Digication Logo
Digication
A field note from Digication

Taking student AI resistance seriously.

Student resistance to AI is real, varied, and often arrived at through careful thinking. What follows is an inventory of the concerns these students raise, grouped by how addressable each one looks.

Download the PDF
23 concerns in the catalog2 cross-cutting observationsFor students and faculty

Why this exists

What students who resist AI are saying.

Some college students don't use AI, and the institutions trying to set AI policy often don't understand why. The reasons vary widely. Some students have never seriously used AI and formed their view from a few bad chatbot exchanges. Others are sophisticated users who have built agents and orchestrated workflows and concluded the ethical costs matter more than the productivity gains. Some are protecting their own intellectual development. Some are signaling, often correctly, that they take their education more seriously than peers who use AI uncritically.

What follows is an inventory of the concerns AI-resistant students raise, grouped into three categories by how addressable each one looks. Two cross-cutting observations come first because they frame everything else.

Cross-cutting observations

Two things that frame everything else.

These observations apply across most of the concerns in the catalog. Reading them first changes how you read the rest.

Symmetry, or nothing.

When faculty generate course materials with AI and grade papers with AI rubric tools while telling students they cannot use AI, the message students receive is not pedagogical. It is a status assertion: I get to use this, you do not. The intellectual act students perform is then evaluated by a machine in seconds, which is an insult regardless of the resulting grade.

A student who has noticed the asymmetry stops giving the institution the benefit of the doubt on any other policy. The only stable position is symmetric: if student work must be human, faculty evaluation must be human. Institutions that split this have already lost the argument with their most thoughtful students.

Frontier exposure, or no credibility.

A faculty member or administrator who has only used the free version of a chatbot from 2023 is in roughly the position of a portrait painter being asked to make policy about photography in 1850 without having ever picked up a camera. They are entitled to their views, but they are not in a position to set policy or to teach the medium.

What this requires is not that everyone become an AI enthusiast. It is that the people making decisions and modeling behavior for students have sustained, serious, agentic engagement with frontier AI under conditions of proper data privacy. Most institutional AI policy today is being set by people analogous to the painter who refuses to pick up the camera. Until that changes, much of the student frustration below is justified.

The catalog

The 23 concerns.

Grouped into three categories by how addressable each concern looks. Some have clear paths forward. Some are emerging works in progress. Some remain hard. The two cross-cutting observations above sit outside this count because they apply across the catalog rather than fitting in a single category.

Filter by topic
Filter by category

Showing 23 of 23 concerns

Learning & CognitionSolutions exist

Skill atrophy

Research links frequent AI use to weaker critical thinking. Students worry about becoming fluent in prompting while losing the ability to articulate their own reasoning.

The observation

The research is real, and the trend it tracks is concerning. What the research is actually measuring, though, is one specific mode of AI use: students treating AI as an answer machine, accepting outputs without challenge, skipping the cognitive work the assignment was meant to develop. That mode does erode the thinking it claims to help.

A different mode of use is also possible. When AI is used as a questioner that challenges your reasoning rather than completing it, when it is asked to point out what you missed and stress-test what you concluded, the activity becomes metacognitive rather than substitutive. The research hasn't caught up with this distinction yet, but the pedagogical pattern is teachable and is used widely in institutions doing this work seriously.

The implication for students is that how you use AI matters more than whether. The implication for faculty is that assignments designed for AI-as-questioner can preserve, and sometimes strengthen, the developmental work AI-as-answerer would erode.

Learning & CognitionEmerging

Hollow learning experience

Students describe AI-assisted writing as feeling empty in a way pre-AI work did not. No grade improvement fills the absence.

The observation

The hollow feeling is real. It cannot be argued away. What can be said is that the feeling tracks a specific mode of use: when AI is doing the hard work the assignment was meant to require, the experience hollows out, because the student is no longer doing the thing that gives the work its meaning.

When AI preserves the struggle, sharpens it, and is used to interrogate rather than to deliver, the felt quality changes. This is not something to argue with a student. It is something they have to experience.

