Where Are the Robot Maids?

What 100 Years of Failed Automation Predictions Can Teach Us About AI

In 1962, American television audiences were introduced to The Jetsons, a middle-class family living in the distant future of 2062. The Jetsons had flying cars, video calls, automated appliances, push-button meals, and, most memorably, Rosie, a humanoid robot who cleaned the house and performed many of the chores traditionally done by humans.

Some of the show's predictions were remarkably accurate. Video calling, flat-screen televisions, tablets, smart watches, robotic vacuums, and various forms of drones eventually became realities. Yet some of its most conspicuous predictions remain conspicuously absent. Flying cars, high-quality instant meals, and general-purpose robot housekeepers have not become ordinary parts of daily life.

More than six decades after Rosie appeared on television, millions of people still pay other human beings to clean their houses. They also pay humans to cut their hair, cook their food, repair their plumbing, teach their children, care for their elderly parents, build their houses, and drive many of their vehicles.

The history of automation is therefore much more complicated than the popular image of machines progressively taking over human work. We have been running an enormous trial-and-error experiment in automation since at least the Industrial Revolution, and its failures may tell us as much about the future of artificial intelligence as its successes.

The Great Automation Experiment

The basic economic proposition behind automation is compelling. If a human performs a repetitive task, build a machine that can perform the same task faster, more reliably, and more cheaply.

This worked spectacularly well in manufacturing. Machines replaced enormous amounts of human muscle. Mechanized agriculture transformed farming, while industrial machinery made it possible to produce textiles, steel, automobiles, and consumer products on scales that manual production could never approach.

The obvious assumption was that the process would continue. Machines had transformed the factory, so soon they would transform the home, the office, the hospital, the school, and the highway. The future would be automated.

But something interesting happened along the way. Some activities proved extraordinarily easy to automate. Others that appeared almost equally obvious proved extraordinarily difficult. Still others could be automated, but people did not particularly want them automated.

Those are three very different outcomes, and understanding the difference is important when thinking about AI.

The Robot Maid That Never Arrived

Housework seemed like one of the most obvious targets for automation. Mid-century visions of the future repeatedly imagined highly automated homes, and The Jetsons simply packaged ideas that had already been circulating for years. Smithsonian's history of the program notes that even a 1950 prediction about life in the year 2000 assumed homes would become so automated that household chores would increasingly happen through buttons, fingertips, or voice commands.

We did automate parts of housework. Washing machines replaced hand washing. Dryers replaced clotheslines. Dishwashers automated much of dishwashing. Vacuum cleaners reduced the physical work of cleaning floors, and robotic vacuums automated a narrow portion of that work even further, but Rosie never arrived.

The reason tells us something important about human intelligence. Cleaning a house seems easy because nearly every normal adult can do it. Yet consider what actually happens when a human cleaner enters an unfamiliar bedroom.

There may be a shirt on the floor, a glass on the bedside table, a book on the bed, a medication bottle, a wet towel, and a crumpled piece of paper. Almost instantaneously, the cleaner makes different decisions about each object. The shirt may go into the laundry. The glass belongs in the kitchen. The book should probably remain where it is. The medication should certainly not be discarded. The towel should be hung up. The paper might be garbage, or it might contain something important.

Then the cleaner notices that a chair needs to be moved to vacuum underneath it, while a sleeping cat should probably be left alone.

To a human, this is mundane. To a machine, it requires perception, physical manipulation, contextual understanding, judgment, and continuous adaptation to an environment it may never have encountered before, the easy human task turned out to be an extremely difficult machine task.

The Automated Kitchen

Cooking seemed similarly destined for extinction as household labor. Machines can measure ingredients more precisely than humans. Temperatures can be electronically controlled, recipes consist of sequences of instructions, and food can be chopped, mixed, heated, and timed mechanically, yet people still cook.

Part of the reason is technological. A general-purpose robot capable of opening a refrigerator, recognizing hundreds of ingredients in different states, peeling an onion, determining whether an avocado is ripe, handling a knife safely, improvising when an ingredient is missing, and cleaning everything afterward is a much harder engineering problem than building a microwave oven, but there is another reason. People like cooking.

For some people, cooking is an obligation. For others, it is recreation, creativity, relaxation, family interaction, or an expression of affection. Automating something a person dislikes can create value. Automating something the person enjoys can destroy part of the value they were receiving from the activity in the first place.

Why Are Humans Still Cutting Hair?

Haircutting would seem like another activity that should be relatively easy to automate. The geometry of a human head is reasonably predictable, computer vision can measure shapes extremely accurately, and robotic systems can make precise movements, yet barbers and hairdressers remain thoroughly human occupations.

Again, the physical problem is harder than it first appears. Hair differs, heads move, customers change their minds, styles are subjective, and a small mistake can matter greatly.

But there is something else going on. For many customers, conversation and personal interaction are part of getting a haircut.

