Here’s a problem: Microsoft has built their “copilot” directly into (at least) the online versions of Word. Students use this *and don’t know they’re using AI*. I’ve had students report this, explicitly. That is, I have had students declare they don’t use AI, but also declare they do use Microsoft’s copilot and see no conflict.

Copilot probably isn’t the only such example. Even if copilot is the only example it won’t be the only example for long. And to be clear: no you can’t ban Word. So perhaps (maybe) you can get through *this* semester with “hey copilot is AI too” and feel you’ve given all your students enough warning. But the fact is you won’t catch all uses of it and even if you did you won’t be dealing with the next thing of this shape to come down the line.

So while I agree it would be best to ban AI, I also think banning AI simply cannot be done. We’ve lost this fight, and we have to find some other way to deal with the moment.

So I was at the International Conference on Academic Integrity last month, and there’s this really interesting approach that’s starting to gain some serious traction among educators dealing with the whole ChatGPT problem. It’s called the “Socratic Faraday cage,” and honestly, when I first heard the name I thought it was a joke, but the more I learned about it, the more sense it makes. Basically, it’s a specialized assessment environment that completely isolates students from any electronic interference. Picture a small room—maybe 10×12 feet—lined with copper mesh that blocks all electromagnetic signals, so no wifi, no cellular, nothing gets through. The walls are transparent from the outside so faculty can observe, but the student inside is completely cut off from any digital assistance. The assessment itself is purely oral and takes place over three intensive days. Students are sustained with basic meals and water, and they engage in continuous Socratic dialogue with rotating panels of faculty members who pose increasingly complex questions across their field of study. No notes, no books, no technology—just pure intellectual engagement.
The pilot programs at a few universities (not yet in the US, but it’s been done already in some EU countries apparently) have shown really promising results. Students (trivially) can’t rely on AI assistance, obviously, but more importantly, the format seems to reveal genuine understanding versus surface-level knowledge in ways that traditional testing can’t match (this is especially so in philosophy). The three-day format, while the students are in the SFC, allows for seriously deep exploration of concepts (what is justice? do we have free will? etc) and really tests intellectual stamina and authentic learning. I know it sounds extreme, (I hope it goes without saying there are consent forms) but given how sophisticated AI tools are becoming, some institutions are seriously considering this as the gold standard for high-stakes assessment. The early data I heard about at the conference suggests it’s actually really effective at distinguishing between students who truly understand their material versus those who’ve been leaning heavily on AI assistance, this will be even more important when Altman rolls out GPT-5 later this summer most likely.

I’m not a professor, but in my mind, this could be solved by Universities or departments willing and able to issue devices that can be used for assignments but are not the property of the students. This can be viewed as an IT issue, or perhaps an issue of finding the willingness to do something new and proactive with IT budgets. If I could speak to Universities or departments as if they were one person, I would ask them to please stop relying on a Bring-Your-Own-Device policy for academic assignments.

In terms of IT tools able to restrict LLM access, I find that solutions like the LockDown browser to be “lose-lose” compromise solutions where students are made to download software that reduces the functionality of their devices, and professors have to spend relatively large amounts of time to achieve the same level of visibility and integrity as a proctored written essay.

As far as I am aware, there is no easy and ethical way for LockDown or other tools to monitor other devices in the student’s vicinity or on the same network, and it is trivial for students to use a second device to cheat while LockDown is none the wiser.

That all being said, much of the legal and ethical difficulties can be done away by providing managed computers where there is no expectation of privacy.

Students trying to reach an LLM during an assignment? Configure a DNS server or service to refuse queries to all undesired sites and flag them. This is done in practice in worksites around the world to prevent and flag access to gambling or adult entertainment while working.

Want to have take-home assignments? Force them to use a VPN on the managed computer to do any work, and all web activity from the managed device can be safely scrutinized. This is a norm for most of those who work remotely with a company-provided device.

Worried about them using their phone or their personal computer? Mandate that the camera is on and recording while working with a mirror placed behind the student for additional visibility of their workspace. While this may seem draconian to many, this was an everyday reality for office workers in Mexico’s equivalent of the IRS, the SAT.

