Yesterday students turned in their final papers for the semester for my history of economic thought class, and I would like to share what I learned how I can detect almost detect if ChatGPT wrote the paper.
Identifying an essay written with AI tools like ChatGPT can be challenging but not impossible. AI-generated essays often exhibit overly formal and consistent styles, lacking the personal voice and varied sentence structures typical of human writers. They may provide surface-level analysis with an abundance of examples that are not deeply explored or connected. Additionally, these essays can be overly polished, with perfect grammar and structure, which might feel unnatural if the author usually makes minor errors.
Repetition and patterned language are common markers, as AI tends to reuse phrases or ideas. Furthermore, AI might include inaccuracies or fabricated details that sound plausible but do not hold up under scrutiny. Such essays also tend to lack emotional depth or a personal touch, making them feel detached and impersonal.
To identify AI involvement, tools like AI detectors and plagiarism checkers can help, as can discussions with the author to test their knowledge of the essay’s content. While no method is foolproof, combining these strategies can often reveal whether an essay is AI-generated.
I don't know. I mean, I think you can probably catch some things but with a decent false positive and false negative rate, enough that I don't really want to do it. I think the more likely big picture solution is to complete redesign from the ground up what a learning objective is, what a learning task is to achieve that objective, all with eyes wide open that there's tools that cannot be monitored. I think a ton of the current catching is because of a cohort effect. Going forward I bet those vanish. I honestly don't remotely envy fields where the goal is to teach students how to write. All I can think of is they have to perform at a level where the performance itself does not use AI on those learning objectives where I think it's critical they learn that thing. Some things you really don't care that technology has automated it. For instance, in certain statistical modeling, the underlying calculations are all done by the computer itself. I don't need to take a determinant or invert matrices anymore for the actual calculation. So having a computer automate that seems fine. There's probably a ton of examples like that throughout time too. IT's just this seems different because here you an actually produce the entire learning output and have really spent 0 time on it. SO if that's the realistic situation, I think the professor has to figure out how to create learning objectives and assignments that require maximizing time use, but I just don't know how to do that.
The in person part seems like it’s got to be right. I think anyway. In person stuff if high stakes will force outside work. And who cares if they use the AI for outside work?
But these still don’t clearly articulate if learning objectives changed. It’s possible they’ve completely changed. I’m not sure what mine are bc they vary by class. But that seems like the first step. And I agree the physical location issues seem like first order things to focus on.
The irony of hallucinations in the final papers despite the crit assignments is amazing.
I've been worried about how LLMs affect education, specifically learning, but I found myself optimistic about your approach for your history of economics course. In retrospect, it seems obvious that some students wouldn't learn/apply what was intended with the criteria assignments... I suspect there's a sizeable proportion of students who don't need the lesson - they're likely predisposed to fact check a summary - and there's probably another sizeable group who don't care and won't ever care. It makes me wonder what how many students need the lesson and are actually motivated by it.
I don't yet have issues with LLMs in my courses, which is surprising given they are entry-level programming courses. I'm not sure exactly why, but I suspect it's because I'm at a technical college and enrollment has been almost exclusively non-traditional students who are looking to upskill after hitting a career plateau.
I look forward to seeing how you and other instructors continue to adapt and what the results are like.
Yeah I wasn’t worried either. But I think maybe what it could be is that I can be “not worried” but only about certain learning objectives. But then maybe it doesn’t transfer as easily to other learning objectives and tasks. I also have them use LLMs for my causal inference class and don’t think twice about it. I did this intensive designing of this class so that I could have a final paper.
But honestly, just bc that happened doesn’t mean it didn’t work — that’s what I was thinking last night. It’s possible that the entire thing did achieve the outcome. Just because the final papers turned out a certain way doesn’t mean that student learning fell. Makes me think we need to experiment and test not just anecdotally reflect, which is what I’ve been doing
Your assignments sound really interesting and I know a lot of other educators are looking for ways to integrate LLMs into assessments. You should consider writing up a summary as a paper for Journal of Economics Teaching!
Thank you for writing this. My institution and public health department are currently grappling with how to handle AI use in the classroom. Many of my students are online-only executive graduate students. I shared this essay with my department. Everyone was quite receptive and appreciated your detailed, nuanced insights. Personally, I think your concluding thoughts will help shape my approach to teaching and assessing students as AI becomes more engrained in education and professional systems. Thank you.
Great piece, thanks for sharing. I just finished TAing a ~70 person class for MPPs at UChicago and can attest to the flattened bottom end of the normal distribution. The worst papers weren’t really that bad, where previous years had people who full-on misunderstood the assignment, or the one or two whose writing, frankly, made me wonder for a moment how they got into grad school.
