I have reached the bizarre point that when I get a properly crappy essay, I think to myself "hey at least they didn't use AI, or it would actually be better", and then I am inclined to give them a slightly higher grade just to reward them for having done this bad work for themselves.
I TAd intro political theory at my university three times for three different professors, and the in-class essays combined with a single final paper worked best: they get to practice the relevant skills in the essays, and then test them out in the final paper. It's not ideal, but I've learned that trusting students to actually care about learning makes them more likely to... actually care about learning. There are 360 students in that class, and it's impossible to punish the AI slop. But the students who use AI are already punishing themselves enough. At least that's how I see it.
I think a lot of difficulties of the open book multi day lockdown browser in class essay are caused by the facts that it is during normal class time on the students’ own laptops. If you can get the institution to treat a writing class like a chemistry class, in that it needs a several hour lab component separate from the main class, then you can have the students write the essays in a testing center on a computer that has all the readings (and whatever supplements you want to give them) but no browser and no AI, and a proctor to make sure they don’t do it on their phone on the side. (Though as someone pointed out recently, I guess you do have to take away their Meta Ray-Bans.)
We need supervised writing labs. It’s the only way to save the college essay. And it must be saved: only through writing can humans think in the most sophisticated ways they are capable of.
As a way to preserve something close to standard essay assignments, I like this approach. But I haven’t heard a convincing pitch for how we should help students build their ability to usefully integrate AI into their writing and thinking. Given that we now live in a world where everyone can and (probably should) include AI in their workflow, these kinds of solutions seem akin for game plans circa 2004 for making sure students use physical libraries instead of Google Scholar. Or maybe you think successful AI use depends on being able to write and think the old way first (don’t run before you walk)?
This last point is right. You can’t use AI to write well if you can’t tell the difference between mediocre and good writing, and by far the easiest way to learn that difference is to do a lot of writing yourself.
I think the appropriate comparison is to making sure that firefighters still lift weights and develop their physical strength even though they have all sorts of heavy machinery to work with.
Eli, thanks, I agree with Kenny and with you -- I think we should *also* teach students to integrate AI into their research (their writing not so much). I wrote about this in an earlier essay (link below) and also tried it out last semester. I have much more work to do to develop this aspect of my pedagogy. (Kenny teaches this though, I think, and should write up his approach!)
It’s more like making sure firefighters life weights at firefighter school only. When these students leave university, any student disposed to use AI to think for them will go right back to using AI again to help them write and think, and the skills they took from philosophy classes will atrophy
I appreciated this piece. In my first year writing course, I’m providing students a Google doc that contains tabs for all phases of the writing process for every project. I hope this helps them work through and value their own writing process. It also makes it easier for us to have a continuous conversation about their writing. I have a few other tricks up my sleeve as well - including a class Substack to give them an audience bigger than just me.
Great! The toupee fallacy I'll remember. I fell for that one for a few quarters. I've gotten out of the policing business altogether since then. But I do regularly talk with my students about their learning habits and the hazards of outsourcing cognitive labor. Having AI to go to college for you can only lead to employers having AI do your job.
Agreed, whatever else you do, talk to them and persuade them of the value of doing the work themselves (and expect that to have no effect on 10-20% of the class)
I've been doing the low-stakes take-home essay (very short) for years, long before AI was an issue. It works. AI provides another reason, but I'll stay focused on my original reasons. I've also, in some classes, done no-stakes homework. (I assume that the stakes in question here are grades.) These assignments are largely focused on 'small' skills ('formulate one objection','do this (logic) derivation', 'identify the strongest reason' -- generally simple things like that).
In either case, they do the assignment (or not) between Tuesday and Thursday (for example), and then in class on Thursday they get to share their answers, ask anything they want about the assignment, etc.. We talk about it for 40 minutes and then they take a short, directly relevant, quiz. (This procedure works extremely well in logic courses but it can work in other contexts.)
There is a long-format variation (couple of weeks).
