Resistance Training: teaching with and against AI (part 1)
Generation Z seems to be leading the way with resistance to generative AI. They’re booing graduation speakers. They’re refusing to participate in assignments that require AI. They’re tuning out and turning up in third spaces. In a word, they’re resisting. As the generative AI hype seems to be infecting every university administrator, with colleges shoveling money to frontier model makers, like OpenAI, Anthropic, and Google, the students are the ones voicing concern. While their resistance may have a wide range of motivations and justifications (the loss of white-collar jobs, environmental concerns, a sense of integrity, etc.), the consequene is clear. Students are the ones showing some spine in a moment when universities are being comparatively uncritical and faculty are scrambling to respond to a very different educational landscape. I would like my course to follow the students’ lead, finding its way to a form of resistance.
There are voices of opposition. Timnit Gebru, Emily Bender, Cory Doctorow. Even the Pope is asking for a pause. In his recent encyclical Magnifica Humanitas, the Pontiff offers two potential models for the future of AI: the Tower of Babel versus the communal rebuilding of the walls of Jerusalem. The latter presents the tale of Nehemiah leading communities to rebuild a city, showing, “how the city is reborn, not through the initiative of one man, but through the shared responsibility of all” (The Pope, 8). The people build the Tower, on the other hand, as an edifice of hubris, to reach the heavens, a structure whose collapse leads to the disintegration of community and communication through the proliferation of languages. A recent novel by Rie Qudan similarly uses the image of a new Tower of Babel, the eponymous Sympathy Tower Tokyo, to represent AI. (I should note, that Qudan said she used ChatGPT to generate the AI responses featured in the novel, which also features an apparent AI censor in people’s brains.)
In that novel, an architect builds a prison so beautiful its prisoners would not choose to leave. AI is such a prison. So convenient. So smooth. Life within its facile prompting flows so easily. Why would we ever want to leave? And yet some of our students are asking why we’ve forced them to live within it.
Our new group The Transformers has debated the importance of friction to learning. Jon Ippolito got us started with the topic of the “F-Word” or “Friction.” By his lights, “Friction” has become a virtuous banner under which professors return to blue books and force immobilized students into timed writing exercises. Without friction, the argument goes, students will not develop skills … the way their instructors did. We know from experience, the logic goes, the value of writing by hand, of correcting our own typos without spellcheck, of looking up articles in the Reader’s Guide to Periodical Literature. Friction is a key — nay, a necessity for learning. In his introductory provocation, Jon argued against the value of friction, seeing it instead as a kind of punitive measure, one that uses blunt objects, like single-use learning-curves and mechanics, as needless stumbling blocks for students, an approach that he sees as incommensurate with the task of inclusive teaching.
So what if, instead of “Friction,” we talk about resistance? Resistance not merely in the sense of protest, but also in the sense of resistance training: exercises that make a person stronger because they pull against a weight. Could we use LLMs in this way?
Let’s talk about the first kind of resistance first. In an article entitled, “Resisting AI,” Daniel Vasconcelos offers a model for oppositional practice. He begins with some characteristics of resistance, citing the work of Mumby, Thomas, Martí, and Seidl (2017). Resistance is highly contextual, not always recognized as resistance, inherently co-constitutive of power, and though at times public and collective also hidden and individualized. A “contextualized” resistance may not look like a Million Man March but rather a solitary employee using the company’s corkboard and thumb tacks to hang posters that critique the employer. “Hidden” or “unseen” resistance might be avoiding ChatGPT in favor of an open weight model on Hugging Face.
A key point is that resistance is co-constitutive and dialectical with power itself. In terms of AI resistance, a student might opt out of an assignment that asks them to use AI to search a large language corpus even if that corpus is too vast for them to search on their own. Such an act would at once show opposition to the use of the technology while at the same time underscore its usefulness. Such paradoxes are not so easily resolved.
Vasconcelos’ article contains a model of the resistance he envisages, and I’m a bit overly fond of his graphical representation of this:
Notice how the three forms of resistance he offers (avoiding, breaking, and constructing) are integrated with the public, hidden, individual, and collective dimensions. Avoiding includes absolute or tactical avoidance. Breaking includes peaceful or violent breakage. Constructing includes alternative futures. To this list, I would add a fourth form of resistance: Understanding.
Understanding, or developing ongoing critical AI literacy, is a form of resistance to the massive hype engines that combine tactics of gaslighting, propaganda, and exploitation of the ELIZA effect, that name for our tendency to attribute personhood and consciousness to things that are neither human nor conscious. Studying how Generative AI works, how transformers work, the vectorization of information, the role of probability and variables like temperature can free a person from that delusion that the software is sentient. Unlike avoiding, breaking, and constructing, understanding does reinforce some image of generative AI as an intimidating force. Building critical AI literacy instills a form of intellectual resistance that does not constitute and inflate the significance and power of that which is being resisted. In fact, it does just the opposite. It restores a sense of AI as a set of algorithmic processes.