What was interesting, though, was a new class of errors that I would’ve glossed as “smart student, but clearly didn’t come to lectures” before chatGPT. The Harris school made attendance mandatory this year, so I know they actually were at lectures — one of my TA duties was the attendance sheet. (Let’s set aside whether requiring attendance from full-on adults is a good idea.)
Especially in a few cases where our professor used words to mean something other than their lay definition, we had papers explaining the lay definition, despite the hours we’d spent on the non-standard definition! I can’t say for sure that it’s AI in the way that I could if it’d said “certainly! let me give three examples:”, but it certainly smelled fishy.
Thanks for sharing your experience—it’s fascinating to hear that you’re seeing a similar flattening of the distribution, but with this new twist of errors tied to terms with non-standard definitions. I know what you mean abt how these errors ‘smell fishy’—it seems like gen AI can produce a polished surface-level answer, but it struggles with nuance unless the user explicitly asks for it. Which I think makes sense but I don’t yet have the words for why. It’s more maybe that these assignments are not merely writing intelligently. I think that’s where students relying on AI might unintentionally reveal a gap in their own engagement with the material.
Your point abt the specialized terms is a great example of where lecture content and deeper synthesis play a critical role. That’s probably something I need to reflect on longer. But my hunch generally was somehow the weights on outside work needs to fall, and inside work needs to rise.
That’s one of the directions that needs to be explored. How to design assignments or evaluations that make do not rely on surface-level AI outputs but which allow a student to use AI as an input to their own learning.
I teach writing in the university (as well as in a dual-enrollment high school class) and yes, it is very, very hard to prevent AI-generated essays. It is also discouraging how few teachers actually try to fight it. My method is to give very precise prompts, so that, as you say, they need to develop a unique thesis, and also to summarize and respond to particular texts. You can almost always tell when AI has been used, but the hard part is proving it. Aside from catching fake citations (which make it easy but are increasingly rare as students get savvier--though my h.s. students do it a lot), I look for exact matches between long strings of words (6-8) in the student essay and online publications (since of course AI is really just plagiarizing, piecemeal, from online publications). This is extremely time-consuming, but if you can find enough such matches and show them to the student, it is usually enough to get them to admit to using AI. The good news is that language is actually extremely individual; very rarely do two people string together 8 words in exactly the same way; so when that happens, again and again (especially when 1 person is a teenager and the other is a published author), you know that it is AI.
That sounds unbelievably frustrating and difficult. If our the majority of our jobs becomes about catching this, I just have to think we need to start over and rethink the overall leaning objectives we want to choose for them. Plus, what you’re describing crossed my mind too but the thing is, I bet it’s also very error prone. I ONLY felt comfortable penalizing the clear violation of my AI policy when I found fabricated cites. But I don’t know if I got this clearly across, but I only caught them in the first place bc the student cited my work. And I KNEW that what they said was wrong — like wrong on many levels, but one being the work of mine they cited didn’t exist. So to see that AND see them discuss the work, I sort of had my mind blown. Because that’s when I thought to myself “wait. Is this happening with other cited works then? The ones I don’t know anything about? Are they citing real works but making up the content of them?” Made me really feel unsure tbh of what to do.
Yes, it is. And right, with citations you have to actually read (at least part of) the sources they cite, to confirm that they are inaccurate. That's why it's best to assign particular sources to discuss, so you know them well and can tell when they're misrepresenting them. (Then you can always just give them bad grades without even calling out on the cheating.) Unfortunately pointing to matches between strings of words isn't good "proof," since the student can simply call it a coincidence (though those of us who study language, and really anyone who stops to think about it, know that coincidences like that don't happen six times in a piece of writing, not to mention the fact that there is an unmistakable quality of professional writing that distinguishes it from student writing); what I try to do is force a confession out of students by presenting them with this evidence. At the very least, I think it's important to call them out. But as you say it is really hard and frustrating and so most teachers, I gather, don't do it. A lot of writing teachers even try to be progressive and incorporate AI into their teaching, which to me is just thinly veiled defeatism, but I wouldn't say that to anyone's face.
It sounds like you’re walking the walk man. Please leave comments on here from time to time of what you’re thinking and I’ll do the same. I’ll be probably trying something different in the spring and will update my thoughts as I go.
Loved the way you integrate LLMs into your classroom! Why anyone would trust an LLM for citations is beyond me, but assignments like these are incredibly important for students and professors to understand the limitations of this technology. This is a great example of something we wrote in favor of recently! I’d love if you gave it a read and let me know your thoughts.
Identifying an essay written with AI tools like ChatGPT can be challenging but not impossible. AI-generated essays often exhibit overly formal and consistent styles, lacking the personal voice and varied sentence structures typical of human writers. They may provide surface-level analysis with an abundance of examples that are not deeply explored or connected. Additionally, these essays can be overly polished, with perfect grammar and structure, which might feel unnatural if the author usually makes minor errors.