More or less all of the students figure out pretty quickly that if they put some care into the low/no-stakes assignment, they fare much better, because class discussions are correction and reinforcement for them -- they aren't learning it (or not) for the first time. It doesn't mean that they all go on to take the appropriate action, but many of them do.
Thanks Michael! I've been doing *medium*-stakes weekly essays for many years and would continue doing them regardless of AI.
This is really helpful. One of the things I want to work on is finding ways to make the weekly essays even better prep for the exams, and maybe reserving class meetings to share essays, give feedback, and answer questions would further that aim.
Victor -- really appreciate this framing (and your last few essays -- Paul Bloom's recommendation turned me onto your work).
One thing I have been experimenting with at WashU (business school) for the oral exam is group papers and group discussions. So, imagine you have a class of 48 students. You divide this class into 12 groups of 4. For a final exam, each group is tasked with a shared written paper. They are welcome to use AI as a research tool to inform their work, though they are encouraged not use it to generate the writing itself (realizing I can't actually accurately assess this). After submitting, I have a final oral component. For this hypothetical class, I would have 3 x 1.5-hour sessions, each with 3 groups. Discussing each paper would take 30 minutes, a process I would lead with each group, pushing them on their ideas and seeing how they respond (individually versus collectively). Each person has at least 1-2 questions directed toward them about the shared work. The first 20 minutes I lead, and the last 10 are led by members of the other teams (who are asked to read each assignment in advance).
Student grades are based on their submitted writing (20%), their responses to live questions (70%), and their thoughtful questions of their peers (10%). Total grading time is thus reading and commenting on 12 papers in advance of the live sessions (2-3 hours?), 4.5 hours of live session discussion, and capturing and codifying the total grades at the end (.5). So... 7-8 hours in total? My hope is that this ends up as a good combination of allowing for advance time to work through an idea (using all available tools), testing if they actually learned about the topic through their own and the support of the available tools through the live discussion, and encouraging the skills of thoughtfully engaging with one's peers in a kind of dialectic.
Interesting, thanks for sharing, Peter. I love hearing about other approaches. I'm not sure collaborative papers work for philosophy, but I can see this approach being viable in other fields. I worry a bit about 70% of their grade being based on just 1-2 questions.
Good call! I’ll play around w percentages. The other adjustment a friend made was having students lead more of the 30 minute… trusting them to guide the conversation.
I think you're right about its value: actually sitting down and expressing ideas and arguments forces a person to clarify their own thinking. You can do something similar by giving a presentation. You can only really explain an idea if you know it, and you'll know it better if forced to try and explain it.
Shooting from the hip, here are some other (possibly dubious) ideas for maintaining the essay:
1. Heavily weighting the grading on originality of thought or the breadth of references. In philosophy, this might look like novel thought experiments and analogies or drawing examples from a wider literature.
2. Each person gives a 5 minute summary of their essay position to the class which is half the mark (introverts would hate this).
3. Invite students to give an auto ethnographical frontispiece to their essays explaining how they came to think about the topic and what their influences were, how it ties into their personal history.
I worry about outsourcing 1 and 2 to AI, but I like the idea of a hybrid essays/presentations. Introverts may hate it, but verbal communication is a skill worth building and evaluating, just like test-taking and essay-writing.
I imagined that people would typically give a five minute talk with the speaking improvised, but yes if they had a script then it could be gamed with an AI, you're right.
But as for #1— I might be wrong, but I think that while the best LLM writing can be difficult to distinguish from some people's handcrafted prose, the reverse isn't true: there are many ways of writing that cannot be mistaken for generated prose. You can read Nietzsche, Mill, or J. L. Austin and hear a voice and style in the text that couldn't be mistaken for the voice of ChatGPT and friends. Demonstrating a distinct style through particular inventiveness or particularly broad and eccentric reading could be graded.
I wonder whether letting students use AI freely, but asking them to at least use their own notes and go over the output, aligning it with their own style of writing, would be good enough to disincentivise them from just copy-pasting and not doing any work by themselves.
I have reached the bizarre point that when I get a properly crappy essay, I think to myself "hey at least they didn't use AI, or it would actually be better", and then I am inclined to give them a slightly higher grade just to reward them for having done this bad work for themselves.