Worried about them successfully cheating by modifying the managed computer or installing an unauthorized program? This is near impossible to do without tripping alerts or alarms in a well-managed IT system.

In my view, a deeper level of collaboration and conversation with one’s IT department could easily eliminate the need for tools like the LockDown browser in the long term, especially if the University has already had to invest in security tools for other areas like Medical and Law schools.

Finally, providing managed devices communicates to students a level of rigor, seriousness, and respect for their property rights. Some of us are sincerely concerned about maintaining our digital freedom in the republican sense as it extends to our personal devices, and software requirements like the Lockdown browser communicate to us that, when push comes to shove, you actually don’t have a choice when it comes to installing Google, Apple, or Microsoft operating systems on your personal property and granting them non-trivial amounts of control over your digital life (Academic integrity software is often built without support for open-source operating systems).

Long story short, I find that relying on software for students’ personal devices to ensure integrity on academic assignments is, at this time, self-defeating and conveys a lack of respect. If cheating really is to be taken seriously, in a similarly grave manner to when I am on systems where I can access sensitive information at work, students ought to be provided managed devices.

I think Prof. Kukla’s syllabus statement is a reasonably good example of what one might use in the case of prohibitive policy on generative AI use in one’s class. It is fairly specific, delineating what is and is not allowed. It acknowledges that there are various types of AI and various uses the technology may be put to, and it clearly spells out sanctions. It also articulates a procedure (the use of Google docs) that everyone can follow and that allows for evidence-based arbitration of academic dishonesty disputes, if and when they arise. (As an aside, at my institution, which uses Microsoft products, there is a way to use and save Word docs to the cloud in a manner that maintains a version history. Perhaps there is a way of replicating the Google doc procedure on a Microsoft campus. I’m not sure about this.)

In my view, Kukla’s sample policy would be improved if it were to be paired with (i) a rationale and (ii) a statement of expectations for the instructor.

(i) may or may not be spelled out in writing. It may be enough to discuss this in class and/or post it in video form on the class site. The idea, though, is that it is crucial to explain to students why one has the policy on has, with reference to the learning objectives in the course and how the assignments and policy relate to that. Even better if one opens up discussion with the class about what the correct policy should be and why, allowing students to weigh in and ask questions. Perhaps best to structure this so as to result in something like a contract developed by all parties involved. It’s not likely that students will try and band together to game the system and try and get the course policy to permit them to use generative AI in any way they’d like; and anyway, as the instructor you have the authority to veto this. And the process of group norm development has several key benefits, including increased buy-in and clear demonstration (not mere paying of lip service) that one takes the classroom (in-person or online) to be a site of mutual inquiry and respect. One could begin the term with something like the above syllabus policy and note explicitly that it will be subject to revision based on mutual agreement.

(ii) is a best practice anyway, and it is important as a means of showing students that one is not a hypocrite (eg, prohibiting their use of gen AI and yet using it oneself, say, to grade their work or write one’s own papers). It also signals to them, once again, that the classroom is a space of mutual respect and inquiry, where everyone is in this together. Making explicit that there are expectations on both sides of the teacher-student relationship does not, of course, mean ceding all authority. It does really change the dynamic, for the better. This is one reason it’s a best practice from times before gen AI tools.

In terms of assignments, I’ll add two brief comments. The first is that we’ve known for a long time that the practice of merely assigning a take-home essay in response to a prompt is not a good way of teaching writing or assessing learning. It doesn’t do the second thing (academic dishonesty issues aside), because one may write a good essay based on prior knowledge or knowledge gained from another course, and so the end product may not reflect anything gained in the relevant course. It doesn’t do the first thing because there is much more to teaching writing than crafting a good prompt and providing students the opportunity to answer it. So, I don’t see much reason to try and figure out how to AI-proof the take-home essay if what’s meant by that is trying to figure out how to go about doing a poor job of teaching writing and assessing learning.

