Let me be clear: there is an attention economy surrounding claims about AI’s transformative impact, which sometimes drowns out nuance and caution. Benchmarks never tell the full story, and there are well-founded objections even to the most popular ones. And there are profound risks — from extraction to economic upheavals, from safety to the concentration of power.

But if most benchmarks point in the same direction and if many serious — and often initially skeptical — people agree that something profoundly changed sometime around late 2025, you should perhaps at least pay attention—and dip your toes in it. Yet, large parts of academia still seem to cycle through the first two stages of grief, mixing outright denial with misdirected anger.

Much as the mirror provides affordances for vanity, LLMs provide affordances for cognitive lethargy. Like sugary snacks at the checkout, they exploit first-order desires at the expense of second-order ones (Frankfurt 1971). Students may want to learn but they also want to finish their assignments. Procrastinating or cutting corners does not require LLMs but is aided and abetted by them. Why use an ‘e-bike for the mind’ if you have a delivery app at your fingertips?

LLMs not only promote ‘metacognitive laziness’: a reduced tendency to plan, monitor, and evaluate one’s own thinking (Fan et al. 2025). They also erode the very thing that makes learning rewarding. Intellectual and personal growth, as Thomas Pfau reminds us, ‘can only ever be the fruit of sustained personal effort’. For education to be more than a ‘relentless series of logistical challenges’, students must experience the struggle and labor of learning.

Let’s not kid ourselves, the AI-enabled ‘democratization of cheating’ means that certain types of assignments are no longer viable—at least not without some extra effort. We need better detection and sanctioning, more oral and in-person assessment, and creativity in designing new assignment formats.¹ This is not because all students cheat or want to cheat—but because some do. When designing rules, we need to think of the knaves.

Yet a supposedly AI-proof constitution for knaves has two central problems:

• It ignores that assignments are not just assessment formats but learning vehicles. Making curricula cheat-proof by eliminating certain assignments is like draining the pool to prevent drowning.

• It ignores that designing around incentives only crowds out intrinsic motivation—distrust breeds distrust which undermines the disposition to learn (Bowles 2016).

In the myth, Damocles envied the king — until he saw the sword hanging above the throne. Instead of giving up on written or other hard-to-cheat-proof assignments, we should create similar Damocles effects: pair every submission with an accountability moment that hangs over the process like the mythological sword.

• A term paper followed by a colloquium where you present and defend your work to peers
• A coding assignment followed by an ‘explain your code’ session
• An essay followed by an oral exam on the same material

AI may be able to write a paper for you. But it can’t save you from embarrassing yourself if you have to explain, defend, and communicate it.

In the myth, Procrustes offered travellers a bed — then stretched or amputated them to fit it. Fixed semester lengths, standardized exam formats, rigid credit structures often do something similar: they force the learning process to fit inherited institutional furniture.

AI is like a forcing function that reveals how poorly learning actually fits these constraints.

• A colloquium after the term paper? That may require scheduling — and grading — beyond the semester.
• Combining written assignments with unannounced in-class tests or follow-ups? Keine Anwesenheitspflicht!
• Wanting students to collaborate across different seminars on a joint project that combines different approaches? Different modules, different grades, different semesters.

In the myth, the young Hercules meets two figures at a crossroads: Vice, who offers ease and pleasure, and Virtue, who offers a harder path leading to genuine accomplishment. With AI, we face this choice over and over.

The, well, Herculean task is to get students to want to learn and convince them (and ourselves) that most everything worth having is on the other side of hard — and to explore and provide the tools and ideas that make the hard choice easier.

To think that AI will, by itself, speed up scientific progress tout court is, as Narayanan and Kapoor put it, like ‘adding lanes to a highway when the slowdown is actually caused by a tollbooth.’

AI can undoubtedly make individual scientists more productive—even if it makes their research interests narrower (e.g. Hao et al. 2026). But more productive scientists will only lead to more scientific progress if we reform the institutions and norms of academic publishing. As of today, AI mostly acts as an ‘accelerant for the existing optimization machine, (…) offering a faster way to produce the artifacts the system rewards’—perpetuating and intensifying the existing reward structure, attention scarcity, and pressure toward safe, incremental work.

