I was recently invited to give a keynote at the third annual Teaching Amid AI Conference - the largest of its kind in Canada (600+ attendees). The feedback on my talk was (mostly) positive so I re-worked the talk into the essay below. I’d love to hear your thoughts in the comments.
Like many of my colleagues, I scrapped take-home essays to protect academic integrity. But then we had to reckon with the fact that fourth-year philosophy students wanting to apply to graduate school didn’t have writing samples. They’d never written a term paper.
This is what a “solution” looks like when ethics ignores trade-offs.
In a previous essay, I argued that the field of AI ethics is one-sided, misleadingly negative, and myopically focused on harms. We’re in the midst of an AI Ethics Winter. Meanwhile, AI capabilities are advancing relentlessly. Nearly every AI policy at every level - from individual syllabi to university guidelines - evokes the ethical use of AI. What would an ethics look like that educators can actually use?
Good vs Good
At the beginning of each semester, I tell my undergraduate students that it’s quite natural to think ethics is about good vs evil. If that were true, things would be easy and I’d be out of a job. We’d just do the good things and avoid the evil things. But ethics is rarely about good vs evil. It’s almost always about good vs good. Ethics fundamentally involves trade-offs between different things that are good in different ways.
If AI was evil, and all it did was destroy the environment, produce biased outputs, and rot our brains, then things would be easy. No one would use it. Yet, one billion people use ChatGPT monthly. It’s possible this is planetary-scale false consciousness. But the much more likely explanation is that many people just find AI useful, warts and all.
This does not negate environmental impacts, bias concerns, or worries about cognitive surrender. It just shows that, like most things in life, AI is a mixed bag. It is extremely unlikely that a technology as complex and far-reaching as AI would fit neatly into an “All Good” or “All Evil” box.
The path out of the AI Ethics Winter towards policy-relevance, then, requires framing ethics less in terms of harms and more in terms of trade-offs. Everything worth talking about in AI involves trade-offs.
I can get Claude to wade through the deluge of AI papers published each week and give me summaries. Or I can download them to the black hole that is my “AI TO READ” folder. Summaries are surely better than nothing, but then again, isn’t the point of my job to read the papers?
What about using AI as an editor for writing? It’s helpful, cheaper, and more readily available than the alternatives. But people who spend a lot of time on Substack can already glimpse the homogenization resulting from Claude providing editorial direction to most of their feed.
The list goes on. What should be clear is that there’s no “solutions” to the problems posed by AI in education. It’s trade-offs all the way down. We owe it to our students to honestly and explicitly name the trade-offs involved with using and not using AI, and to model careful reasoning about them. We abdicate our responsibility as educators when we substitute this with simplistic, tribalistic, good vs evil sloganeering.
But if it really is trade-offs all the way down, how do we resolve them?
A Positive Vision
Nathan Labenz, host of the popular AI podcast, The Cognitive Revolution, often says that the scarcest resource today in AI isn’t GPUs, compute, or researchers, but a positive vision. This is as true for AI companies as it is for universities.
A positive vision is not unreflective techno-optimism, nor a manic insight about how the latest AI capabilities will upend higher ed. It is instead a slow, nuanced, deliberative account of what we’re trying to teach and why. It’s the secure base from which we can confidently explore how, if at all, AI fits with who we are and what we do.
I’m very lucky to have my first sabbatical this Fall, and in the rest of this essay, I’ll share how I’m going to be re-thinking my approach to teaching philosophy at a large, public Canadian university. Your mileage may vary.
Teaching OOD
The first component of a positive vision is a compelling answer to the question: What can students get from my class that they can’t get from AI?
It is well-known that LLMs degrade out-of-distribution (OOD): novel problems, genuinely local knowledge, fine-grained particulars, and embodied and institutional knowledge. So, one guiding ideal is to teach out-of-distribution, as much as possible.
Another way to think about it is the difference between artisanal goods and commodities. The main reason I’m willing to pay more for farmer’s market heirloom tomatoes is that they taste better. But I also like hearing from growers about micro-climates in the field and how this year’s crop compares to last. The human story and connection constitute the value. Contrast this with USDA-graded supermarket tomatoes. They’re mealy, fungible commodities abstracted away from any particular plant or field.
So, another idea is to offload the commodity-content to LLMs and focus class time and energy on the artisanal good-equivalents. Textbook accounts of consequentialism are commodities. Have the students chat to (carefully prompted) LLMs about it. That way, they can ask for examples and clarifications relevant to them. Then, class time is about my experiences using consequentialist arguments for animal welfare at the Thanksgiving table with in-laws.
New Affordances
A positive answer to the question: “what can students get here that they can’t from AI?” also needs to go beyond existing teaching and learning practices. We need to ask, what can I do now with AI that I couldn’t do before? What are the new affordances?
Andy Hall’s student-led evals project is one of the best answers I’ve seen to this question. Though not as impressive, here’s one example I trialled last semester with an AI-forward graduate student.
Say I allocate ~45 minutes per student for feedback on term papers. One thing I can do now is jot down very rough, directional comments on the paper. And instead of spending ~15 minutes polishing and editing those comments, I can now send the paper with my comments to an LLM and ask for two things: First, a polished version of my comments. Second, critical feedback on my comments, with an emphasis on my idiosyncrasies and blind spots, plus novel suggestions and counterarguments. Then, with the time I didn’t spend polishing, I can write meta-comments on the AI’s feedback. Now, the student gets three layers of feedback, plus a model of philosophical exchange.
If our duty as faculty advisors is to give students the best feedback we can, then it’s probably wrong to NOT use methods like this. Indeed, I’m quite certain that, at least for now, Gus + Claude > Gus or Claude. Not using AI has real costs.
Judgment All The Way Down
Careful readers might have noticed a tension. Teaching OOD seems to suggest a retreat: “find what AI can’t do, and plant your flag there”. New affordances seems to suggest an advance: “pick up the tool and do more”. Which is it?
Like most questions worth asking, the answer is, “it depends”. And what it depends on is what Aristotle called phronesis, variously translated as practical wisdom or judgment. It is the opposite of applying rules. It’s the capacity to see what a particular situation demands once the rules have run out. Thinkers from Aristotle to Autor have recognized that this kind of judgment resists codification and commodification.
“No AI in my classroom” and “unfettered AI access” both fail as policies because they substitute a rule for judgment. Abandoning take-home essays made the same mistake and it cost our majors their writing samples. I’ve since tied take-home writing to in-class exams, but (surprise!) it is imperfect and involves trade-offs.
When AI makes prediction cheap, the value of its complements - data and judgment - rises. When intelligence is abundant, judgment is what’s scarce. Cultivating judgment is arguably the most important goal for education in the age of AI.
We need to be honest with ourselves, and with our students. Using LLMs involves trade-offs. They tend to erode the very skills that are needed to use them well. But not using LLMs also involves-trade-offs. They are the greatest tool ever invented for learning. They are also the greatest tool ever invented for not learning. There’s nothing in the technology itself that decides which wins out. We do. One judgment at a time.
While I don’t disagree with your advice per se, I think it’s wrong to look at this through the lens of individuals. We need to set up institutions and norms, and have companies design AI products, that channel AI in productive ways and minimize its downsides. Individual decision-making won’t work at scale.
Compare with social media. Yes, there are ways to use it that help with networking and being exposed to new ideas and experts. But we know most people who use it, even fairly sophisticated and intelligent users, spend a ton of time getting sucked into the drama of the day, engagement bait, short-form video binges, and so on. This is because the social media apps are designed to maximize engagement first and foremost.