In the previous episode, I talked about how AI is moving ever deeper into our decisions. It may assemble a list of candidates, shape decisions about people, prepare analysis, or suggest an answer on which we then base a decision. We also said that, when a decision truly matters, a person has to understand what happened, be able to verify the result, and, above all, assume responsibility for that result. But after that episode, I was left with another question: how do we form the person—and the competencies—capable of doing all this today?
I want to share a personal story. My son is twenty years old and has been studying applied computer science at Warsaw University of Technology for a year. It is a strong, highly competitive program. My son has always genuinely enjoyed learning. Yet lately I have heard him say more and more often: “Mom, I don’t understand why I need to learn this.” Or: “Mom, I don’t understand why we need these courses. You should hear our lecturers.” I do not hear an unwillingness to study in his words or his voice.
I hear a loss of meaning. I see that he is disoriented. In part, this may be a question facing a whole generation—a generation that now lives inside two systems. On one side, students at universities study algorithms, write code, debug errors, and master fundamental principles. On the other, AI is already beside them, in an environment where an application can be written, an algorithm explained, or a ready-made solution to almost any task proposed in seconds. That raises a basic question: what does it mean to know something in this situation?
Does a person still need to be able to do what a tool can already do? It is like asking whether you still need to know how to drive a manual transmission when an
automatic exists—and whether you need to know how to drive an automatic when there is an autopilot. That is what I mainly want to discuss today: how a professional now emerges, and how a person develops new competencies. On July 22, Business Insider journalist Thibault Spirlet published an article about new findings from Jacob Light. Light assembled data on more than fifty million course sections at more than one thousand U.S. colleges and universities going back to 1996. The striking finding was that, in the 2025–26 academic year, computer-science enrollment fell by 4.6 percent at the average institution—the first decline in roughly twenty years.
This is not a catastrophic collapse. Light himself says several changes may have coincided, and the data do not prove that AI caused the decline. At the same time, data from the National Student Clearinghouse Research Center—a U.S. nonprofit that collects and analyzes education data—point in a very similar direction. In fall 2025, enrollment in traditional four-year undergraduate programs in computer and information sciences fell by 8.1 percent. Data-science and data-analytics programs, however, continued to grow.
So the question my son is asking—and that, I think, many other young people are asking—is not merely personal. Students are beginning to reassess the educational contract itself. What exactly will I get here? Will it still matter four years from now? And why should I spend months or years on a task that AI can now complete in a minute? I am not suggesting that we debate whether basic skills provide access to higher-level thinking. That is obvious, at least for now. But something else is also obvious: the old instruction—“Learn it because that is what you are supposed to do,” or “Learn it because you have to”—is no longer enough.
A university must now be able to explain which capability each difficult task develops, and how and where that capability connects with AI. While preparing this episode yesterday, I asked my son, Danila: “How many times during the past year did the subject of AI come up in any of your courses?” This is applied computer science at Warsaw University of Technology. He answered: “Mom, zero times.” Not one lecturer raised the issue once in an entire year, in any subject. That is a collapse.
I initially set the thought aside, but it stayed with me. In substance, it is a catastrophic breakdown. Of course, it is not the same everywhere. When you prepare a topic like this, you can see the difference between some universities in the United States and, for example, some universities in Europe. Still, the example is revealing.
On July 23, a new preprint from a University of Michigan research center examined the relationship between access to generative AI and university grades. The team analyzed data from 156,000 students and nearly 88,000 individual course offerings at one large flagship U.S. university between 2015 and 2025. The university was not identified in the paper. The researchers compared periods before and after ChatGPT appeared, and looked separately at courses in which homework and essays were especially easy to hand over to generative AI.
They found no statistically significant additional increase in grades in those courses. Nor did they find a separate effect among students whose prior academic performance had been lower. In other words, students’ self-assessed understanding of the subject did not change either. The research points to something important. Access to AI by itself was not enough to produce an additional, measurable increase in grades or in self-reported understanding on the indicators the researchers observed.