Repetition and patterned language are common markers, as AI tends to reuse phrases or ideas. Furthermore, AI might include inaccuracies or fabricated details that sound plausible but do not hold up under scrutiny. Such essays also tend to lack emotional depth or a personal touch, making them feel detached and impersonal.
To identify AI involvement, tools like AI detectors and plagiarism checkers can help, as can discussions with the author to test their knowledge of the essay’s content. While no method is foolproof, combining these strategies can often reveal whether an essay is AI-generated.
See what I did there....
I don't know. I mean, I think you can probably catch some things but with a decent false positive and false negative rate, enough that I don't really want to do it. I think the more likely big picture solution is to complete redesign from the ground up what a learning objective is, what a learning task is to achieve that objective, all with eyes wide open that there's tools that cannot be monitored. I think a ton of the current catching is because of a cohort effect. Going forward I bet those vanish. I honestly don't remotely envy fields where the goal is to teach students how to write. All I can think of is they have to perform at a level where the performance itself does not use AI on those learning objectives where I think it's critical they learn that thing. Some things you really don't care that technology has automated it. For instance, in certain statistical modeling, the underlying calculations are all done by the computer itself. I don't need to take a determinant or invert matrices anymore for the actual calculation. So having a computer automate that seems fine. There's probably a ton of examples like that throughout time too. IT's just this seems different because here you an actually produce the entire learning output and have really spent 0 time on it. SO if that's the realistic situation, I think the professor has to figure out how to create learning objectives and assignments that require maximizing time use, but I just don't know how to do that.
Universities will have to audit.
That is, students will have to answer questions orally to prove they have the knowledge.
I don't know what else to suggest.
The in person part seems like it’s got to be right. I think anyway. In person stuff if high stakes will force outside work. And who cares if they use the AI for outside work?
But these still don’t clearly articulate if learning objectives changed. It’s possible they’ve completely changed. I’m not sure what mine are bc they vary by class. But that seems like the first step. And I agree the physical location issues seem like first order things to focus on.
Love the honest perspective. Thank you for sharing your experience in lucid detail. Loved reading it!
The irony of hallucinations in the final papers despite the crit assignments is amazing.
I've been worried about how LLMs affect education, specifically learning, but I found myself optimistic about your approach for your history of economics course. In retrospect, it seems obvious that some students wouldn't learn/apply what was intended with the criteria assignments... I suspect there's a sizeable proportion of students who don't need the lesson - they're likely predisposed to fact check a summary - and there's probably another sizeable group who don't care and won't ever care. It makes me wonder what how many students need the lesson and are actually motivated by it.
I don't yet have issues with LLMs in my courses, which is surprising given they are entry-level programming courses. I'm not sure exactly why, but I suspect it's because I'm at a technical college and enrollment has been almost exclusively non-traditional students who are looking to upskill after hitting a career plateau.
I look forward to seeing how you and other instructors continue to adapt and what the results are like.
Yeah I wasn’t worried either. But I think maybe what it could be is that I can be “not worried” but only about certain learning objectives. But then maybe it doesn’t transfer as easily to other learning objectives and tasks. I also have them use LLMs for my causal inference class and don’t think twice about it. I did this intensive designing of this class so that I could have a final paper.
But honestly, just bc that happened doesn’t mean it didn’t work — that’s what I was thinking last night. It’s possible that the entire thing did achieve the outcome. Just because the final papers turned out a certain way doesn’t mean that student learning fell. Makes me think we need to experiment and test not just anecdotally reflect, which is what I’ve been doing
Your assignments sound really interesting and I know a lot of other educators are looking for ways to integrate LLMs into assessments. You should consider writing up a summary as a paper for Journal of Economics Teaching!
Okay you nudged me. I just wrote an editor somewhere about whether they might be interested in something.
Thank you for writing this. My institution and public health department are currently grappling with how to handle AI use in the classroom. Many of my students are online-only executive graduate students. I shared this essay with my department. Everyone was quite receptive and appreciated your detailed, nuanced insights. Personally, I think your concluding thoughts will help shape my approach to teaching and assessing students as AI becomes more engrained in education and professional systems. Thank you.
Oh very cool! Thank you for telling me. Sorry for all the typos. I left them so people would know ChatGPT didn’t write it. :)
Great piece, thanks for sharing. I just finished TAing a ~70 person class for MPPs at UChicago and can attest to the flattened bottom end of the normal distribution. The worst papers weren’t really that bad, where previous years had people who full-on misunderstood the assignment, or the one or two whose writing, frankly, made me wonder for a moment how they got into grad school.
