expand the population of writers and create more literature. A greater more diverse pool of literature creates more opportunities for genius. Literature will become inherently better just as chess is better, benefiting authors and readers. Chess players are not making less money because computers play better.
For younger grades, especially grades 4-8, I'm worried. There is absolutely nothing stopping them from doing all their take home written work via AI. Those kids are already more tech savvy than their parents, and can find a way to access anything that's available. The time, resources, and agility required for elementary and middle schools to prevent this themselves seems like it's not there and parents broadly do not have the ability either (some will, but as a whole, nope). And these are crucial ages for writing skills that are life skills, not academic ones. I fear we'll have many students that come into that age range proficient in reading and writing and then slowly drift downward over time before it's obvious they've been using AI and need remedial help.
I think the question about higher Ed is much less important than younger grades. For higher Ed, so much depends on whether AI can break into the realm of consistently great analysis rather than the mere good writing it produces today (both possibilities seem plausible to me). If it can do that, we're talking about a major breakthrough and the ability to measure college students grades is small potatoes in comparison. If it can't break through, colleges will tweak assessment methods or raise standards and it'll be fine.
yeah but the homework in grade 4-8 is pretty BS, isn't it? In our district regular homework doesn't really even start until 7th grade.
Great point
Another path is they become a tool just like spellcheck. I can't spell but since I always type it's not a huge set-back. I could also see these tools becoming the norm for research, starting structure, proof-reading, grammar puncher-uppers. They would level-up the entire class.
If writing skill is affected by AI in the same way that spellcheck has affected spelling, I think we will be worse off. Writing is much more generalizable than spelling and is a conduit for novel ideas and analysis in a way that spelling is not.
I don't see how that would be the case. Professional writers have editors. Editors create value. A LLM powered "personal editor" would provide some approximate value.
I think we're talking about different things here. For adults, I see powerful and useful writing AI as a clear good. It's the impact on the formation of writing skills in children that I think would be bad.
Could be different things. I guess just don't see a problem if children are learning how to write in tandem with such an editor-like tool. I think it would have a rising-tides-lifts-all-boats effect. I do agree if children are just copy and pasting LLM outputs that's a problem. I think teachers will be able to solve for that though.
Yes. Excellent points.
I am aware of that. However, students would have to upload the relevant texts first to do that and most of them aren't that clever. They just pump in generic questions and provide generic answers which can then be critiqued for their lack of specificity.
These programs also require students to pay attention to the examples/evidence provided, and many don't do that either. For example, one student was "writing" about juvenile justice reform in the United States, and she ended up discussing an Indian law about juvenile justice at length. If she'd only clicked on the article she "cited," she would have seen that it had nothing to do with reform in the United States.