What kinds of knowledge will save you from AI?

For some professions, “AI is coming for our jobs” is no longer a vague threat about future events. Timothy McKeon, who spent years translating to and from Irish for the European Union, knows this better than most. As machine translation has improved, the ability to produce a text that is “good enough” has taken a huge bite out of his livelihood—costing him roughly 70 of his income as his EU work dried up. “The more it learns, the more obsolete you become,” he told CNN. And McKeon is not an outlier. 43 of translators have seen their incomes drop thanks to the increasing presence of AI alternatives in the marketplace. What is happening to translators is an early sign of an evolution that is now underway across the knowledge economy. For decades, much of the value produced by white-collar work rested on a straightforward proposition: you knew things or could find things or assemble things that most people could not, and others were willing to pay to gain the benefits of that knowledge. AI is collapsing the value of a broad swath of this market. In an increasing number of fields, a chatbot can now deliver in seconds work that is close to, or in some cases better than, that of an average professional. The bulk of the knowledge economy, the broad base of competent-but-unremarkable cognitive work, is being priced downward toward zero. It is tempting to think that the threat stops at the door of the merely average —that deep, specialized expertise is safe in a way that ordinary competence is not. That is only half right. The useful question is no longer whether AI will reshape knowledge work; it plainly will. It is which kinds of knowing hold their value when the machine can do so much. The way things were For most of the modern era, your market value as a professional came from your stock of knowledge: the tax code you had memorized, the case law you could marshal, the market data you had at your fingertips, the language you had spent a decade learning to render fluently. The work was, in large part, knowing things other people did not and being paid to retrieve and apply them. AI has learned to imitate that work in an increasingly convincing way. A frontier Large Language Model has read more tax code, more case law, and more market reports than any individual ever could, and it can hand most of it back on demand, fluently and instantly. The once-widespread idea that knowledge workers will be saved by the tendency of AI models to hallucinate is falling away. Once commonplace, hallucinations are becoming increasingly rare, and they can be mitigated in many contexts by effective prompting. Reliable LLM access isn’t quite free or frictionless, but when compared to human labor the cost is becoming negligible. The natural move for many knowledge workers in the face of these developments is to retreat upmarket: cede the simple work to the machine and stake their future on depth. Specialized expertise, the thinking goes, is the high ground. And there is real evidence for this. Translators, for example, have found that the surviving work is migrating upward: the volume jobs have gone to the machine, but the literary translators and the high-stakes legal and diplomatic interpreters—the people whose errors carry real consequences—still find their phones ringing. The specialists look safe . . . for now. But the ground they are standing on is less solid than it appears, and the line between the work AI can take and the work it cannot is not where most people assume it to be. Two kinds of knowing The problem is that depth of this kind is only a temporary refuge. To a machine, rare knowledge is nothing special, and there is no reason it can’t drill down to it so long as it is made available in a recorded form. The obscure corner of tax law is, to an LLM, just another corner. To ensure that your knowledge holds a more enduring type of value, you can’t rely on depth or rarity. You need different types of knowledge altogether. Two stand out. The first is contextual judgment. A seasoned consultant’s value was never just the industry detail in her head; it was knowing which detail mattered for this client or that board, which background fact guided how to read the problematic balance sheet, how to understand the half-articulated fear the CEO mentioned in passing. Deep expertise, however rare, involves reasoning over material that exists in the record (the obscure corner of tax law is written down somewhere), and that is something these models now do well. Contextual judgment is different. The decisive cue—what this silence means, why this board will balk—isn’t something that’s in the record precisely because this situation has never arisen in quite this form before. This kind of judgment relies on something real but fleeting, something the individual reads from the room in that specific moment. That can’t be looked up, and current models are far less reliable at this type of inference than at the recorded-knowledge reasoning they have already mastered. It may not stay out of reach forever, but it is not the threat knowledge workers face today. The second is procedural knowledge. Some philosophers make a useful distinction between “knowing that” and “knowing how.” You can know every proposition in every physics textbook and still be unable to keep your balance on a bicycle. You can absorb everything ever written about music theory and still not be able to play the violin. The same holds in business. A comprehensive store of facts and opinions about leadership is not enough to make someone a great leader. Reading every book on negotiation doesn’t translate into the ability to hold your nerve, time the concession, and keep your footing when the other side pushes. This kind of knowing lives in the doing: it can be acquired only through practice and experience, and at the highest levels it is bound up with things—trust, authority, the ability to read and relate to other humans—that exist only between people. That is not a stock of facts anyone could hand you, and it is not work you can hand off without becoming the bottleneck you were trying to remove. Neither of these types of knowledge can be downloaded. But both can be built deliberately. And that is where the serious effort of career development now belongs. Building survivable knowledge Here are three moves that can help put you on the right side of this historic change in what it means to be a knowledge worker. Own outcomes, not outputs. An AI model produces outputs: a draft, an analysis, an answer. So stop building your career around competing on this front. Audit what you’re actually paid for—your core value proposition—and strike everything that a good model can now do in minutes. What’s left are the outcomes only you can deliver: the messy problem carried from the initial diagnosis through to a result you can stand behind or the insight into what the client really needs that goes beyond what he says. Reorganize your role or your offer around these outcomes. Results—not a stock of facts—are your real moat. Build judgment in the room, not on the page. Situation-specific judgment can only be picked up firsthand by being present for consequential decisions and watching how they actually turn out. It resists mechanical replacement because what mattered in those rooms can never be fully summarized and passed into the kind of record an LLM can read. The people who advance fastest won’t be the ones who can store the most information, but the ones who find ways to improve their contextually grounded judgment. Delegate the routine; protect the practice. Procedural know-how lives in the doing, so the work you hand entirely to AI is work you stop getting better at. Push the genuinely rote tasks onto the model but keep doing the high-skill work yourself—the negotiation, the argument you think through—even when the model could turn out a passable version faster. Convenience now is paid for in capability later. Conclusion Timothy McKeon’s verdict about AI—the more it learns, the more obsolete you become—holds for certain types of knowledge, and those are the types that most professionals have built their careers around for decades. But there are other types of knowledge that are less vulnerable. Some may even be impervious to AI, at least in the forms available today. That kind of knowing can’t be downloaded. It is knowledge you embody rather than possess—earned in the doing, carried in the person, and yours in a way a stock of facts never was.