What was interesting, though, was a new class of errors that I would’ve glossed as “smart student, but clearly didn’t come to lectures” before chatGPT. The Harris school made attendance mandatory this year, so I know they actually were at lectures — one of my TA duties was the attendance sheet. (Let’s set aside whether requiring attendance from full-on adults is a good idea.)
Especially in a few cases where our professor used words to mean something other than their lay definition, we had papers explaining the lay definition, despite the hours we’d spent on the non-standard definition! I can’t say for sure that it’s AI in the way that I could if it’d said “certainly! let me give three examples:”, but it certainly smelled fishy.
Thanks for sharing your experience—it’s fascinating to hear that you’re seeing a similar flattening of the distribution, but with this new twist of errors tied to terms with non-standard definitions. I know what you mean abt how these errors ‘smell fishy’—it seems like gen AI can produce a polished surface-level answer, but it struggles with nuance unless the user explicitly asks for it. Which I think makes sense but I don’t yet have the words for why. It’s more maybe that these assignments are not merely writing intelligently. I think that’s where students relying on AI might unintentionally reveal a gap in their own engagement with the material.
Your point abt the specialized terms is a great example of where lecture content and deeper synthesis play a critical role. That’s probably something I need to reflect on longer. But my hunch generally was somehow the weights on outside work needs to fall, and inside work needs to rise.
That’s one of the directions that needs to be explored. How to design assignments or evaluations that make do not rely on surface-level AI outputs but which allow a student to use AI as an input to their own learning.
I teach writing in the university (as well as in a dual-enrollment high school class) and yes, it is very, very hard to prevent AI-generated essays. It is also discouraging how few teachers actually try to fight it. My method is to give very precise prompts, so that, as you say, they need to develop a unique thesis, and also to summarize and respond to particular texts. You can almost always tell when AI has been used, but the hard part is proving it. Aside from catching fake citations (which make it easy but are increasingly rare as students get savvier--though my h.s. students do it a lot), I look for exact matches between long strings of words (6-8) in the student essay and online publications (since of course AI is really just plagiarizing, piecemeal, from online publications). This is extremely time-consuming, but if you can find enough such matches and show them to the student, it is usually enough to get them to admit to using AI. The good news is that language is actually extremely individual; very rarely do two people string together 8 words in exactly the same way; so when that happens, again and again (especially when 1 person is a teenager and the other is a published author), you know that it is AI.
That sounds unbelievably frustrating and difficult. If our the majority of our jobs becomes about catching this, I just have to think we need to start over and rethink the overall leaning objectives we want to choose for them. Plus, what you’re describing crossed my mind too but the thing is, I bet it’s also very error prone. I ONLY felt comfortable penalizing the clear violation of my AI policy when I found fabricated cites. But I don’t know if I got this clearly across, but I only caught them in the first place bc the student cited my work. And I KNEW that what they said was wrong — like wrong on many levels, but one being the work of mine they cited didn’t exist. So to see that AND see them discuss the work, I sort of had my mind blown. Because that’s when I thought to myself “wait. Is this happening with other cited works then? The ones I don’t know anything about? Are they citing real works but making up the content of them?” Made me really feel unsure tbh of what to do.
Yes, it is. And right, with citations you have to actually read (at least part of) the sources they cite, to confirm that they are inaccurate. That's why it's best to assign particular sources to discuss, so you know them well and can tell when they're misrepresenting them. (Then you can always just give them bad grades without even calling out on the cheating.) Unfortunately pointing to matches between strings of words isn't good "proof," since the student can simply call it a coincidence (though those of us who study language, and really anyone who stops to think about it, know that coincidences like that don't happen six times in a piece of writing, not to mention the fact that there is an unmistakable quality of professional writing that distinguishes it from student writing); what I try to do is force a confession out of students by presenting them with this evidence. At the very least, I think it's important to call them out. But as you say it is really hard and frustrating and so most teachers, I gather, don't do it. A lot of writing teachers even try to be progressive and incorporate AI into their teaching, which to me is just thinly veiled defeatism, but I wouldn't say that to anyone's face.
It sounds like you’re walking the walk man. Please leave comments on here from time to time of what you’re thinking and I’ll do the same. I’ll be probably trying something different in the spring and will update my thoughts as I go.
Loved the way you integrate LLMs into your classroom! Why anyone would trust an LLM for citations is beyond me, but assignments like these are incredibly important for students and professors to understand the limitations of this technology. This is a great example of something we wrote in favor of recently! I’d love if you gave it a read and let me know your thoughts.
https://open.substack.com/pub/aalokbhattacharya/p/how-teachers-should-delve-into-ai?r=4hgw27&utm_medium=ios