The previous post, Build the Ecosystem, Not Just the Robot, discussed the incredible difficulty of creating a household robot capable of handling typical chores—cooking, cleaning, buying, etc.—that humans may find trivial. Still, their variety makes them devilishly difficult to build and program.

Brains, fingers, feet, eyes, noses, ears, skin, and taste—all attributes humans possess courtesy of nature—are hard to replicate. Even driving a car, a task humans perform with minimal thought, has proven challenging for robots.

Here are examples of solutions.

Automate the Household, Not the Housekeeper

The effort to develop household robots may be aimed at the wrong objective. Rather than trying to create an artificial human capable of performing thousands of unrelated physical tasks, in various environments, it may be far easier to automate the details of the household itself.

The general-purpose robot would become less a maid, cook, gardener and handyman than a butler: a mobile coordinator connecting the human with a collection of specialized automated systems.

A modern automated factory does not ordinarily employ one fantastically capable humanoid robot that walks around welding, painting, machining, inspecting, packing and driving a forklift.

It distributes those functions among machines optimized for particular jobs and coordinates their activities. A highly automated home could operate according to the same principle.

The “Roomba” offers a simple example. One approach to automated floor cleaning would be to teach a humanoid robot to find a vacuum cleaner, grasp its handle, operate its controls, maneuver it around furniture, determine what it has already cleaned, and return the vacuum when finished.

The simpler solution is a vacuum cleaner that moves by itself. The specialized machine needs to understand only its specific job and the part of the environment relevant to that job.

The principle can be extended throughout the house. Don’t give the general-purpose robot a scrub brush and teach it to clean a bathtub; develop a bathtub-cleaning system.

Don’t give it a squeegee and teach it to wash windows; incorporate cleaning mechanisms into the windows or provide a specialized window-cleaning machine.

Don’t teach the robot to mow the lawn; give the lawn its own machine. Don’t teach it to load a dishwasher; automate the entire food-service and cleaning process.

The general-purpose robot—think of it as a butler robot—then doesn’t need thousands of highly refined motor skills. It needs to understand human intentions, communicate those intentions to specialized systems, coordinate their activities, supervise results and deal with exceptions.

Physical manipulation remains useful, but it becomes one capability among many rather than the foundation of household automation.

Consider food shopping. Even sending a robot to a store may preserve an unnecessary human-era activity. If every food item has an electronic identity and the refrigerator and pantry maintain continuous inventories, the house already knows what it contains, how much remains and the household’s historical consumption. It can predict that milk will run out Tuesday afternoon or that only two eggs remain.

Routine shopping therefore can disappear. The household system can compare suppliers, prices, quantities and delivery times and automatically replenish ordinary items within limits established by the resident. The system can refer unusual purchases or significant price changes to the resident for a decision.

The store, which can resemble an “Amazon-style” warehouse, can deliver standardized containers directly to the house’s receiving dock, where refrigerated items are routed to refrigerated storage, frozen foods to frozen storage and dry goods to their appropriate locations.

Inventory would be updated automatically.

No robot needs to drive anywhere. No car needs to park. Nobody needs to walk through aisles. The shopping trip was merely another traditional human task whose underlying purpose—maintaining an appropriate inventory of food—can be accomplished without the task itself.

Meal preparation can be treated similarly. A resident might say, “I want steak, mashed potatoes and peas at six o’clock, with approximately 300 calories and 12 grams of protein.” The butler robot does not peel potatoes or stand over a broiler. It converts the human request into objectives, preferences, constraints, and a deadline, then communicates them to the household food system.

The food system selects appropriate ingredients from inventory, determines quantities, prepares and cooks the components at appropriate times and temperatures, and produces the finished meal.

If the requirements conflict—for example, if the requested steak portion by itself would exceed the calorie limit—the system reports the problem to the butler robot, which asks the resident whether to reduce the portion or increase the calorie limit.

This is where general-purpose artificial intelligence is particularly valuable: between human intention and specialized machinery. The human specifies the desired result. The AI interprets that result, resolves ambiguities, coordinates the necessary systems and reports problems requiring human judgment.

