The Deskilling Ledger: What You Lose Every Time You Let AI Think For You
The Deskilling Ledger: What You Lose Every Time You Let AI Think For You
There is a version of the AI deskilling argument that is obviously wrong: that using any cognitive tool degrades the relevant skill. By that logic, calculators have made us worse at arithmetic (true but irrelevant), GPS has degraded spatial navigation (partly true, generally fine), and writing has weakened memory (Plato worried about this in the Phaedrus). Tool-mediated cognition has always involved tradeoffs, and the tradeoffs have generally been worth it.
But there is a more careful version of the argument that the evidence is beginning to support, and it does not require any of the naive anti-tool premises. It simply requires taking cognitive science seriously: not all skills are equally recoverable once lost, not all forms of cognitive offloading are equivalent, and the rate at which current AI tools encourage offloading appears to be historically unusual.
The question is not whether to use AI. It is whether you are keeping an accurate ledger of what you are spending.
What the Data Actually Shows
A Microsoft and CMU study presented at CHI 2025 examined how AI assistance affects critical thinking in knowledge work tasks (n=319 participants). The finding that deserves careful attention: 40% of AI-assisted tasks involved zero observable critical thinking activity by the human participant. Not reduced critical thinking. Not diminished critical thinking. Zero.
This is not simply that workers were delegating tasks to AI — that would be unproblematic if the tasks were genuinely rote. The issue is that the tasks in question included ones that, when performed without AI assistance, consistently elicited critical evaluation, alternative hypothesis generation, and error-checking behavior. The AI assistance did not just speed up the work. It switched off a class of cognitive activity entirely.
A preprint from MIT Media Lab (arXiv:2506.08872) offers neurological correlates for this pattern — though it requires significant methodological caveats before drawing strong conclusions. The study measured brain connectivity in three groups: participants using LLMs, participants using search engines, and participants relying on memory alone. The LLM group showed the weakest brain connectivity across the relevant regions. More provocatively, the effect appeared to persist after AI use stopped, suggesting possible carryover rather than immediate recovery.
Critical caveat: this is a preprint with n=54. The sample is small, the methodology has not been peer-reviewed, and the persistence finding in particular requires replication before anything firm can be concluded. Treat this as a hypothesis-generating result, not an established fact. The CHI 2025 critical thinking data is on much stronger footing; the MIT brain connectivity finding is interesting but preliminary.
Taken together, these studies suggest a consistent directional signal: AI assistance, under at least some conditions, reduces the cognitive engagement that produces the learning and skill maintenance benefits of doing work.
Productive Struggle: What You Give Up When Work Stops Being Hard
Learning science has a well-established concept that is directly relevant here: productive struggle. The research — across mathematics education, language acquisition, motor learning, and complex problem-solving — consistently shows that difficulty, frustration, and error are not unfortunate byproducts of learning. They are mechanisms of it.
When a learner encounters genuine resistance — a problem they cannot immediately solve, a concept they cannot readily apply — the cognitive work required to push through that resistance is precisely what encodes the skill. This is why “desirable difficulties” (a term from Robert Bjork’s laboratory at UCLA) — spaced practice, interleaved problems, retrieval practice — consistently outperform more comfortable learning formats over the long run, even when they feel less effective during the learning session.
AI tools, at their best, eliminate unproductive struggle — the friction that comes from not knowing syntax, not finding the right documentation, not being able to hold a large context in working memory. That elimination is genuinely valuable. The problem arises when AI tools eliminate productive struggle — the friction that encodes understanding, builds mental models, and develops judgment.
The challenge is that productive and unproductive struggle are phenomenologically indistinguishable. They both feel like not making progress. They both generate the urge to reach for assistance. The discipline required to sit with productive struggle is hard to maintain, and AI tools make the escape from difficulty effortless.
The Offloading Decision Tree
Rather than a blanket prohibition or blanket permission, a more useful framework treats cognitive offloading as a decision with costs that need to be explicitly weighed.
Ask three questions before offloading a cognitive task to AI:
- Is this a skill I need to maintain or develop? If you are a senior engineer whose value lies in architectural judgment, offloading architectural reasoning to AI costs more than offloading boilerplate generation. The cost scales with the relevance of the skill to your professional identity and future work.