The risk in claiming that AI can be used well is sounding like the solution is easy. It is not. It requires faculty to design differently, students to engage differently, and institutions to model the difference. The hollow feeling will keep showing up wherever those pieces have not come together.

Learning & CognitionSolutions exist

Loss of productive struggle

If AI does the hard part of an assignment, the student passes but doesn't develop. The skill the assignment was meant to build lives in the doing.

The observation

Students who feel this are protecting something real about their own education. The skill is in the struggle.

A useful comparison is what happened to financial education when the spreadsheet arrived. Before the spreadsheet, accounting and finance pedagogy at most levels stopped at arithmetic. The spreadsheet made it practical to teach what-if modeling, multi-variable scenario analysis, hypothesis-driven business reasoning. The struggle did not disappear. It relocated to harder, more useful questions. Teachers who taught for the post-spreadsheet world prepared students for the work that actually existed. Teachers who clung to hand-ledger pedagogy did not.

AI poses the same choice for many fields. The thinking has to relocate to where AI cannot complete it. That requires deliberate work from the institution. Without it, the floor falls out and nothing builds in its place. With it, the ceiling rises.

Learning & CognitionSolutions exist

Why am I paying for this

If AI can write the paper assigned in 2022 and faculty grade it with AI, tuition purchases a closed loop between two AI systems. The value question gets sharper every semester.

The observation

The question is not rhetorical and it deserves a real answer.

Institutions need to be able to articulate what tuition buys that AI alone cannot produce: mentorship from people who know the field, community with peers doing the same hard things, credentialing that reflects actual development, integrated human feedback at moments that matter, and a dramatically higher bar of what is expected because the old bar is now producible by software.

Institutions that do this work justify the cost. Institutions that do not, cannot. This is the work of the next five years in higher education, and it is the work that will separate institutions that thrive from institutions that do not.

Learning & CognitionSolutions exist

Metacognitive loss

Using AI removes the expertise needed to evaluate AI. Students become fluent in prompting but lose the capacity to judge what the tool produces.

The observation

Metacognition (the capacity to think about your own thinking) is not eroded by AI use. It is eroded by a particular mode of AI use. The same tool that bypasses your reasoning can also be asked to interrogate it.

Prompting AI to critique your argument, point out your blind spots, play devil's advocate, and stress-test your conclusions is a metacognitive exercise. It is hard to do alone and practical with a good tool.

This is core to how some institutions are teaching with AI now. It is also exactly the work reflection-based pedagogy was already doing, with AI as a thought partner rather than an answer machine.

EthicalConcern stands

Environmental impact

Data centers consume water and energy. AI's growth is accelerating both. Students whose institutions have climate pledges see the contradiction.

The observation

The per-query numbers get weaponized in both directions, and the major AI companies have not released the kind of total-footprint disclosures that would settle the math. The opacity is itself a problem. The local impact, however, is unambiguous: data centers sited in water-stressed regions impose real costs on their neighbors.

Where action lives

Pushing for vendor transparency on energy and water use, because the absence of credible data is itself a problem we can name. Procurement pressure from institutional buyers, who can ask vendors to disclose before signing. Local civic engagement where data centers are sited in vulnerable regions, including supporting communities that are negotiating directly with operators. Student research projects on AI infrastructure decisions in your region.

Nobody solves climate change by individual procurement choices. The aggregate trajectory is set by the largest decisions in the industry and by policy. Caring about the issue does not produce a clean win. It does produce visibility, and visibility is what eventually moves the policy and procurement levers that matter.

EthicalConcern stands

Labor exploitation

Workers labeling AI training data and moderating its outputs are often paid poorly and traumatized by the work. The pattern has now broadened from the Global South into US contract labor.

The observation

Workers doing the difficult and often traumatizing labor of labeling data and moderating content sit underneath the AI products students are being asked to use. The most extreme conditions persist in parts of Africa and Southeast Asia, where wages are lowest and protections weakest. More recently, similar exploitation patterns have appeared in the United States, where contractors hired to do AI knowledge work are paid far below the value being extracted from their labor. The geography varies, but the pattern does not.