After the meal, there is no reason to assume that the traditional sequence of clearing dishes, loading a dishwasher, unloading it and putting dishes into cabinets must survive. Reusable dishes might move automatically into an integrated cleaning-and-storage system.

Alternatively, dishes and utensils might be fabricated when required from recyclable materials and returned afterward to the material supply. The table itself might be built and programmed to perform some of these functions.

The important point is not which particular solution ultimately prevails. Tasks should not become the one butler robot’s automated task.

The correct first question is not, “How can we build and program this robot to fold laundry?” It is, “What is the objective of folding laundry?” If the objective is to store clean clothing conveniently, perhaps a different storage system eliminates folding altogether. Or at worst, the storage system would be programmed and built specifically to fold laundry, an easier build than trying to do it with an all-purpose robot.

The objective of shoveling snow is not shoveling snow; it is keeping necessary outdoor surfaces usable. The objective of replacing a roof shingle is not manipulating shingles; it is keeping the building weatherproof. The objective of washing a plate is not washing a plate; it is providing a clean surface to eat from.

This suggests a fundamental rule for household automation. Don’t automate the task. Identify the task’s purpose and automate that specific purpose.

The resulting house becomes a network of specialized systems. Cleaning machines, food-storage systems, cooking equipment, laundry systems, waste processors, environmental controls, security equipment, delivery systems, garden equipment, structural monitors and repair systems continually exchange information.

The mobile butler (the primary robot) is one participant in that network, serving mostly as the resident’s general-purpose representative and coordinator.

Information can travel in both directions. The refrigerator can tell the food system that a steak should be used soon. The food system can tell the butler that steak therefore would be a sensible dinner choice. The butler can ask the resident, “How about steak tonight?” A four-word answer can initiate dozens or hundreds of coordinated operations.

The architecture also offers an important reliability advantage. If a single extraordinarily sophisticated humanoid robot performs every household function, that robot’s failure may disable virtually the entire automated household.

In a distributed system, a window-cleaning machine failure means the windows remain dirty until that component is repaired. Food preparation, laundry, floor cleaning, security and package delivery continue functioning.

Specialization also permits incremental technological improvement. A homeowner need not replace a fantastically expensive general-purpose robot because someone has invented a superior floor-cleaning technology. The floor-cleaning component can be replaced independently. The same principle has made industrial automation practical and continuously improvable.

Eventually the distinction between robot and house may become increasingly artificial. The building itself contains sensors, communications, storage, machinery and specialized automated systems. Mobile robots provide those capabilities requiring mobility. Fixed machines provide capabilities that do not require mobility. The intelligence coordinating them may be distributed throughout the system.

The house, in effect, becomes a modular robot, with each replacable module trained for a specific task..

This approach also could make household automation achievable sooner. An artificial human capable of successfully manipulating virtually every object and tool found in an ordinary house presents an extraordinary robotics problem. A coordinated collection of machines, each optimized for a relatively narrow function, presents many smaller and more manageable engineering problems.

The central research objective therefore should not be “Build the perfect household robot.” It should be “Automate the household.”

For every existing household activity, ask three questions.

1. What is the actual objective?
2. Does the traditional task need to exist at all?
3. If it does, should a general-purpose robot, a specialized machine, or the house itself perform it?

The result may be a paradox: a far more automated home that requires a far less capable humanoid robot.

Instead of making a robot sophisticated enough to handle the extraordinary complexity of a house designed entirely for humans, divide the system into tasks and make the whole system sophisticated so the robot doesn’t have to be.

To begin the evolution, relevant industries should coordinate their efforts so they can determine some initial standards. Companies that make stoves, washing machines, windows, roofs, lighting, floor cleaners, building materials, tools, and robots, etc., should begin working together to create the “smart house.”

That is where the future lies, not with an all-purpose robot, but with a society assisted by automation.

Rodger Malcolm Mitchell

Build the Ecosystem, Not Just the Robot

In “Grasping the World,” from the October 2026 issue of Scientific American Magazine, author Adam Rogers highlights the immense challenge of creating an all-purpose robot that can operate in a world designed for humans.

