51.8% of Korean Workers Now Use AI on the Job

63.5% of Korean Workers Have Tried Generative AI

A survey released by the Bank of Korea last summer shows that 63.5% of Korean workers have tried generative AI, and narrowing it to work purposes alone, 51.8% are actively using it. That is roughly double the adoption rate in the United States (26.5%). The pace of diffusion is even more startling. 3 years after the commercial debut of the internet, its adoption rate stood at 7.8%; AI adoption today is 8 times higher.

Even more striking is the time spent. Korean workers dedicate 5~7 hours a week to AI, compared to 0.5~2.2 hours in the US. Heavy users who rely on it for more than an hour a day account for 78.6%, more than double the US figure (31.8%).

When AI is used this frequently and for this long, it ceases to be a mere matter of tooling—it rewrites how work itself gets done. In an environment like this, the reflexive advice to “just use it carefully” rarely works as promised.

AI Has Stopped Being an Optional Tool and Become a Working Condition

In the same survey, 61.2% of respondents said their organizations encourage AI use (Nowsurvey, sample of 1,000 office workers). That means well over half of working professionals are opening AI under pressure to use it.

We need to re-examine the word “tool” here. By definition, a tool is an object you can pick up and put down at will. When you are done with a hammer, you can simply toss it back in the drawer. But can we genuinely call something we fire up every single day, for 5–7 hours a week, a tool we can set aside? Before its arrival, the workday began with a blank document; today, it begins in a prompt window. If a tool is something you choose whether to use, something that becomes the baseline condition of work is an environment. AI today is far closer to the latter.

There is another curious figure. The Bank of Korea estimated that AI has trimmed workers’ hours by an average of 3.8%—about 1.5 hours in a standard 40-hour workweek. Yet in the same survey, 54.1% of respondents reported that their working hours did not decrease despite using AI. More than half of these workers are not saving any time, yet they keep using it anyway. Why? They are not using it to save time; they are using it because they are already operating inside an environment that demands it. Once the workflow around you assumes AI by default, choosing not to use it on your own becomes nearly impossible. You can hardly start from scratch on a blank page when every colleague around you is churning out drafts with AI.

In practice, this environment exerts a quiet, insidious pressure. In a related survey, 27.3% of respondents reported higher stress levels following the rollout of AI, with the No. 1 reason cited as “pressure over the pace of change.” If something is merely a tool and you dislike it, you can just walk away. An environment, however, makes the very act of opting out feel like falling behind. The moment not using something makes you anxious, it has already moved well beyond an optional tool you can take or leave.

Even Cautious Teams Are Losing Their Judgment

Media theorist Marshall McLuhan pointed this out decades ago: the moment we say of any medium, “It all comes down to how you use it,” we are already under its spell. He famously compared a medium’s “content” to a juicy piece of meat a burglar tosses to a watchdog. While the dog is busy gnawing on the meat, the intruder quietly loots the house. Our complacency—assuming that “as long as we get the content right, we’ll be fine”—is precisely that piece of meat. While we deliberate over what prompts to feed the model, the architecture of our thinking is quietly being altered.

The philosopher Martin Heidegger captured this dynamic with the concept of Gestell (enframing)¹. A tool is not merely something we use; it simultaneously herds us into being a particular kind of subject. Consider a chair. A chair quietly issues commands to the body: “Sit down, face forward, stay still.” We believe we are using the chair, but the chair is actively shaping us into a sedentary creature. AI works the same way: by organizing answers and framing choices, it steadily directs our judgment and imagination.

One tech-criticism essay cites a case study from a 600,000-person enterprise: the teams that used AI most cautiously were actually the first to lose their ability to distinguish between a “safe choice” and a “good choice.” In day-to-day work, the pattern looks like this: AI reliably produces safe, average answers. Accept those answers repeatedly, and the subtle instinct that once separated the merely serviceable from the truly exceptional begins to atrophy. People simply get fewer chances to make an unconventional call and see it through on their own judgment. Crucially, this side effect does not spare cautious users—in fact, it catches them first, because the conviction that they are being careful doubles as a justification to stop thinking independently.