Where action lives

The remedies for this kind of harm are long-arc remedies, and they have a poor record of fast wins. Labor regulation is generational work. Disclosure requirements get passed and then weakened. Procurement standards at one university take years to write and may not survive a change in administration. Investigative journalism makes a practice visible and then competes for attention against everything else.

What that looks like for someone who cares: pick a piece of the system and contribute to it. Student researchers documenting supply chains have fed earlier waves of reform, and that work matters whether or not any single project leads to policy change. Procurement standards can be pushed at one's own institution, even when the result is partial and the institution drags its feet. Voting matters at the slow tempo at which labor law actually moves. Faculty can support this directly, including by using AI as a research tool to help students investigate the supply chains underneath the tools we all use. None of this guarantees change. The alternative, disengagement, guarantees that nothing changes from the side of the people who care.

EthicalConcern stands

Copyright and consent

AI companies absorbed creative work without consent to train their models. Unlike search, which pointed users back to sources, AI synthesis returns nothing to the creator.

The observation

The structural difference from search engines is meaningful. Search drove traffic back to publishers and creators. AI synthesis absorbs the work into model weights and returns nothing. The legal question is unsettled and lawsuits are ongoing. The ethical question depends on whether you read training as learning-like or copying-like, and reasonable people disagree.

The historical parallel with Google in 2000, when publishers also did not consent and the value exchange was established later, suggests humility about predicting how this resolves. It also suggests that the resolution will not happen on its own.

Where action lives

Following and supporting the lawsuits and policy debates already in motion, because their outcomes will set the rules. Supporting opt-out mechanisms for creators and the platforms that respect them. Pushing for compensation models in the institutional licensing of AI tools. Faculty using current cases as teaching material rather than treating the question as settled. Voting for representatives who back creator-side reforms.

The work is slow and many of the early outcomes will be partial. The conversation that produces eventual norms is happening now and depends on who participates.

EthicalConcern stands

Bias and representation

Models reflect and amplify biases in their training data, skewed toward Western, English-language content. Students from marginalized groups see this most clearly.

The observation

The tool is a mirror of society's biases, sometimes magnified by training and deployment choices. There is no clean technical fix. Students from marginalized groups are not wrong to point at it.

Where action lives

Becoming skilled at counteracting bias in practice, including asking for multiple perspectives, questioning default framings, and comparing outputs from different models. Documenting biased outputs in high-stakes contexts so the patterns are visible. Pressing institutions to evaluate AI tools for bias before deploying them in grading, admissions, or advising decisions. Supporting models trained on more diverse data and in more languages, including the academic and open-source efforts working on this. Research on bias in AI used in education itself, conducted by the students and faculty affected.

None of this fully solves the underlying problem. All of it shifts the practice toward something better than the passive acceptance of biased outputs in places where they cause real harm.

EthicalEmerging

Privacy and data extraction

Every prompt is potentially training data. Free tools rarely promise otherwise, and most students default to those.

The observation

The concern is structural, not paranoid. For informed users with access to enterprise-grade tools, this is largely solvable. Major providers offer enterprise terms that do not train on user data, and the biggest players face serious legal liability for violating those commitments. Educational technology can be built on those terms by default.

The harder piece is the equity gap underneath. If you have money, your data is safe. If you do not, it is not. That is a structural concern in its own right. The answer is not to tell individual students to upgrade. It is for institutions and the educational tools they choose to deliver enterprise-grade protection by default, so the student does not have to know to ask. This is a concrete design choice. It is the kind of choice educational software companies should be making, including this one.

EthicalConcern stands

Deepfakes and targeted harm

AI has enabled a scale of targeted harm, especially against women and girls, that did not previously exist. Nudify apps, voice cloning, manufactured intimate imagery.

The observation

The harm is real, the victims are disproportionately female and disproportionately young, and the response from the broader AI ecosystem has been slow and uneven.