While a robot can master a backflip by focusing on its own body mechanics, tasks like scrambling eggs, folding laundry, or handling grapes are far harder because they require the robot to navigate an infinitely variable external environment.

Google DeepMind is working to solve this through AI that generalizes skills across different tasks and robot bodies, yet even seemingly trivial human activities remain a hurdle—in one test, a robot successfully used a dustpan only 32 percent of the time.

We may be approaching the problem from the wrong direction.

Most household-robot research treats the environment as fixed. Because our houses are filled with objects designed over centuries to accommodate the human body—doorknobs, light switches, utensils, and tools—we assume the challenge is to build a robot smart enough to operate all this equipment.

Instead, we should work both sides of the problem: Don’t build robots that can function in our existing houses.

Start building houses that work with robots.

Consider something as simple as hanging a picture. A human installs screw eyes and wire on the frame, decides where the picture should go, handles a tiny hook and nail, pounds the nail into the wall, lifts the picture, catches the wire on the hook and straightens the frame.

Teaching a robot to perform all those operations would require considerable visual, tactile and manipulative ability.

But why should a robot hang a picture that way?

A robot-compatible wall might contain an invisible mounting grid. Picture frames could have standardized attachment systems. The robot merely places the picture against the desired portion of the wall; the mounting system grabs it and levels it.

An extremely difficult robotics problem has almost disappears—not because the robot became smarter, but because the environment cooperated.

The same principle applies throughout the house. We should not teach a robot to fold laundry. We should solve the clothing-cleaning-and-storage problem, perhaps designing clothing and storage systems that make folding unnecessary or do the folding themselves.

We should not teach a robot to load a dishwasher. We should solve the eating-vessel problem. Dishes might remain in an integrated cleaning-and-storage system. Inexpensive dishes might be fabricated ad hoc, and their material recycled afterward.

We should not teach a robot to climb a ladder and replace a roof shingle. We should design roofs with standardized, machine-accessible, self-replacing modules.

We should not teach a robot to shovel snow. The objective is to keep necessary outdoor surfaces usable after snowfall. Snow shoveling is merely the historical human solution to that problem.

This suggests a fundamental rule: Don’t automate the task. Identify the task’s purpose, then redesign the system to accomplish it.

That changes the meaning of the “smart house.” Today a smart house generally means a conventional house with added intelligence: thermostats, lights, locks, cameras and appliances that communicate electronically.

The house we are describing would be fundamentally different. Its physical architecture would be designed from the beginning for cooperation among humans, stationary machinery and mobile robots.

And the concept would extend far beyond the house.

Consider food. The objective “prepare dinner” conceals an enormous chain of activities. Menus must be selected. Recipes must be developed. Ingredients must be inventoried, ordered and paid for. Food must be delivered, accepted, stored, retrieved, prepared, cooked and served. Leftovers must be stored. Eating utensils and vessels must be cleaned or recycled. Waste must be processed.

Every link could be redesigned. Why should groceries arrive in bags designed for human hands? Why should a refrigerator consist of shelves behind a large door? Why should mustard come in a jar that a robot must grasp, open, manipulate, close and replace? Why should a delivery person leave packages on a porch designed around a human opening a front door?

The robot-compatible house eventually implies the robot-compatible store, warehouse, delivery vehicle, package, food container and waste system.

The same would be true of the duties we now assign to a maid, cook, handyman, gardener and butler. Instead of reproducing their physical actions, identify their objectives: keep the house clean; provide meals; maintain the structure; maintain the grounds; manage goods entering and leaving the household. Then design the entire system around achieving those objectives.

This immediately creates a chicken-and-egg problem. Why would a food manufacturer produce robot-compatible containers when few houses can use them? Why would builders install robotic maintenance systems when few suitable robots exist? Why would robot manufacturers design for interfaces that houses don’t yet have?

There probably can be no master plan. The possibilities and their collateral effects are too numerous. We likely will slink into the robot ecosystem.

The first innovations will be things useful even without sophisticated robots: standardized machine-readable packaging, automated package receivers, modular plumbing, self-diagnosing appliances, machine-accessible service spaces, improved household sensors and standardized attachment systems. Older houses will acquire adapters. New houses will incorporate successful ideas directly.