I TAd intro political theory at my university three times for three different professors, and the in-class essays combined with a single final paper worked best: they get to practice the relevant skills in the essays, and then test them out in the final paper. It's not ideal, but I've learned that trusting students to actually care about learning makes them more likely to... actually care about learning. There are 360 students in that class, and it's impossible to punish the AI slop. But the students who use AI are already punishing themselves enough. At least that's how I see it.
I think a lot of difficulties of the open book multi day lockdown browser in class essay are caused by the facts that it is during normal class time on the students’ own laptops. If you can get the institution to treat a writing class like a chemistry class, in that it needs a several hour lab component separate from the main class, then you can have the students write the essays in a testing center on a computer that has all the readings (and whatever supplements you want to give them) but no browser and no AI, and a proctor to make sure they don’t do it on their phone on the side. (Though as someone pointed out recently, I guess you do have to take away their Meta Ray-Bans.)
We need supervised writing labs. It’s the only way to save the college essay. And it must be saved: only through writing can humans think in the most sophisticated ways they are capable of.
I love the chemistry lab analogy btw. That’s what we need to say when we ask for funding for this.
As a way to preserve something close to standard essay assignments, I like this approach. But I haven’t heard a convincing pitch for how we should help students build their ability to usefully integrate AI into their writing and thinking. Given that we now live in a world where everyone can and (probably should) include AI in their workflow, these kinds of solutions seem akin for game plans circa 2004 for making sure students use physical libraries instead of Google Scholar. Or maybe you think successful AI use depends on being able to write and think the old way first (don’t run before you walk)?
This last point is right. You can’t use AI to write well if you can’t tell the difference between mediocre and good writing, and by far the easiest way to learn that difference is to do a lot of writing yourself.
I think the appropriate comparison is to making sure that firefighters still lift weights and develop their physical strength even though they have all sorts of heavy machinery to work with.
Eli, thanks, I agree with Kenny and with you -- I think we should *also* teach students to integrate AI into their research (their writing not so much). I wrote about this in an earlier essay (link below) and also tried it out last semester. I have much more work to do to develop this aspect of my pedagogy. (Kenny teaches this though, I think, and should write up his approach!)
https://openquestionsblog.substack.com/p/you-dont-need-an-ai-policy-you-need
It’s more like making sure firefighters life weights at firefighter school only. When these students leave university, any student disposed to use AI to think for them will go right back to using AI again to help them write and think, and the skills they took from philosophy classes will atrophy
I appreciated this piece. In my first year writing course, I’m providing students a Google doc that contains tabs for all phases of the writing process for every project. I hope this helps them work through and value their own writing process. It also makes it easier for us to have a continuous conversation about their writing. I have a few other tricks up my sleeve as well - including a class Substack to give them an audience bigger than just me.
this looks winning all round - the retvrn of the oral exam, with a twist
https://www.behind-the-enemy-lines.com/2025/12/fighting-fire-with-fire-scalable-oral.html
Great! The toupee fallacy I'll remember. I fell for that one for a few quarters. I've gotten out of the policing business altogether since then. But I do regularly talk with my students about their learning habits and the hazards of outsourcing cognitive labor. Having AI to go to college for you can only lead to employers having AI do your job.
Agreed, whatever else you do, talk to them and persuade them of the value of doing the work themselves (and expect that to have no effect on 10-20% of the class)
I've been doing the low-stakes take-home essay (very short) for years, long before AI was an issue. It works. AI provides another reason, but I'll stay focused on my original reasons. I've also, in some classes, done no-stakes homework. (I assume that the stakes in question here are grades.) These assignments are largely focused on 'small' skills ('formulate one objection','do this (logic) derivation', 'identify the strongest reason' -- generally simple things like that).
In either case, they do the assignment (or not) between Tuesday and Thursday (for example), and then in class on Thursday they get to share their answers, ask anything they want about the assignment, etc.. We talk about it for 40 minutes and then they take a short, directly relevant, quiz. (This procedure works extremely well in logic courses but it can work in other contexts.)