People return to the same barber for years. They discuss their families, work, sports, politics, and local gossip. The barber knows how they like their hair without receiving detailed instructions every time.

The ostensible product is a shorter haircut. The actual product can be much broader: the haircut, conversation, familiarity, touch, attention, and human relationship all form part of the experience.

A robot could eventually become better at producing the haircut while still producing an inferior overall experience.

Construction and Repair

Factories became heavily automated, so it was reasonable to expect construction to follow. Instead, building sites remain filled with people, the difference is environmental structure.

A factory can be designed around the machine. Every object can arrive at a predictable location. Lighting can be controlled, surfaces can be standardized, and repetitive actions can occur thousands of times under consistent conditions, a 70-year-old house cannot be redesigned around the plumber who arrives to repair it.

The plumber may encounter pipes installed by different people in different decades, inaccessible spaces, corrosion, previous repairs, unusual fittings, furniture in the way, and problems that were incorrectly described by the homeowner. Electricians, carpenters, mechanics, and repair technicians confront the same general problem.

A mediocre human tradesperson possesses something technologically extraordinary: a general-purpose human body connected to a general-purpose human brain.

That person can climb stairs, crawl under something, move a box, recognize an unfamiliar object, change tools, apply variable force, answer a question, and improvise when the original plan fails. Because billions of humans possess these abilities, we regard them as ordinary. From an automation perspective, they are remarkable.

Where Are the Self-Driving Cars?

Driving provides another lesson. The idea is surprisingly old. Experiments with driverless automobiles go back about a century, including a radio-controlled car demonstrated in New York in 1925.

Much later, as computing and machine vision improved, autonomous driving repeatedly appeared to be just around the corner. Enormous progress has occurred. Autonomous taxis now operate in some places, while increasingly sophisticated driver-assistance systems handle portions of ordinary driving, but universal autonomous driving proved harder than many optimistic predictions suggested.

The reason is that routine driving is relatively predictable. The difficulty lies in everything that is not routine: construction zones, strange road markings, emergency vehicles, pedestrians behaving irrationally, objects falling into the road, weather, human drivers breaking rules, or a police officer giving hand signals that contradict a traffic light.

Driving requires not merely controlling a vehicle but interpreting an open-ended physical and social environment. That last few percent of unusual circumstances can be vastly harder than the first 90 or 95 percent of the problem.

Even in 2026, the transition is still occurring incrementally. Some commercial robotaxi services are expanding, while other deployments remain supervised or geographically restricted.

Doctors and Teachers Did Not Disappear Either

Computers were also supposed to transform professions based primarily on knowledge. Medicine appeared particularly promising.

A computer can remember vastly more medical information than any physician. It does not forget obscure diseases. It can calculate probabilities and compare thousands of variables, yet physicians remained.

The reason is partly that medicine was never simply an information-retrieval problem. Patients describe symptoms imperfectly. Several diseases can occur simultaneously. Tests can conflict. Treatments involve tradeoffs, and a frightened patient may behave differently from a textbook case. Someone still has to decide which information matters and accept responsibility for acting on it.

Teaching presents a similar problem. If education consisted simply of transferring information from a knowledgeable source to a student, recorded lectures should have eliminated much of teaching decades ago. They did not.

A teacher also notices confusion, motivates a reluctant student, changes explanations, maintains discipline, answers unexpected questions, and provides social reinforcement again, the formal description of the task captures only part of the actual job.

Sometimes We Do Not Want the Machine

There is an even deeper reason some activities resist automation: the activity itself can be part of the value.

People garden despite having supermarkets. They run despite owning cars. They bake bread despite being able to buy it for a few dollars. They paint despite cameras being able to reproduce reality far more accurately. They play chess despite computers having become vastly stronger chess players. They lift heavy objects in gyms despite spending centuries inventing machines specifically to eliminate heavy physical labor.

Some people even find cleaning, ironing, or other repetitive household activities calming, this exposes a flaw in the conventional way we think about automation. We often treat human effort as inherently negative, it is not.

Sex and the Ultimate Automation Paradox

Consider an extreme example: sex, from a purely biological perspective, sex performs the enormously important function of reproduction. But technology can separate reproduction from sexual activity. That does not make sex obsolete.

The obvious reason is that reproduction is not the only, or usually even the immediate, value people seek from sex. The experience, intimacy, physical pleasure, and interaction with another person are themselves central to its value, automating the biological output therefore does not automate what the participants actually value.

The same principle, in less dramatic forms, applies throughout life. A machine can prepare dinner, but if I enjoy preparing dinner, eliminating the activity has not necessarily improved my life. A machine could someday cut my hair perfectly, but if I enjoy spending half an hour talking to my barber, eliminating the barber also eliminates part of the product I was purchasing.

A machine can defeat me at tennis, but the purpose of playing tennis is not to determine whether a machine can hit a ball better than I can, sometimes the activity is the product.