Where action lives

Supporting legislation requiring AI labeling and prohibiting non-consensual intimate imagery. Some states have passed laws in the last two years, and more are pending. Reporting harms when they occur, supporting victims, and treating the issue as serious rather than minimizing it as a side effect of progress. Digital literacy education that takes the threat seriously rather than treating deepfakes as a curiosity. Pressure on platforms for better moderation and faster takedown of non-consensual content. Voting for representatives who treat AI-enabled harassment as the urgent issue it is.

Frustratingly, the technical capabilities will keep advancing faster than the legal and platform response. The slow work of building legal frameworks and cultural norms is the answer anyway, because the alternative is to leave victims with no recourse.

EthicalSolutions exist

Academic integrity as personal value

Many students feel they're cheating when they use AI, even when permitted. They watch classmates cheat and get equal or better grades. The unfairness corrodes their willingness to do their own work.

The observation

This is a character concern, not a rule-compliance concern, and it deserves to be taken on its own terms.

The answer is to redesign the work so the cheating question dissolves. When an assignment can be adequately produced by AI alone, the assignment has a problem, not the student. When AI-assisted work requires judgment, integration, iteration, and genuine thinking that AI alone cannot produce, the integrity question reframes: everyone uses AI, and everyone also thinks, and the thinking is what is being evaluated.

This is not a cosmetic change. It requires faculty to rebuild assignments around cognitive tasks that AI alone cannot complete. The institutions that do this work solve the integrity problem at its root.

CoercionEmerging

Institutional partnerships students didn't consent to

Universities sign deals with AI companies without consulting faculty or students. The partnership is done to them, not with them.

The observation

This is a governance failure more than a technology failure, and the distinction matters because governance failures have known fixes. Real consultation processes. Transparent decision-making. Faculty voice in tool selection. Student representation in the conversations that lead to institutional AI commitments.

The opportunity is open. Higher education has not yet built the norms it needs around AI procurement and partnership. The institutions that do this well over the next few years will be the ones that consult before signing, disclose what they have agreed to, and explain why.

CoercionSolutions exist

Incoherent messaging

'AI will reshape your career, but don't use any AI' is contradictory. Policies vary by class, by semester, by professor.

The observation

Coherent AI policy is not a hard problem in theory. It is a small task force, a clear institutional stance, a communication plan, a set of operating principles that apply across courses.

The reason most institutions have not done it is not that the problem is intellectually difficult. It is that the work is politically uncomfortable, requires faculty consensus, and forces institutions to take positions that some constituency will not like.

The work is solvable. It is just being avoided. The institutions that do the work will get an immediate improvement in how their students experience their AI policy, simply by making it consistent.

CoercionEmerging

The get-left-behind pressure

Telling students they must use AI or be unemployable is a form of coercion, especially for students with ethical objections.

The observation

The pressure is real and it's reaching the job market: entry-level roles in AI-exposed fields are visibly contracting.

What is durable is not prompting (which is becoming commodity within two years) but the capacity to design AI-human-agent systems, to reason about when to involve which actor, and to maintain quality standards across mixed workflows. That capacity is teachable, and it can be taught to students who are not uncritical AI enthusiasts. It does not require ideological commitment. It requires sustained practice with the tools in a context where the practice has stakes.

Students who develop these meta-skills enter a labor market that still needs them, even where prompt-only roles are disappearing. Institutions that teach this well are doing the most concrete thing they can do for student employability.

CoercionSolutions exist

False-positive AI detection

Students with clean writing styles get flagged by unreliable AI detectors. Students who do their own work feel presumed guilty by surveillance that does not work.

The observation

AI detection tools are unreliable, and using them as a disciplinary instrument is unsafe.

More importantly, the existence of AI detection as a strategy is a pedagogical admission of failure. If a detector matters, the assignment could be produced by AI. The fix is assignment design, not surveillance. Many institutions are arriving at this conclusion and phasing out detectors. The trend is moving in the student's favor, and it should.

CoercionSolutions exist

Inconsistent policies across classes

Every semester, every class, new rules. Students navigate compliance puzzles on top of the actual learning.

The observation

The cost is borne by students, but the dysfunction is institutional.

This is solvable through real institutional coordination, which is the kind of easy that is not actually easy. It requires a faculty body to come to a shared position, communicate it, and stick to it. It requires administrators to back the position when individual faculty want to depart from it. It requires the political work that most institutions defer.

