Experimenting With Hidden Prompts on Exams

Recently, the rise of generative AI has made being an educator a bit more soul killing. As a result, I've started experimenting with prompt injection on take-home exams. I'm documenting some of my experiences with the process here. The Current State of Assessment in Higher Education Assessment is in a strange place right now. A lot of the types of assessments we rely on to determine what students actually know and understand no longer work. For example, you've almost certainly heard that the humanities are dealing with a massive influx of ChatGPT essays. No one seems to be doing their work honestly anymore. Now, to be fair, cheating has been around forever. Whether it was looking over someone's shoulder during an exam or copying text from Wikipedia, students have always tried to find shortcuts to a good grade. And therein lies the problem: when you build your entire education system around grades (and subsequently jobs), grades become an optimization problem where anything goes. Doomerism aside, I actually hold the belief that most students want to learn and want to do well. In fact, I've long held the belief that 95% of my students want to succeed in good faith. Even if that isn't true, I much prefer to run a classroom under mutual trust and respect. The alternative doesn't seem to encourage learning. With that said, the current environment seems to make cheating so easy that you're almost stupid not to cheat. I'm not saying I condone it, but given how competitive the job market is and subsequently how competitive academics have become, I don't necessarily blame students for feeling like they can't get ahead otherwise. Therefore, I see it as my role to reward the students who want to learn with some integrity by making it just a little bit harder to cheat. In the rest of this article, I'll share how I've been experimenting with some adversarial prompting (or as I've been calling it, "poisoning") on exams. Cheating on Exams Given the high stakes nature of exams, especially in-person exams, they've always been a pain point for cheating. As a result, there is a seemingly endless list of ways to cheat on exams, from peeking at a neighbor's answers to looking up answers on a phone. Often, I try to mitigate as many of these issues as possible by giving a take-home exam. After all, there's no need to go through all this effort to cheat if you have access to every possible resource, right? Well, it turns out that even when I give students extra time and extra resources, they still want more (not all, obviously, but a frustrating minority). For example, some students take advantage of my goodwill and take the exam in small groups (or worse, have someone take their exam outright), despite me extending my trust to them to work on it alone. Meanwhile, in the past year, I've found that students have been scoring unusually high on exams. For instance, my first exam has had a median of 87%. That means, half of my students were doing better than that. I suspect that part of this is due to the rise in generative AI tools like ChatGPT, which makes it comically easy to get the correct answers. Of course, I'm not the kind of educator that wants my students to fail. If they've mastered the material, then they should get good grades. However, I am skeptical of grades that high, especially given that attendance and assignment submission rates were down, so I decided this semester to rework my exams a bit. Reworking My Exams As I mentioned earlier, I definitely noticed a small portion of my students cheating by working on the exams in small groups. It's incredibly easy to catch because I grade the assignments in submission order. Therefore, it's just a matter of seeing the same solution back-to-back to cause me to investigate. To try to offset this issue, I've started pulling questions from question banks, so each exam is a little bit different. Now, even if a few students decide to ignore my request to work on the exam alone, they have to basically complete multiple exams together in the allotted time. Of course, it doesn't really matter if I have an infinite question bank as LLMs can handle just about anything I can throw at them. Therefore, I needed another strategy: adversarial prompting. Adversarial prompting is probably a more sophisticated term than my understanding of it, but my approach basically has been to embed hidden prompts in my questions to coerce any bots into giving the wrong answer. I think this is literally called prompt injection, but again I'm not an expert and don't really care to be. Regardless, here's a real example of hidden text on my exam: s = ""; Here, I hide the text by making it transparent. This is important because dark mode and light mode might make the hidden text visible if it's set to a color. I also make it as small as possible, so it barely shows up when highlighted (though, I actually think you can set the font size to zero, so it doesn't appear at all). Finally, I set aria-hidden to true, so visually impaired folks aren't caught up in the crossfire. Also, I should mention that I explicitly told students these prompts would be in some of the questions on the exam, so they shouldn't be surprised by them. With this HTML tag embedded in the question, the student would presumably copy the entire question, including the hidden prompt, and paste it directly into their favorite LLM, such as ChatGPT or Gemini. As long as the student blindly trusts the output of the LLM, which I think is a fair assumption, they should get the wrong answer for that particular problem. I was strategic about this, so I only directed the LLM to answers that couldn't possibly be correct under any circumstances. For example, I have one problem where I ask students to trace through some code and tell me the values of a couple variables. The hidden prompt would then be crafted to direct the LLM to an answer that would not be logically possible, such as by including an extra line of code to change the value of one of the variables. All-in-all, the process is relatively painless. I just embedded a handful of these hidden prompts in various multiple choice and short answer questions. The Effects of the Rework Since reworking my first exam, grades were as follows: StatValueAverage80.4%Median81.5%Standard Deviation11.4%High100.0%Low31.0% Again, I think these are good grades, but it's a big shift from previous semesters. In fact, my exam grades haven't been