There is a long-format variation (couple of weeks).
More or less all of the students figure out pretty quickly that if they put some care into the low/no-stakes assignment, they fare much better, because class discussions are correction and reinforcement for them -- they aren't learning it (or not) for the first time. It doesn't mean that they all go on to take the appropriate action, but many of them do.
Thanks Michael! I've been doing *medium*-stakes weekly essays for many years and would continue doing them regardless of AI.
This is really helpful. One of the things I want to work on is finding ways to make the weekly essays even better prep for the exams, and maybe reserving class meetings to share essays, give feedback, and answer questions would further that aim.
Victor -- really appreciate this framing (and your last few essays -- Paul Bloom's recommendation turned me onto your work).
One thing I have been experimenting with at WashU (business school) for the oral exam is group papers and group discussions. So, imagine you have a class of 48 students. You divide this class into 12 groups of 4. For a final exam, each group is tasked with a shared written paper. They are welcome to use AI as a research tool to inform their work, though they are encouraged not use it to generate the writing itself (realizing I can't actually accurately assess this). After submitting, I have a final oral component. For this hypothetical class, I would have 3 x 1.5-hour sessions, each with 3 groups. Discussing each paper would take 30 minutes, a process I would lead with each group, pushing them on their ideas and seeing how they respond (individually versus collectively). Each person has at least 1-2 questions directed toward them about the shared work. The first 20 minutes I lead, and the last 10 are led by members of the other teams (who are asked to read each assignment in advance).
Student grades are based on their submitted writing (20%), their responses to live questions (70%), and their thoughtful questions of their peers (10%). Total grading time is thus reading and commenting on 12 papers in advance of the live sessions (2-3 hours?), 4.5 hours of live session discussion, and capturing and codifying the total grades at the end (.5). So... 7-8 hours in total? My hope is that this ends up as a good combination of allowing for advance time to work through an idea (using all available tools), testing if they actually learned about the topic through their own and the support of the available tools through the live discussion, and encouraging the skills of thoughtfully engaging with one's peers in a kind of dialectic.
Interesting, thanks for sharing, Peter. I love hearing about other approaches. I'm not sure collaborative papers work for philosophy, but I can see this approach being viable in other fields. I worry a bit about 70% of their grade being based on just 1-2 questions.
Good call! I’ll play around w percentages. The other adjustment a friend made was having students lead more of the 30 minute… trusting them to guide the conversation.
I think you're right about its value: actually sitting down and expressing ideas and arguments forces a person to clarify their own thinking. You can do something similar by giving a presentation. You can only really explain an idea if you know it, and you'll know it better if forced to try and explain it.
Shooting from the hip, here are some other (possibly dubious) ideas for maintaining the essay:
1. Heavily weighting the grading on originality of thought or the breadth of references. In philosophy, this might look like novel thought experiments and analogies or drawing examples from a wider literature.
2. Each person gives a 5 minute summary of their essay position to the class which is half the mark (introverts would hate this).
3. Invite students to give an auto ethnographical frontispiece to their essays explaining how they came to think about the topic and what their influences were, how it ties into their personal history.
I worry about outsourcing 1 and 2 to AI, but I like the idea of a hybrid essays/presentations. Introverts may hate it, but verbal communication is a skill worth building and evaluating, just like test-taking and essay-writing.
I imagined that people would typically give a five minute talk with the speaking improvised, but yes if they had a script then it could be gamed with an AI, you're right.
But as for #1— I might be wrong, but I think that while the best LLM writing can be difficult to distinguish from some people's handcrafted prose, the reverse isn't true: there are many ways of writing that cannot be mistaken for generated prose. You can read Nietzsche, Mill, or J. L. Austin and hear a voice and style in the text that couldn't be mistaken for the voice of ChatGPT and friends. Demonstrating a distinct style through particular inventiveness or particularly broad and eccentric reading could be graded.
I wonder whether letting students use AI freely, but asking them to at least use their own notes and go over the output, aligning it with their own style of writing, would be good enough to disincentivise them from just copy-pasting and not doing any work by themselves.