Output Value and Experience Value

This suggests a useful distinction. An activity can produce at least two kinds of value: output value and experience value.

Output value is the value of what the activity produces. Washing a dirty sewer pipe, for example, consists overwhelmingly of output value. Most people would happily receive the same result without personally doing the work.

Experience value is the value experienced while performing or participating in the activity. Playing with your child can have enormous experience value. Having a robot play with the child on your behalf may free an hour of your time, but it does not necessarily produce the same value.

Activities therefore fall along a continuum, the economically relevant question is not simply, "Can we automate this?", It is, "Which portions of this activity does the human beneficiary actually want automated?"

Why AI May Be Different

This history should make us skeptical of confident predictions that AI will simply eliminate entire professions. We have heard similar predictions before, but AI really does differ from many previous waves of automation in an important respect.

Earlier machines were exceptionally good at narrow physical repetition. Industrial robots thrive in factories precisely because factories can be structured around them. Computers then became exceptionally good at calculation and structured information processing, generative AI attacks something different: routine cognitive competence.

It can write an acceptable paragraph, produce ordinary computer code, create a conventional advertising image, summarize a document, translate text, analyze routine information, prepare a presentation, or generate a standard contract.

These activities occur primarily inside the digital environment, which eliminates much of the physical complexity that protected plumbers, cleaners, and construction workers. In many cases, the customer also does not particularly care whether a human performed the task, that combination matters.

The Vulnerability Test

History suggests that an activity is especially vulnerable to automation when several conditions occur together. The desired output can be clearly specified, the environment is digital or highly structured, routine cases constitute most of the economic value, machine errors can be detected or tolerated, the customer does not care who performed the work, and the human does not particularly value performing the activity.

Routine cognitive work increasingly satisfies all six conditions, that is why AI may threaten an average graphic designer before it threatens an average plumber.

The graphic designer may require considerably more formal training, but the designer produces a digital artifact in an environment AI can directly access. The plumber walks into an unpredictable physical environment and manipulates reality.

Four Shelters From Automation

The long history of automation suggests at least four broad reasons human activities survive, the first is exceptional capability. A human performs at a level sufficiently unusual that people value that particular person's judgment, creativity, athletic ability, leadership, or other capability.

The second is physical complexity. The task occurs in an uncontrolled environment where a general-purpose human remains cheaper or more capable than specialized machines.

The third is human interaction and provenance. People value having another human involved. This protects portions of caregiving, entertainment, hospitality, personal services, athletics, and many other activities.

The fourth is activity value. People value performing the activity themselves. Cooking, gardening, exercise, crafts, sports, and countless hobbies survive regardless of whether machines could produce the nominal output more efficiently.

These protections can overlap. A great chef provides exceptional skill, human provenance, and an experience. A massage therapist provides physical manipulation and human interaction. A professional athlete provides exceptional performance and human provenance. A person cooking dinner at home may simply enjoy doing it.

AI Should Automate Burdens, Not Life

The Jetsons imagined technological progress largely as the elimination of effort. Push the button and the machine does the work.

It was an understandable vision. Much of human history involved exhausting labor that people were delighted to escape, but eliminating effort is not the same thing as maximizing human welfare.

People voluntarily climb mountains. They run marathons. They restore old cars. They grow tomatoes. They learn musical instruments. They solve puzzles. They build furniture. They spend hours preparing meals that could have been delivered in 20 minutes.

Difficulty can produce accomplishment. Repetition can provide relaxation. Creation can provide meaning. Physical effort can feel good. Interaction with other humans can be valuable independently of whatever task is supposedly being accomplished, the ideal use of AI is therefore not necessarily maximum automation. It is selective automation.

We should automate what people regard as burdens, assist with activities where they want assistance, preserve the parts they value doing themselves, and preserve human interaction when human interaction is itself part of the desired experience.

The Lesson of Rosie

Rosie the robot remains useful precisely because she has not arrived, She reminds us how easy it is to look at an activity from the outside, identify its apparent output, and assume that producing that output mechanically constitutes successful automation, history says otherwise.

Some seemingly simple tasks contain hidden physical and cognitive complexity. Some occupations consist largely of handling exceptions rather than performing the routine task described in their job title. Some activities derive their value from interaction with another human. And some things remain human simply because humans enjoy doing them.

AI will undoubtedly automate an enormous amount of work. In information-based occupations, it may move far faster than many earlier automation technologies because the environment is already digital and the outputs can often be reproduced at negligible marginal cost.

But the history of automation provides a warning against confusing technological capability with human value, the future will not be determined simply by asking, "What can machines do?"

The more important question is, "What do humans still want to do, and what do they still want other humans to do with them or for them?"

Technology determines what can be automated. Economics determines what is worth automating. But ultimately, human beings determine what they actually want automated.