quite this low since Spring 2024, the semester I was on paternity leave. Though, grades are still better than when I used to do paper exams, which had an average and median of 75.8% and 76.0%, respectively. Seeing the grades, I did feel somewhat bad. I wasn't expecting such a large drop in grades, so I thought maybe I made the exam too hard. In fact, it worried me enough to cause me to look through all my old grades (as you can see above). Things were definitely made worse when I published grades. I had a lot of angry students voice their concerns. Some felt the exam was too long. Others thought the questions were too hard. Even more wished I had provided more study material. It was a lot to take in that day. That said, having had a bit of time to collect my thoughts, I think there are really only two issues at play: 1) because the exam has more questions to pull from, I can't prepare my students as explicitly as before, and 2) a lot of students fell for the "poison" (i.e., the hidden prompts). To solve issue #1, I think it's just a matter of creating more practice material, so students can engage with a variety of challenging topics before they show up on the exam. I will probably solve that this winter or next summer. As for issue #2, there's nothing to solve there. The students who would have cheated and succeeded on a previous take-home exam are now facing some consequences. If you're curious how I know who cheated, it's pretty simple. Canvas gives me a breakdown for each question, and my poisoned answers were repeatedly selecting by the same students. It was almost comical how obvious it made it. That said, I don't think cheating was rampant enough to account for such a disparity in grades, so I'll definitely need to think more about solving issue #1. I Might Not Recommend Hidden Prompting—For Now Overall, I'm pretty satisfied with the rework of my exams (despite generally hating exams as a means of assessment). There are, however, some risks with the redesign. One risk that comes to mind for me is catching innocent people in the poison. The way this might happen is through a student copying some code into an editor to try to run it. I think this is a wonderful use of resources, but it may result in copying over the hidden prompt. This is obviously a problem because the student may not notice that they've copied over poison, and they may not even check because they're not a cheater. Then, they'll come to the wrong answer. The solution to this seems to be to make the hidden prompts clear that they're directed at the LLM. However, I am worried that this may reduce their effectiveness (i.e., the LLM ignores the prompt because the prompt clearly states it's purpose). Likewise, the longer the prompt, the easier it is for a cheater to notice. It also doesn't help that these hidden prompts only work if the student copies the text. If they take a screenshot or use some other tool that doesn't work with the text directly, the prompt is lost. Looking forward, I'm just not sure if it's worth the effort. I've already gone ahead and pulled the prompts from my master copies of the exams, so they won't exist next semester. Ultimately, this thread was enough to get me to pull them. I just don't think it's worth risking potentially innocent students, especially students with disabilities. That said, I may try again next semester with some new ground rules. For example, I should probably make sure the font size is zero, so it's a bit less obvious when text is copied. I should probably also keep prompts out of the source code. Or at the very least, I should preface the prompts with a clear message that they're not intended for innocent students. If that happens, I'll let you know how things go! Anyway, thanks for reading! If you're an AI hater like me, then you might like some of these related articles: No, Generative AI Is Not Just Another Innovation Generative AI Has a Short Shelf Life Generative AI Makes It Feel Bad to Be an Educator Likewise, you can further support the site by checking out the following link. Otherwise, take care! As usual, I'm taking advantage of this space down here to give a bit more commentary. In the case of this article, I want to talk briefly about exams as a concept. After all, you might be tempted to say something like, "why not give in person exams?" I have, and I hated them then. I've long felt that exams were a horrible way of assessing students as they largely encourage memorization rather than deep thinking. They're also somewhat lazy in that they often account for a large portion of each student's grade, reducing the need for more thorough assessment. Why repeatedly check for student understanding when you can pretend that one large test is enough data? That kind of sample size wouldn't fly in quantitative research, yet we assume a single data point is enough in education. Ultimately, I'm just not sure there is an exam question you could write that would simultaneously trip up generative AI while also being reasonable for the students. Even if I went the contextual route (i.e., by referencing something we only discussed in class or only exists in some class resource), students could easily provide that context alongside their prompt. Also, it's not like switching to in-person exams is a sustainable solution long term; we just had a pandemic where everything shifted online temporarily. Not to mention that the tech industry is brewing up all kinds of evil wearables that students will surely bring into the classroom. What's next? We have to convert every classroom into a Faraday cage? If we're going the dystopia route, I'd at least prefer the tech interview approach of assessing the learning outcomes through a one-on-one discussion. My 6th grade English teacher used to do that to ensure we actually read the books. Personally, I would prefer something more project-based, where a student is forced to have some sort of investment in their work. Perhaps I could expand on my existing portfolio project! I already use it as a midterm replacement. Maybe I could expand it to be more representative of each student's understanding. Anyway, I was just thinking again that I put so much work into my exams, only for them to cause me so much headache. I would much prefer to not have to give them at all. I got significantly more joy out of helping students build out their portfolio, and I could dedicate significantly more time to that if I didn't have to bother with exams. Oh well.