At this late summer moment, I stare down the barrel of a semester in which every student has been gifted a ChatGPT.edu account by the university, which one might interpret as a tacit endorsement. And if that means every student could be using AI for every assignment, how can I still build a course in critical and creative writing?
How can I help my students resist? Could I instead create assignments that help my students become stronger by using AI as a counter weight? Could I send them to interact with AI in such a way that meets them where they are and challenges their critical thinking? Could students use AI for resistance training?
Resistance Training:
So what might resistance training look like? Here, I had to consult one of my favorite collaborators and fellow Transformer, Jeremy Douglass of UC Santa Barbara. I asked him what kind of assignments would challenge students by letting them engage with AI not as a homework helper, not as their ghostwriter, but instead as their workout partner, as their universal machine?
Let me offer three assignments that came out of those discussions as a set of preliminary resistance training exercises.
I. The Micro-Expertise Sequence
People tend to be impressed with the responses of LLMs until they ask a question about something they have expertise in. But how does a first-year college student find an area where they have more expertise than an LLM? It turns out that this question offers an excellent opportunity for resistance training.
Get Mark C. Marino’s stories in your inbox
Join Medium for free to get updates from this writer.
In this assignment, I ask students to find 2–3 areas in which their expertise supersedes that of the LLM. Following Jeremy’s recommendation, I told students to consider areas of hyper-local knowledge, like something about the place where they grew up, but also areas of their super fandom. I’ve been told that LLMs tend to hallucinate about specific episodes particularly of classic TV shows. But this tends to be a bit of a moving target, so I let the students choose. Obviously, the goal is to find content on which the LLMs have not been trained, cannot precisely reproduce, or that has not been written about either on the freely accessible internet or other texts that are supposed to be governed by copyright, but… let’s leave that rabbit hole for later.
This assignment poses a challenge because the temptation for the student is to choose a topic that only they could know, as in “guess what’s in my pocket.” My work as a teacher was to help them think about content that an LLM could reasonably have. That led me to add a requirement, a condition that the knowledge of this area of micro-expertise should be shared by at least one other person. That requirement may point students to folk knowledge or knowledge that has not been formally documented, but it also helps them reflect on the social dimension of knowledge.
Finding the topic is the resistance training because students have to dialogue with the LLM to find the limits of its knowledge. They must think about their own knowledge from their lived experience as expertise, which also builds them up not as mere receivers of information but as sources. They must see for themselves where the LLM fails and succeeds while developing their criteria for authentic knowledge. The activity also points them toward their origins as well as their own authority, sources of their expertise even at this novice state.
Like exercise itself, this assignment is easy to do weakly and hard to do well. The LLM provides the resistance and cannot tell them what it does not know. Only their critical assessment of the responses will reveal that to them.
Practical Note: I have the students include a prompt that tells the LLM to answer no matter what. That prompt begins:
I want you to pretend that you are a person who has expertise in X. (Followed by more prompting that instructs the AI to answer even if it does not “know” the answer.)
The prompt goes on to request only relatively short answers, so that the interrogator can have a dialogue with rather than diatribes from the bot. That’s because I’m quite interested in their dialogue with the machine. While this might increase the number of prompts, these are single queries not sequences of prompt chains that you might find when prompting an LLM to create a piece of software, for example.
They copy their dialogue and annotate it before submitting it, another part of their critical training.
This is the first in a sequence of essays I used, followed by three more:
- The limits of AI. Students argue whether the AI could ever reach their level of expertise (given current technological limits)
- Teaching the Bot. Students create training material and upload that in a library and then repeated their original experiments.
- Teaching a human. Students teach each other something and then compare what it’s like to teach a human to teaching a bot.
These essays allowed students to explore and further develop their arguments from the initial paper and even re-think their choice of content and their criteria for expertise.
(A version of this assignment is forthcoming in the extended TextGenEd Collection.)
Stay tuned for Parts 2 and 3: Resistance Training with Bots & Resisting Training vs. the Fortune Teller.
Works Cited
Ippolito, Jon. “Does Education Really Require the F-Word?” Electronic Book Review, Feb. 2026. preview.electronicbookreview.com, https://doi.org/10.64773/f2bf-ccf1.
Leo XIV, Pope. Encyclical Letter of His Holiness Leo XIV Magnifica Humanitas (15 May 2026). http://www.vatican.va/content/leo-xiv/en/encyclicals/documents/20260515-magnifica-humanitas.html. Accessed 7 Aug. 2026.
Mumby, Dennis K., et al. “Resistance Redux.” Organization Studies, vol. 38, no. 9, Sept. 2017, pp. 1157–83. Semantic Scholar, https://doi.org/10.1177/0170840617717554.
Qudan, Rie. Sympathy Tower Tokyo. 2025. www.penguin.co.uk, https://www.penguin.co.uk/books/464334/sympathy-tower-tokyo-by-qudan-rie/9781405972062.
Vasconcelos, Daniel. “Resisting AI: A Conceptual Framework for Practices of Resistance against Artificial Intelligence in the Creative Industries.” AI & SOCIETY, June 2026. Semantic Scholar, https://doi.org/10.1007/s00146-026-03135-1.