That is probably the central conclusion. It does not prove that nobody learned more, but it also gives us no basis for believing that access to a powerful tool automatically creates a new, measurable competence.
We have discussed this many times on our channel: using AI does not automatically give a person a superpower or some superior knowledge that elevates that particular person to a different level. First and foremost, AI is a tool in the hands of the person using it. Perhaps a student completes the same task faster. Perhaps AI helps students bypass part of the effort. AI can certainly increase the speed and scale of action. But it does not automatically create judgment. It does not automatically create context.
And, as we have said before, it does not automatically create the ability to answer for consequences, assume responsibility, and develop that new competence. This connects directly with our previous episode about responsibility for decisions. It also leads to a broader question: where should we look to understand the real power of this extraordinary new tool that is now in our hands? I caught myself noticing something in my own life. After the new ChatGPT app appeared, after the continuing updates to Codex, and after the chat app and Codex were effectively brought together, Codex stopped being a tool I occasionally opened.
I have talked before about moving information from one environment to another. Today, Codex has become the environment in which I live and work—both professionally and for many personal tasks. By my own estimate, about ninety percent of my tasks now begin not with Google, for example, but there. The task may be domestic—figuring out a washing-machine setting, looking for a recipe, choosing a café—or it may be a work project, research, analysis, document preparation, or decision review.
At that point, it can look as though I have learned to perform the work of many different specialists. But have I actually acquired those competencies, or has the territory of my actions expanded farther than my knowledge? AI gives us breadth of action, but it does not transfer depth.
It does not transfer the professional knowledge that would allow me to make sound judgments in a field. My own experience meets what is now happening in the labor market. Several interesting surveys and research findings have appeared recently. On July 6, Atlassian’s Teamwork Lab published the results of a survey of one thousand U.S. knowledge workers. Ninety-two percent said their responsibilities had moved beyond their original job descriptions. That does not mean AI was the cause for everyone.
But the heaviest AI users were almost twice as likely to take on tasks belonging to other teams and to perform specialized work without an additional expert. Their areas of work and responsibility had expanded substantially. Then, just two days ago, another layer of evidence appeared. OpenAI’s research team published the first report in its Work at the Frontier series. It analyzed more than 800,000 work-related messages from U.S. ChatGPT users. After generic tasks such as email and scheduling were excluded, more than 43 percent of occupation-specific requests concerned work historically associated with another profession.
Among human-resources users, the share reached 69 percent. Tasks from three areas spread especially widely: finance, including calculations and interpretation of financial data; engineering and technology; and marketing. This brings us directly back to the issue we have just discussed. The findings do not mean that people have automatically mastered new professions. OpenAI’s research does not tell us whether the answer was ultimately used, whether it was any good, or whether a specialist reviewed it.
But it records the same shift Atlassian’s team described: people are taking on tasks from adjacent functions while their job title—and, in practice, their compensation system—remains the same, or may remain the same today. The authors of this research ask: who is prepared to do this work? Who is permitted to do it?
Who is paid for it? For me, this is no longer a technological question. It is an organizational one, and it can be viewed in two ways. On the one hand, a broader role can be called development. A person gains more autonomy, sees the work more broadly, discovers new horizons, performs new tasks, expands their area of responsibility, and connects several professional fields. But another scenario is also possible. A company may compress into one person the tasks that several specialists previously performed.
The person has more duties, but no additional time, training, or compensation. Then the change in the labor market becomes an intensification of work. At the same time, it becomes a blurring and restructuring of professional boundaries—a fundamental redesign of the work for which a person is responsible. Neither Atlassian nor OpenAI measured workload or compensation, so we cannot claim that this has already happened. But the evidence forces us to ask: if a role has become broader, what changed with it?
And what did not change? How does access to professional support change? What should students learn at universities, and how should they learn it? How do organizational structures inside companies have to change? There is also the question of professional depth, which follows directly from our earlier discussion. If AI allows me to produce a result in an adjacent field in which I previously had no experience, does that mean I actually know that field? Do I possess depth of knowledge?