Recent brain-imaging research has begun to document this at the neural level. In an experiment conducted by an MIT research team with 54 participants (these are preliminary, non-peer-reviewed findings, so caution is warranted), subjects were divided into three groups to write essays—an AI group, a search-engine group, and an unassisted “bare-hands” group—while their brainwaves were monitored. The AI group exhibited the weakest neural connectivity. The researchers termed this cognitive debt². Even more striking: 2/3 of the participants who used AI could not even accurately recall the topic of the essay they had just written. A finished product was generated, but virtually nothing took root in memory. When the researchers asked them to write unassisted again in a final session, their neural connectivity remained depressed. Stepping away from AI, even temporarily, did not produce an immediate rebound.

Interestingly, a similar warning came from the religious world. In Magnifica Humanitas, an encyclical³ issued this past May, Pope Leo 14 observed that technology is never neutral, because it faithfully mirrors the character of those who build, regulate, and use it. The Pope urged young people to engage with AI in a way that leaves them capable of thinking for themselves even if AI were to disappear tomorrow. That counsel applies not only to teenagers, but to any of us logging 5 to 7 hours with AI every week.

The ‘Curbs’ to Hand Over to AI, and the ‘Mountains’ to Climb Yourself

So what should we do? One researcher of tool design divides the friction we try to eliminate into two categories.

The first is the “curb.” Like a street curb blocking a wheelchair user, it is a barrier with no justification to exist. Cutting it down is simply the right move; accessibility designers call curb cuts progress. Repetitive formatting, scheduling meetings, converting file types, and drafting 1st-pass summaries fall into this bucket. Handing these over to AI costs you nothing. In fact, you should reclaim that time and spend it on things that actually matter.

The other is the “mountain.” With a mountain, the climb itself is the entire point. Think of the gym: having a robot lift the weights for you is completely pointless. The goal is not to move the iron; the goal is to transform the person lifting it. Grasping an unfamiliar field from scratch, structuring the logic of a tough decision on your own, and shaping the framework of a piece of writing from the ground up belong here. Hayao Miyazaki once noted that if all of life’s hassles disappeared, we would end up missing them. Certain difficulties build competence precisely through the struggle of working through them.

The problem is that most professionals lack the vocabulary to separate mountains from curbs. So they offload both to AI indiscriminately. On the surface, both look like the exact same thing: a chore. Tedious formatting (a curb) and the exercise of building an argument from first principles (a mountain) are both outsourced with a casual “AI, do this for me.” Handing off the former is fine, but handing off the latter strips away the opportunity to cultivate your own judgment. The real trouble is that this loss is not visible immediately that day. It reveals itself months later, in a quiet, unnerving moment: “Why does a decision I used to make effortlessly feel so paralyzing now?”

Before delegating, there is only one question you need to ask yourself: If I hand this task to AI, am I merely saving time, or am I erasing the friction that helps me grow? If it only saves time, delegate without hesitation. If it removes growth-inducing friction, you must protect the order of operations: reach your own conclusion first, and bring in AI only to cross-check your work. Once this single question becomes a habit, it becomes immediately clear what to delegate and what to defend.

Of course, the line between mountains and curbs is not always sharp. What was a mountain yesterday can become a curb today once you master it. When entering an unfamiliar domain, you have to wrestle with it yourself; once you know the terrain, the repetitive mechanics can be handed off. This distinction is never a one-time choice. It is a boundary you must redraw continuously as your skills expand. That might sound like extra work, but the very act of redrawing that line is proof that you are still growing.