One way to improve writing instruction is to scaffold writing assignments. This also–and this is my second comment–allows one to establish a baseline by which to judge a student’s progress and learning. So, to answerJustin’s question about assignments, one thing to do is to include a number of low-stakes writing assignments early on in one’s course (eg, minute papers) that both allow students the opportunity to practice their writing on philosophical topics (which may be strange both in terms of form and content) and allow the instructor to get a basic sense of each student’s skills and knowledge. You can assign these as pencil-and-paper (or use a lockdown browser for an online course). Later writing artifacts, even those completed at home, can be compared to these earlier ones, both for assessing progress and in cases of suspected dishonesty. It’s not a perfect system, but it is based on good teaching practices and is not overly onerous. Perhaps it will be helpful.

By way of closing, I’ll add that my experience is that none of the above takes so much time, either class time or grading time, so as to be prohibitive. And in fact, my experience is that well-structured assignments, repeated practice, and scaffolding end up reducing the amount of time I spend grading, since they result in much better papers by my students. So, there are benefits for me, as well as for my students.

Designing assessments that are relatively LLM-resistant is ideal, when you can do it.

Here’s some evidence that the online argument mapping assessments included in my ebook+ With Good Reason are pretty LLM-resistant:

https://docs.google.com/presentation/d/1ZCEDafDzQDUQ5uO8rVKeLDMn_UMuM80-EV56nSyrY-I/edit?usp=sharing

Anything graded is now done in front of me. We simply cannot trust any at-home assessment not to be done by AI. In most classes, this means blue book exams and quizzes. Phones are gone, smartwatches are gone, earbuds are gone. In freshman seminar, where we’re required to assign essay writing, I do all in-class writing, in Blackboard with the Respondus thing that locks out all internet, along the lines of that long post here a couple months ago. Other than freshman seminar, though, the basic short essay is gone. But I have actually liked using blue book exams – it enables me to assess them more comprehensively. I give them a list of questions in advance, a subset of which will be on the exam. This obliges them to study and prepare across a wide range of topics, not just the 2 things that they’d need to write a paper.

On the subject of bathroom breaks during exams, there is an easy way around the dilemma: give your exams in parts. When a student goes to the bathroom, they’re required to turn in the part they’re on, picking up the next part when they return. If you break your exams into two or three parts, it’s possible to accommodate those with urgent bathroom needs without worrying about LLM-use in the stalls.

I’m a TA in our math department, so things are a little different, but I’ve been pushing for more oral examinations, not just as a testing method, but as a way of getting students to “prove” they didn’t use AI. This has a few drawbacks (math anxiety affecting certain student populations more than others), but it has some real upsides.

First, the cognitive impacts of AI usage are well documented, and even just asking students to summarize the work “they” have done is a very easy way to weed out AI usage. Data shows that students understanding of work done with AI is very limited, and so just asking for a summary of technique (or arguments made in a paper) can be really helpful.

Secondly, this develops a certain level of higher level thinking that blind usage of AI cannot replicate. If a student reads all the text generated about how to solve a problem, rather than just skipping to the worked problem (they rarely do), they can still develop an understanding of the technique. If a student is going to AI to learn, rather than merely produce work, I am not too upset about AI usage. Oral questions, or making a student work something on the board, gives us space to correct the students numerical work (without a penalty to the student for missed negative signs and the like) to show an understanding of the concept underlying the computation.

More importantly, this incentivizes learning even when a student may have cheated on an assignment. If the student can learn what they needed for the exam between the exam and being asked oral questions, we have succeeded in teaching them the material.

Ultimately, the issue with AI usage isn’t that there’s something highly immoral we need to prevent and discipline against, or that the work is not theirs; the biggest issues are that cheating and AI actively hinders the learning process for students. Catching cheating is not our goal, teaching is, and IIRC there was an article on here discussing whether or not strict policies were an unethical overstepping. We will never be able to catch all students who are cheating (focus on AI ignores online paperwriting services, but there is less focus on that), and some pragmatics have to be considered in our enforcement. I would rather an anti-AI policy incentivize students to learn (even if in violation of the policy) than be extremely effective in catching all students who have touched AI.