- Is the difficulty I am experiencing productive or unproductive? Unproductive difficulty: you don’t know the pandas DataFrame merge syntax. Productive difficulty: you can’t figure out why your distributed system has inconsistent state under certain race conditions. The first is worth offloading. The second is exactly the kind of struggle that builds the reasoning patterns you need for the next hard problem.
- How recoverable is this skill if I lose it? Some skills recover quickly with deliberate practice (syntax recall). Some skills require extended engagement with hard problems to rebuild (architectural intuition, debugging mental models, mathematical reasoning). The less recoverable the skill, the higher the cost of offloading.
The Compounding Problem
Individual offloading decisions feel low-stakes in the moment. The compounding effect is where the ledger becomes concerning.
Consider a senior engineer who, over 18 months, consistently uses AI for architectural suggestions, debugging assistance, and code review. Each individual delegation seems reasonable — the AI is helpful, the deadline is real, the output is good. But 18 months of avoiding the cognitive engagement required for deep debugging means 18 months of not maintaining the debugging mental models that took years to build. The skill does not vanish overnight. It attenuates gradually, invisibly, in exactly the way that aerobic fitness attenuates when you stop exercising — you feel fine until you need to perform at your previous level.
The CHI 2025 finding that 40% of AI-assisted tasks involved zero critical thinking is not just a snapshot. It is a map of what is not being practiced. And what is not being practiced is not being maintained.
An Anecdote Worth Taking Seriously
In April 2026, Axios reported that Andrej Karpathy — one of the most technically accomplished AI researchers working today, a founder of OpenAI and former head of Tesla AI — described what he called “AI psychosis”: running AI agents for 16 hours a day and writing essentially zero code himself. He framed it partly as exploration, partly as a concern.
Karpathy’s situation is not representative of most engineers’ workflows. But it is a useful data point precisely because he is not naive about AI tools. If someone with his level of expertise and self-awareness can find himself in a workflow where his own coding skills are going completely unpracticed, the gradient toward deskilling is steeper than most people assume.
The 20% Struggle Rule
A practical heuristic that follows from the learning science: commit to doing at least 20% of the cognitive work on any important task yourself, without AI assistance, even when AI assistance would be faster.
The 20% is not arbitrary. It is calibrated to be high enough to maintain productive struggle and skill practice, and low enough to remain compatible with the genuine productivity benefits of AI tools. The specific number matters less than the principle: treat unassisted cognitive engagement as a maintenance cost of professional capability, not as inefficiency to be optimized away.
For debugging: form your own hypothesis about the bug’s cause before asking AI to diagnose it. For architecture: sketch your own design before reviewing AI alternatives. For writing: draft your own opening section before using AI to develop the argument. The constraint creates the struggle; the struggle does the learning.
This is not anti-AI. It is pro-competence. The goal is to remain capable of evaluating AI output with genuine expertise, rather than becoming a well-compensated reviewer of AI suggestions without the underlying skills to know when the suggestions are wrong.
Questions Worth Sitting With
- Is there a threshold of AI use — hours per day, percentage of tasks, duration of sustained usage — beyond which skills don’t recover to baseline even with deliberate retraining? The MIT preprint hints at persistence effects, but we have no longitudinal data on recovery trajectories.
- The productive-struggle framework assumes that difficulty is a reliable signal for skill-building. But some difficulties are genuinely unproductive — irrelevant friction, environmental obstacles, poor tool design. How do you tell the difference in the moment?
- If AI tools become indispensable infrastructure (as calculators became in mathematics), does the deskilling question become irrelevant — or does it become more urgent, because the skills being lost are the ones needed to evaluate and repair the AI itself?
Try the 20% Rule for One Week
Pick one domain where you rely heavily on AI assistance — code, writing, analysis, whatever is most central to your work. For the next week, enforce a rule: before you consult AI on any non-trivial task, spend at least 20% of your estimated time on that task working through it yourself first. Note what happens — not just to output quality, but to how the subsequent AI interaction feels, what you catch, and what you miss. The exercise is a calibration tool as much as a productivity experiment. What you discover about your own skill maintenance will tell you more than any general study can.