As it becomes easier and easier to produce decent — or at least decent-looking — papers, we need to start thinking about the epistemic and attentional externalities of ‘monstrous scientific overproduction’: each paper requires attention from editors, reviewers, and readers, and any further increase risks overwhelming scientists’ collective capacity for validation and digestion (for good starts, see here and Munger et al. 2026).

Like other forms of pollution, the answer must involve some form of tax on production, so that publishing only pays if a paper’s contribution justifies the burden it places on others. Whether that tax takes the form of stricter norms, submission caps, or incentives that reward depth over volume is an open question. But the direction seems clear: we need fewer and better papers—and maybe rethink the paper format altogether!

While AI disrupts and destabilizes the current publishing equilibrium, it can also help — if supported by the right institutional changes — to move us toward a new one of fewer and better papers. The key is to stop thinking of AI as a tool allowing you to produce more and start thinking of it as a collaborator allowing you to do better.

Claes Bäckman, for example, has set up a collection of Claude Code skills that helps you with everything from spelling and academic style to tables and figures, and even provides (journal-specific) adversarial reviews. Tools like this can dramatically speed up feedback and democratize access to expertise.

If set up right, AI can place you in rooms full of experts: theorists, methodologists, editors who are intimately familiar with your work and can not just write, code, and iterate through different analyses, but also provide pushback, ask follow-up questions, and surface connections you missed.

Where people walking see walls, handrails, and gaps, parkour athletes see vertical floors, pathways, spaces to be jumped. This is called ‘parkour vision’ — seeing potential for interaction and exploration where others see nothing at all. Jasmine Sun argues that there is an analogous thing called ‘software vision’ that allows programmers to see ‘software-shaped problems’ everywhere. Social scientists need something similar: the ability to see ‘research-shaped problems’.

As AI makes it easier to find answers, the most interesting ideas and insights will come from those who can ask the right questions — or ask them in the right way. This fundamentally requires the ability to recast old problems in new ways, reframe something obvious as puzzling (or the other way around), or ask what it would mean for something familiar if we assumed something unfamiliar about it (Peirce 1997). Mills (2000) called this the ‘sociological imagination’ — but we can also call it parkour vision for social scientists.

Just as AI makes execution less important, it makes deep expertise more important. AIs still frequently get things wrong, miss or misunderstand context, go down irrelevant rabbit holes, start from questionable assumptions, and get stuck in local optima. As Kevin Baker put it, ‘absurdity is not self-correcting’ — and it requires competent researchers to detect when something is either plain wrong or just doesn’t smell right.

Only having real expertise allows you to tell, as John List put it, the difference between ‘nearly right’ and ‘right’, which will be ‘more valuable than ever’. Expertise also allows you to use AI much more effectively, pointing it in the right directions, helping it avoid unproductive lines of reasoning, or providing essential context. Just like an experienced gamer can do a lot more with the same cheat code, or an experienced piano player much more with a music model, expert researchers benefit disproportionally from AI.

Frontier LLMs are surprisingly poor writers, in some ways worse than earlier, much less capable models. With their rough edges sanded down during post-training, their style is strangely disembodied, emotionally lobotomized, compulsively even-handed. This is not straightforward to fix as the very whimsicality that makes good writing may not be what you—or your customers—want from a model in other domains. In addition to writing bland prose, LLMs, perhaps due to their very architecture, also shrink linguistic diversity and amplify ‘certain sophisticated linguistic patterns while narrowing the range of individual expression’ (Sourati et al. 2025, 34).

Given that many people are already getting sick and tired of AI slop, having your own voice, your own distinct style may be a major competitive advantage: a breath of fresh air in a symbolic landscape suffocating from sameness, a well-recognized gift to a variety-starved world.

The great paradox we are facing is this: parkour vision, deep expertise, taste, a unique voice are all built through the labor of learning — tedious reading of technical texts, long days (and nights) with data, the productive suffering of writing, the accumulated ‘scar tissue’ from many trials and errors. You need to learn about different methods, theories, and domains in order to get a sense of what could make sense — and to assess whether something does. You can’t prompt what you can’t verbalize, and you can’t use what you don’t understand.

The problem, however, as Timothy Burke put it, is that AI is ‘brutally short-circuiting the processes by which people gain enough knowledge (…) to use the potential of generative AI correctly.’ Rightly or wrongly, junior researchers will increasingly feel that they have little incentive to spend long hours acquiring these skills—nor do senior researchers have much incentive to invest in training them.