Have breadth of output and depth of competence become the same thing? Or have they not? A tool can close a gap in my knowledge, but in doing so it may conceal from me the fact that the gap exists at all. I must not fall into the illusion that ChatGPT—or any other model—allows me to take over another person’s work completely. Or does it? If someone uses AI to enter another profession’s territory, who will verify the decisions? Who will examine the consequences? And how can that person understand the degree and depth to which they are qualified to make decisions in that other territory and assume responsibility for the outcome?
The studies we are discussing are still largely descriptive. I am convinced that every week we will see new reports and new angles. But they already reveal a powerful tension: a job may still have its old title while the actual territory of the work has become entirely different. The figures suggest this is happening across a fairly large area.
Now connect the pieces. The path of learning is becoming less clear. The future role is becoming broader. And where, in that system, is the first step into a profession? I want to use software engineers only as a clear example, because the profession and its changes are highly visible and sit near the front line. Seoul National University studied students’ experiences and published a small but very illuminating qualitative study. It does not establish the scale of the phenomenon—the researchers conducted about fourteen student interviews—but it explains the mechanism very well.
The authors call the mechanism “absorption.” A simple task that was once given to a junior employee does not disappear from the company. Instead, a senior employee now completes it much faster with AI. For the business, the work has clearly become more efficient. But the beginner has lost the opportunity to try, make a mistake, receive feedback, talk with a more skilled mentor or colleague, and learn to distinguish a correct result from an incorrect one, a good result from a bad one.
That is precisely the paradox we have just identified. The learning path becomes less understandable. What should we teach? How? For what purpose? The first rung of the professional ladder becomes very narrow. And this is not only about programmers; the software example simply makes the mechanism especially easy to see. At the same time, the future job is broader. It requires new tasks and an expansion of the territory in which a person works. The market now expects a young specialist to use AI immediately, enter adjacent functions, verify results, and act like someone with experience.
But where was that person supposed to acquire the experience? Who will become the next senior specialist? A student does not understand why an old task still matters. A grade does not show whether we have developed the new abilities and competencies. Employers, meanwhile, offer a new profession—a role that already extends beyond its previous boundaries. The answer to this dilemma is not to ban AI. I am certainly not arguing that companies must preserve routine, basic tasks simply because junior employees and recent graduates once learned on them.
My point is different. Universities and businesses both need to redesign the path by which competence is created. They also need to determine exactly which competence they want to develop in the new person, the new professional, the specialist we now need to cultivate.
This week I also read what I first thought was a research paper, but was actually a note by a pro-vice-chancellor for education at a British university. The author argued that, beginning in September 2026, AI must be integrated into every university curriculum. It should not matter whether the subject is civil engineering, marketing, finance, or something else: AI has to be addressed within every specialty students study. A curriculum built this way must answer three questions.
What must a person be able to do without AI? What must a person be able to do with AI? And how must that person verify the result produced at the end? The same kind of test is needed inside companies. If an employee’s role has expanded, do they have access to an expert from the adjacent function? Are they authorized to present or use a result, and to what degree? Do they receive feedback? Does their compensation change with their new responsibility? Without those elements, role expansion will look like simple overload in real life.
When I connect all the studies I have discussed today, I do not see proof that AI threatens
professions or education. I see a lack of synchronization. Technology has already changed how a result is produced, and it can now produce that result very quickly. But the university in our example has not always changed how it checks understanding. Inside companies, the title of the job has not yet changed even though the boundaries of the role already have. A person still has to work out what to learn, how to enter a profession, and what the next step is. That is the central dilemma now taking shape.
It is being raised by many researchers—not only AI researchers, but people working across very different areas of human life. That is why we should not discuss technology separately from the person. We need to keep asking: who makes the decision, who is responsible for it, and how do we know that a person working with AI truly possesses competence rather than merely obtaining the required result?