The Last Human Job: What AI Genuinely Cannot Do

The Moravec Paradox, Revisited

In 1988, roboticist Hans Moravec articulated something that had been bothering AI researchers for years. The things that are hard for humans, chess, calculus, legal reasoning, medical diagnosis, turned out to be relatively easy for computers. The things that are easy for humans, walking across an uneven surface, picking up a glass without breaking it, recognizing a face in bad lighting, understanding what a two-year-old is trying to say, turned out to be extraordinarily difficult for machines.

This became known as Moravec's Paradox, and for decades it shaped how researchers thought about the limits of automation. The insight was that human competence in sensorimotor tasks represents millions of years of evolutionary refinement. Our brains devote enormous computational resources to the seemingly simple act of reaching for a coffee cup. We do not notice this because we have been doing it since infancy. But the underlying computation is staggering.

Large language models have now demolished the upper half of Moravec's Paradox. The abstract reasoning, the language processing, the pattern recognition in structured data, AI handles all of this with a facility that would have seemed impossible ten years ago. But the lower half of the paradox, the physical world, the unstructured environment, the body in space, that remains stubbornly, expensively, frustratingly hard.

The Plumber Problem

Consider the plumber. Not a glamorous profession. Not one that appears on lists of jobs requiring advanced degrees or specialized knowledge. And yet, as of 2026, there is no commercially available robot that can do what a plumber does.

The problem is not the knowledge. AI can tell you exactly how to fix a leaking pipe, which fittings to use, what pressure ratings to check, how to solder copper correctly. The problem is the environment. A plumber works in spaces that were not designed for robots, under sinks, behind walls, in crawl spaces, in basements with low ceilings and poor lighting and pipes that were installed by different people over different decades using different standards. Every job is different. Every space is different. The physical dexterity required to work in these environments, to feel when a fitting is tight enough, to navigate a body through a space that barely fits a human, to improvise when the situation does not match the textbook, this is precisely what AI cannot do.

The same logic applies to electricians, HVAC technicians, carpenters, and roofers. These are not low-skill jobs. They require years of apprenticeship, a deep library of embodied knowledge that cannot be fully articulated, and the ability to solve novel physical problems in real time. The Bureau of Labor Statistics projects that employment of electricians will grow 9 percent from 2025 to 2035, much faster than the 3 percent average for all occupations, not because AI is not advancing, but because the physical world is not cooperating.

What Surgeons Actually Do

Surgery is often cited as a profession AI will transform, and in some respects this is already happening. Robotic surgical systems like the da Vinci platform have been in operating rooms since 2000, and AI-assisted imaging has genuinely improved diagnostic accuracy in radiology and pathology. But the surgeon has not been replaced. The surgeon has been augmented.

The reason is instructive. What a surgeon does in the operating room is not primarily a matter of knowledge or even technical skill in the abstract sense. It is a matter of judgment under uncertainty, in real time, with a body on the table. Tissue does not always behave as expected. Anatomy varies between patients in ways that no textbook fully captures. Unexpected bleeding requires immediate improvisation. The decision to proceed, to pause, to change approach, these are made in fractions of seconds, drawing on pattern recognition built from thousands of previous cases, filtered through the specific conditions of this patient, right now.

AI systems can assist with planning, with imaging analysis, with robotic precision in executing specific movements. But the judgment layer, the moment-to-moment decision-making in an uncontrolled environment, remains human. And the liability structure of medicine means that even when AI could theoretically make a decision, a human is required to make it and own the consequences.

The Therapist's Chair

Mental health is perhaps the most contested frontier in the AI-versus-human debate. AI therapy apps have proliferated rapidly. Wysa and dozens of similar products offer cognitive behavioral therapy techniques, mood tracking, and conversational support at scale and at low cost. For mild anxiety and depression, the evidence suggests these tools provide genuine benefit. Access to mental health support has historically been limited by cost and availability, and AI is genuinely expanding that access.

But the limits become visible quickly. A therapist working with a patient who has experienced severe trauma is doing something that no current AI can replicate. The therapeutic relationship, the specific, particular relationship between this therapist and this patient, built over months or years, is not a delivery mechanism for techniques. It is itself the treatment. The experience of being genuinely understood by another human being, of having your pain witnessed by someone who is also vulnerable, who also suffers, who also does not have all the answers, this is not a feature that can be engineered into a language model.

Researchers studying therapeutic outcomes consistently find that the quality of the therapeutic alliance, the relationship between therapist and patient, is a stronger predictor of positive outcomes than the specific technique being used. This finding has been replicated across decades and across different therapeutic modalities. It suggests that what therapy is actually doing, at its core, is something that requires two humans in a room.

Teaching Children

Education technology has been promising to replace teachers for as long as there has been education technology. Radio was going to do it. Television was going to do it. The internet was definitely going to do it. And now AI is going to do it.

The reality is more complicated. AI tutoring systems, particularly those built on large language models, have shown genuine promise in specific domains, mathematics, language learning, coding. A patient, infinitely available tutor that adapts to a student's pace and never gets frustrated is genuinely valuable. Khan Academy's Khanmigo and similar tools are providing real educational benefit to students who would otherwise lack access to personalized instruction.

But a classroom teacher does something that extends far beyond content delivery. A good teacher reads a room. She notices which students are disengaged and why. She understands that the boy in the back who is not paying attention is not stupid, he did not sleep last night because his parents were fighting. She adjusts not just the content but the emotional temperature of the room. She models curiosity, intellectual humility, the willingness to say "I don't know, let's find out together." She is, for many children, the first adult outside their family who takes their ideas seriously.

These are not supplementary features of teaching. They are, for many students, the most important thing that happens in a classroom. And they require a human being who is genuinely present, genuinely invested, and genuinely capable of being affected by what happens in the room.

The Jobs AI Made More Human

There is a counterintuitive development happening alongside the automation wave that does not get enough attention. In many professions, AI is not replacing human judgment, it is elevating the level at which human judgment operates.

A radiologist in 2026 spends less time on routine scan review, because AI flags the normal cases with high confidence and surfaces the ambiguous ones for human attention. This means the radiologist spends more time on the hard cases, the unusual presentations, the situations that require genuine expertise. The job has not been automated. It has been concentrated into its most demanding and most human elements.

The same pattern is emerging in law, in financial advising, in architecture. AI handles the research, the document review, the initial drafts, the routine analysis. The human handles the judgment, the relationship, the ethical reasoning, the creative synthesis. The proportion of the job that requires genuine human capability has increased, not decreased.

This is not a comfortable story for everyone. It means that the humans who survive automation are the ones who are genuinely excellent at the human parts of their jobs. The mediocre lawyer who was coasting on the ability to do research faster than clients could verify is in trouble. The exceptional lawyer who builds trust, reads situations, and makes judgment calls that change outcomes is more valuable than ever.

What This Actually Means

The jobs that are most resistant to AI automation share a cluster of characteristics. They involve working in unstructured physical environments where every situation is different. They require building genuine relationships with specific individuals over time. They demand real-time judgment under uncertainty where the cost of error is high and the variables are not fully knowable in advance. They involve being present in a way that matters, not just processing information, but being a human being in a room with other human beings.

This is not a comforting message dressed up as optimism. Plenty of jobs that do not have these characteristics are being automated right now, and the disruption is real and uneven. But it does suggest that the question "will AI take my job" is less useful than the question "what is the irreducibly human part of what I do, and am I actually good at it?"

The last human job is not a single occupation. It is a quality of presence, judgment, and relationship that machines have not yet found a way to replicate, and that, if the history of this technology is any guide, will remain just out of reach for longer than the headlines suggest.