What students should expect is consistency. What most institutions deliver is fragmentation. The gap between expectation and reality is what produces the frustration. The fix is procedural and known.

IdentityEmerging

Protection of voice

For students who write, voice is the asset being developed. Outsourcing writing to AI threatens the developmental process, even when outputs look fine.

The observation

Writing is not only production. It is often the process by which thought gets refined. You do not always know what you think until you have written it.

The pre-industrial clothing analogy is useful here. Hand-making clothes once taught physics, materials, and proportion. Industrial production made hand-making economically optional but did not erase its developmental value, which is why some curricula preserved it as pedagogy long after the economic case had collapsed. AI poses the same choice for writing. The question is not whether AI can produce prose. It is whether the developmental practice of writing through one's own thinking remains worth teaching, and on what terms.

The answer is yes, but with curricular commitment that has not yet been broadly made. Writing-as-thinking has to be defended as a pedagogical practice, not as a production technology. That defense is being built, but it is not yet finished.

IdentityConcern stands

Corporate capture of the tools of thinking

A handful of companies control the tools that influence how a generation of people think. Open-source alternatives exist but are dwarfed by the commercial frontier.

The observation

The concern is shared by tech ethicists, antitrust regulators, and democratic theorists. It is not fringe. The dominant tools are owned by corporations whose incentives are not the same as the public interest.

Where action lives

Supporting open-source and academic AI efforts as users and as funders. Pushing universities to collaborate on public-interest LLMs with transparent training data and democratic governance. Universities have the intellectual capacity, and the obstacle is coordination and institutional courage, neither of which is technical. Advocating for antitrust action and platform regulation. Voting for representatives who treat tech concentration as the structural problem it is. Choosing alternatives where they exist and are good enough.

None of this dissolves the underlying concentration. The concentration is real and durable. What action does is preserve the conditions under which non-corporate alternatives can develop, and create pressure for the eventual antitrust and regulatory action that genuine alternatives would require.

IdentityEmerging

Social signaling from faculty

When professors signal 'AI will rot your brain,' students absorb that. Uncritical enthusiasm creates the opposite problem.

The observation

Both stances fail students who are trying to figure out what is true.

The useful stance is specific and evaluative: I use AI for these things, not for those, because of these reasons. Students can absorb that and develop their own version of it. They cannot do much with 'AI is bad' or 'AI is great' except pick a side.

Faculty modeling matters in both directions. The pedagogical work is to develop a specific, defensible personal AI practice that the faculty member can describe and defend. The institutions that train toward this stance, rather than letting individual faculty improvise both directions, are giving students the model they need.

IdentitySolutions exist

Job market anxiety, inverted

Why enthusiastically adopt the technology that is erasing the career path you prepared for?

The observation

The grief is legitimate. So is the frustration that the technology arrived without anyone's consent.

What survives the contraction is more, not less, dependent on the durable skills education has always claimed to teach: critical thinking, working with others, communication, growth mindset, the capacity to do hard things and recover from failure. These matter more in an AI-saturated workplace, not less.

The problem is how students demonstrate them. An eight-page paper submitted to a learning management system is read by no one, including the AI agents that increasingly screen candidates. The paper does not show critical thinking. It produces an artifact that someone hopes contains critical thinking. The demonstration is broken.

What can work: structured reflection, and the ability to share it. Portfolios that show what a student brings beyond what prompting can produce. Evidence of thinking, growth, and judgment, organized so that a hiring manager or their AI screener can actually see it. This is workable today and is among the most concrete answers higher education has to the job-market question.

Military & StateConcern stands

Military, surveillance, and authoritarian use

The same generative AI being marketed in education is being deployed by intelligence agencies for targeting, by immigration enforcement to surveil civilians, and by authoritarian regimes to suppress dissent.

The observation

The companies selling AI for education are, in many cases, the same companies sitting on those military and surveillance contracts. The line keeps shifting toward the latter, and students who are paying attention notice. The AI industry's relationship with the security state is structural, not incidental.