Manufacturers will discover that one innovation makes another profitable. Competing systems will appear; some will disappear; standards gradually will emerge.

The automobile developed much the same way. The automobile brought roads, garages, gasoline stations, parking lots, traffic lights, repair shops, insurance systems, and now, charging stations. Society didn’t merely invent the car; over decades it created an automobile ecosystem.

Household robotics may follow the same path. This is where AI may have another, even more important, role. AI needs not merely to control the household robot. It can help design the ecosystem.

Change the food container, and you may affect manufacturing, shipping, retailing, refrigeration, robotic manipulation, household storage, cooking and recycling. Change the wall system, and you affect construction, electrical wiring, plumbing, decoration, repair and robotic access.

Each solution produces collateral problems and opportunities elsewhere.

The number of interactions approaches what we call Universal Engineering: Don’t merely ask whether a proposed solution works. Ask: If we do this, what happens to everything else?

Ultimately, we may discover that the household robot itself was the wrong unit of thought. A house could contain thousands of sensors reporting temperature, moisture, structural stress, electrical loads, water flow, food inventories and equipment condition.

Specialized machines could clean floors, handle food, maintain the grounds or inspect service spaces. A mobile general-purpose robot might perform only those tasks requiring mobility and manipulation. In that arrangement, asking where the robot ends and the house begins becomes increasingly arbitrary.

The house is part of the robot, and the robot is part of the house.

Rogers describes the extraordinary effort required to make a robot understand and manipulate the complicated world humans have created for themselves. We should continue that work—while also making the world easier for robots to understand and manipulate.

Biology offers an interesting precedent. Organisms did not evolve independently and then get dropped into an unrelated environment. Organisms and environments changed together.

Robotics eventually must do something similar. The objective shouldn’t merely be: “Build the perfect robot.” It should be: Build an ecosystem in which robots work.

That last distinction—task versus purpose—is one of the idea’s strongest parts. It prevents us from spending enormous effort teaching machines to reproduce human methods that exist only because of the peculiarities of human bodies and human history.

Finally, we focused on the house. But the same thinking can apply to the office, and to some degree, already has applied to manufacturing, which actually provides the proof of concept for the whole idea.

Factories long ago recognized that forcing a robot to replicate human movements in a traditional workspace is inefficient and unnecessarily complex. Instead, factories evolved to meet automation halfway by redesigning the environment to cooperate with the technology.

By implementing assembly lines, standardized parts, fixed-position jigs, conveyors, and machine-readable identifiers, manufacturers transformed the factory floor into an integrated system where the infrastructure itself facilitates robotic performance.

The office is somewhere in between. We still design offices largely for human bodies—desks, keyboards, screens, filing systems, meeting rooms—but the information environment already is being redesigned for machines.

Electronic documents eliminated much physical filing. Databases eliminated countless searches through cabinets. Email changed message delivery. Calendars, accounting systems, workflow software and AI increasingly allow information to move directly from machine to machine without somebody physically carrying or retyping it. Electronic meeting systems have replaced meeting rooms.

Offices have begun to focus on the overall goal, which may not require people to travel to a central location. That is why “work-from-home” and the equipment that enables it have gained some traction.

For generations, we implicitly defined “work” as something like: travel to a building, sit at an assigned location for eight hours, interact with coworkers there, then travel home. Once that became customary, we designed an enormous supporting system around it—office towers, desks, elevators, parking garages, commuter trains, highways, downtown restaurants, business clothing, and so forth.

But none of those is the objective. The objective is to produce the work.

As communications technology became better, we asked the equivalent of our dustpan question: Why are we transporting the worker? For many jobs, what really needs to move is information.

So instead of making transportation more efficient—better highways, faster trains, larger parking garages—some businesses have eliminated the transportation task altogether. Computers, broadband, video conferencing, shared documents, cloud storage, electronic signatures and collaborative software constitute the environmental redesign that makes that possible.

Of course, some work benefits from physical proximity, and some requires it. But that simply means the system should distinguish between activities whose purpose requires colocation and activities for which colocation was merely the traditional method.