Oswarld’s Lens

To be completely candid, what strikes me as most terrifying about this story is that “the most cautious teams collapse first.” Having observed AI adoption across numerous organizations while developing GTM strategies, I have repeatedly run into this paradox: the more meticulously governed a team is, the more uncritically they tend to accept AI outputs. The more elaborate the guidelines, the more reassurance sets in: “This is an approved method, so it must be fine.” That reassurance shuts down judgment. The most dangerous red flag I have witnessed is when a team stops examining raw data directly and begins making decisions based solely on AI-generated summaries. Summaries are always deceptively seamless, concealing the exceptions and counterexamples sliced away beneath the surface.

There is also a dimension here that no individual can shoulder alone. Dismissing the issue with platitudes like “AI is all about how you use it” is convenient, but framing it as an individual’s mindset only obscures systemic problems. The environment is shaped not by the individual, but by corporate adoption policies, evaluation methods, and regulations. That is why I believe an organization’s design of what to automate and what to reserve for humans is a far bigger issue than individual ‘smart usage.’ To expect individuals to distinguish mountains from curves, the organization must first set the criteria for which tasks can be handed over to AI and which humans must handle themselves. I often share this advice during consulting engagements: the teams that decide where not to use AI before asking where to deploy it are the ones that ultimately go the distance. A list of what not to automate is, after all, the exact territory an organization has chosen to preserve for human judgment.

Closing

Let me sum things up. First, South Korea is the world’s most intensive adopter of AI, making it the first to undergo the shift from a tool to an environment. Second, inside an environment, prudence alone is no defense. Third, what professionals need in practice is the judgment to tell which tasks are curves you can cut, and which are mountains you must climb yourself.

Just for today, think back to one task you handed over to AI this past week. Was it a curve worth cutting, or a mountain whose entire point lay in the climb? You do not need grand resolutions; simply making that single distinction establishes a clear standard for your next delegation.

Have you ever found yourself thinking, “I shouldn’t have handed this to AI,” regretting it later? Share in the comments which tasks turned out to be your personal “mountains.” If there is enough response, I will break down the categories of “tasks you should never delegate” in the next issue.

---

💬 Let me know in the comments which tasks were your “mountains.” I will weave your thoughts into the next issue.

📨 If you know someone who relies heavily on AI, please share this issue with them.

---

Recommended past issues

• How Thinking Muscles Collapse in 10 Minutes
• The Verifier Bottleneck Is Only Half the Story

References & Further Reading

Primary sources

• Bank of Korea, “Rapid AI Adoption and Productivity Effects: Based on Household Survey Data,” BOK Issue Note No. 2025-22, 2025. Source ··· All the figures for Korea cited in today’s letter come from this report. A single glance at the cross-country comparison reveals just how exceptional Korea’s situation is.
• Nataliya Kosmyna et al., “Your Brain on ChatGPT: Accumulation of Cognitive Debt,” arXiv preprint, 2025. Source ··· With a sample size of 54 and peer review still pending, we shouldn’t jump to conclusions. Still, the finding that “AI-assisted participants couldn’t recall their own writing” gives serious food for thought.
• Pope Leo XIV, Encyclical Magnifica Humanitas, 2026. Source ··· The single sentence “Technology is not neutral” forms the philosophical backbone of this essay.

Background

• Frank Elavsky, “Stop saying that AI is just a tool” ··· The original essay distinguishing mountains from curves, and repetitive drudgery from meaningful struggle.
• L. M. Sacasas, “Your AI Is Not a Tool” ··· The starting point for framing AI as an “environment rather than a tool.” It charts an intellectual lineage connecting Marshall McLuhan and Ivan Illich.

---

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

Footnotes

1. Gestell (enframing): A concept coined by Martin Heidegger describing the “framework” through which modern technology compels us to view and order reality in a specific way. The idea is that while we use technology, technology simultaneously shapes who and what we become. ↩

2. Cognitive debt: A metaphor comparing the cost of offloading our thinking to AI to financial debt: delegating cognition feels convenient in the short run, but our capacity to think independently diminishes down the line, returning with heavy interest. ↩

3. Encyclical: An official pastoral letter issued by the Pope to the worldwide Catholic Church, outlining authoritative teachings on critical moral, social, and ethical questions. ↩