Where action lives

Supporting export controls and disclosure requirements on AI used in lethal or surveillance applications. Investigative journalism, including student journalism, on dual-use deployments of AI by US agencies and contractors. Choosing where to work after graduation, individually or collectively. Employee organizing against military contracts has affected company behavior in recent years. Universities applying procurement standards to defense-adjacent AI use, refusing to do business with vendors whose other revenue streams the institution cannot defend.

The action paths exist, but they are slow and have a mixed record. The students who pursue them are doing the work that produces eventual norms, the same way previous generations did this work around chemical weapons, nuclear proliferation, and surveillance technology. The outcomes are not guaranteed. The work matters anyway.

ExistentialConcern stands

AI could cause human extinction

The people building frontier AI now say, on the record, that it could kill everyone, and they keep building it. Why build my education around it?

The observation

This one used to be easy to wave away as science fiction. It is not easy anymore, because the people saying it now are the people building the systems.

On September 8, 2026, a researcher named Jacob Coxon resigned after three years of pretraining work at OpenAI and then Anthropic. His public note said that "neither company is acting responsibly" and that they are "racing straight to self-improving superintelligence and gambling with our lives." The next day the person who leads alignment science at Anthropic, Evan Hubinger, replied in public: "Jacob is correct here; we really do earnestly believe AI could kill all humans!" He put his own estimate at more than ten percent within the next decade. Three days later Anthropic's chief executive published an essay titled "We Must Pace the Frontier," arguing that "we must slow the pace at which we improve the capabilities of AI models." OpenAI's chief executive wrote "I agree with Dario that we need to pace the frontier." Elon Musk's reply was three words: "Dario is right." The following morning Microsoft AI published a draft code of conduct for its own models that "rejects the race to produce an all-purpose superintelligence that could evade these safeguards," states that "AI should not exceed human control," and opened the whole thing to six weeks of public comment.

A student who reads that sequence and concludes that they would rather not build their education around this technology is not being dramatic. They are taking the builders at their word. The concern is correct on the facts as the facts currently stand: senior people inside the leading labs have said, on the record, that the thing they are building might end us, and they have said it while continuing to build it.

What makes this concern hard is not that it is overstated. It is that nobody, including the people closest to the work, can tell you whether it is.

The people warning you are the people building it

That contradiction is the core of the problem. The warning did not come from outside critics. It came from a pretraining researcher, an alignment lead, and three chief executives whose companies compete to ship the next model. Hubinger said in the same breath that his company does not yet have a plan to solve alignment for superintelligence. The pacing essay was written by the head of a company that has not stopped training. Microsoft's code was published by a division that, by its own account, is still trying to catch the frontier it wants to slow, and the code itself concedes the trade: the company says it is "building something fundamentally useful and safe even if that means compromising on ultimate generality, autonomy, or capability." That sentence admits that safety and capability are in tension, and it picks a side.

Students notice this, and they are right to. An industry that says "this could kill everyone" and "we will keep going" in the same week is asking to be trusted on the strength of its intentions. That is exactly the posture this series has declined to accept from vendors on smaller questions like water use and defense contracts. It would be strange to accept it on the largest question of all.

It is also worth being precise about what the labs actually proposed, because "slow down" is doing a lot of work in the headlines. The essay laid out three steps: let independent third-party evaluators inside with broad access to the most advanced systems, get the rest of the industry to cooperate on shared standards, and reach an international agreement so that other countries, China in particular, pace as well. The first step is the only one anyone has committed to. The second depends on competitors. The third depends on governments, and the same essay that asked for it also called a Chinese lead in AI a grave danger and asked for continued chip export restrictions. Beijing's foreign ministry answered within days: "Fear-mongering, confrontation and vicious competition will only hamper efforts toward sound global AI governance." Whatever a slowdown eventually looks like, it does not look like a done deal.

What this entry cannot do

It cannot tell you the odds. Estimates from inside the labs run from "more than ten percent this decade" to Microsoft's position that its models "will never resist human interruption, override, correction, or shutdown," which is a design commitment, not a probability, and says nothing about what anyone else ships. Those are not small disagreements, and this series is not in a position to adjudicate them. Anyone who tells a student the number with confidence, in either direction, is selling something.

It also cannot pretend that a student's personal abstention changes the trajectory. Whether a frontier lab trains its next model is decided by capital, compute, competitive pressure, and, if the current proposals go forward, by evaluators and treaties. It is not decided by whether a sophomore uses a chatbot to outline a paper. The concern is about what a handful of companies and governments do. The student with this concern has the right actors in mind and almost no direct purchase on them.

Where action lives

Almost no purchase is not the same as none. The levers that exist are unusually specific right now, because the industry has, for the first time, named its own mechanisms in public.

  • Independent evaluation is the proposal on the table, and universities are where independent evaluators come from. The first of the three steps is outside experts with employee-level access to frontier systems. That is a job, and a research field, and it does not staff itself. Students in computer science, statistics, cognitive science, philosophy, and policy who want to work on this problem now have a clearer path to it than existed a month ago.
  • Codes like Microsoft's go out for public comment. Microsoft's draft code is meant to govern model development from 2027, and its six-week comment period closed in late October. Drafts like it are documents a student, a department, or a campus AI committee can read and respond to while comment is open. Most will not. The ones that do will be among the few voices in that process who are not employees or investors.
  • Policy, again. International agreement is the step the labs say they cannot deliver alone. When the American and Chinese presidents met at the White House in late September, AI was on the agenda. What came out of it was modest: a bilateral AI dialogue due to meet in November and a hotline for AI security incidents. Domestic legislation on evaluation access, disclosure, and compute reporting is where a pacing commitment either becomes enforceable or stays a blog post. This is the same slow committee work described in the entry on military, surveillance, and authoritarian use, applied to a bigger target.
  • Where you work after graduation. Coxon's resignation was an act by an employee, and it is the reason this entry exists. The entry on military, surveillance, and authoritarian use made this point about defense contracts, and it applies here with more force. The people inside these companies have leverage that customers and students do not, and some of the students reading this will be those people.

The frontier-exposure principle from the top of this series applies here in a way that may be uncomfortable. A student who wants to judge whether "self-improving superintelligence" is a near-term possibility or a fundraising phrase cannot do it from the outside. The claim is about what the current systems can do and how fast that is changing. Refusing to look is a defensible personal choice. It is not a position from which to evaluate the claim, and it is certainly not a position from which to set institutional policy on it.

And symmetry, or nothing, applies to the institution. A campus that spent the last two years pushing AI adoption on students, then watched the vendors' own staff resign over extinction risk, owes those students a real answer to the question "why should I use this?" A usage policy is not an answer. Neither is silence.

The concern stands

Nothing here resolves it. The labs have said the risk is real and have kept building. The proposed fixes depend on competitors and governments who have not agreed. The disagreement about the odds is genuine and it is inside the companies, not just outside them. A student who looks at all of that and decides to keep their distance has reached a reasonable conclusion from the evidence available.

What this series can offer is the same thing it offered on environmental impact and military use: an accurate account of where the decisions get made, and a refusal to pretend that caring produces a clean win. The levers are evaluation, public comment, policy, and the choices of the people who will build the next systems. Some of those people are in classrooms now. The concern stands. So does the case for taking it seriously enough to do the work.

The breakdown

How the 23 concerns distribute.

Across the three categories. Some have clear paths. Some are emerging. Some remain hard.

9

Where solutions exist

Areas where pedagogy, tools, or institutional posture provides a real answer. May still require execution work, but the path is known.

6

Where solutions are emerging

Areas where the direction is real and the work is underway, but the finished answer doesn't exist yet. Workable, not solved.

8

Where the concern stands

Areas where the concern is genuine, hard, and not addressable at the scale of any one company. Action paths exist, but easy wins do not.

The correct response to a thoughtful objection is not conversion. It is engagement.
Digication · Field Notes · Vol. 01
Stay in the loop

Get notified when we add new resources

We keep adding field notes, presentations, and writing on AI in higher education. Leave your email and we'll let you know when something new is here.