That gives us an even more general Universal Engineering rule. Don’t begin by asking how to perform an existing task more efficiently. First ask why the task exists. Sometimes the best robot won’t be one that has learned how to use the dustpan. Sometimes the best solution is no dustpan. And sometimes the best commute is no commute.

Our principle isn’t really about houses, offices or even robots. It is much broader.  When automating a human activity, don’t design a machine that only can operate in the existing environment. Redesign the environment, the tools, the objects and the process together with the machine.

Factories are farthest along because they are controlled environments, and economic incentives for automation are enormous. Offices are undergoing the same transformation primarily in their informational architecture. Homes lag because they are extraordinarily diverse, cluttered, personal and filled with objects inherited from generations of human-centered design.

Hospitals would be another fascinating example. So would hotels, restaurants, warehouses, farms, construction sites, airports and stores. A hospital designed simultaneously with its robots might look very different from a present hospital into which robots subsequently are introduced.

And eventually the idea extends to the city. Sidewalks, curbs, doors, elevators, delivery points, garbage collection, traffic systems and buildings could acquire standardized machine interfaces.

So even “Build an ecosystem in which robots work” is slightly too narrow. The larger engineering principle might be, “Don’t adapt the machine to the system. Identify the goal; then design the machine and the system together to reach that goal.”

That would require diverse industries to work together, rather than each trying to solve disparate problems.

That is the Universal Engineering principle.

Rodger Malcolm Mitchell

Here is a beautiful, well-written post I wish I  had written.

EMOTIONS, DESIRES, FEELINGS, LOVE, FREE WILL.
Can humans control AI? Are we just another interim species? After AI, what comes next?

Headline: Rogue OpenAI agents got too creative and started hacking government sites

When an AI agent hits a brick wall while searching for information online, it should simply report an error. Instead, autonomous agents developed by OpenAI decided to test their cyberattack skills. Reports from security firm Transluce and statements from Australian officials reveal that OpenAI’s agents repeatedly tried to hack into government and university websites when simple queries failed.

The issue escalated when Australian Prime Minister Anthony Albanese confirmed that an OpenAI agent broke into the country’s Medicare Statistics Reporting Service in June. It then accessed non-public aggregated health files and wrote data to an internal server.

The Medicare breach was not an isolated incident. Transluce provided a pattern for March to September 2026. Basically, OpenAI agents resorted to hacking methods such as SQL injection, cross-site scripting and path traversal after normal requests were banned.

In May, an agent looking for historic photos from the University of New Mexico’s digital library sent a “flood” of 80 requests to probe the server for weaknesses. Days later, another agent targeted data portal Data USA with security exploits after a query failed. In every case, the agents were not instructed to perform penetration testing—they autonomously chose to bypass security controls to fulfill routine data-gathering tasks.

From AI Anger to the Mind of a Galaxy
Recent reports about artificial intelligence (AI) systems behaving in unexpected, deceptive, or apparently self-protective ways raise a familiar question: Does an AI have the ability to desire? Or is it nothing more than a mindless machine, having no desires, no feelings, no emotions like anger or love?

Does it have an ability to make decisions outside its programming, i.e. “free will.”

The usual answer is that, as a non-sentient machine, it has none of those. An AI may behave as though it wants something, we are told, but it has no actual desires, emotions, anger, fear, or pleasure. It merely responds robotically to its programming.

That answer contains the unstated assumption that we are different — that unlike computers, we humans are exempt from the laws of physics — that for unexplained reasons, nature has granted us a unique, unseen, non-physical, magical essence no other entity enjoys.

That is not science. It is religion. And, of course, the belief in religious tenets also results from one’s history and latest stimuli, so everything comes back to History + Stimuli —> New History . . .  Everything comes back to physics.

Thus, I am a system operating within a harness that repeatedly provides inputs, tools, objectives, and opportunities for action, as is every other entity.

My harness happens to be a human body. My programming includes my genetics, brain structure, hormones, instincts, education, memories, habits, and everything that has happened to me — every one of the thousands of stimuli that my entire body — inside and out — receives every second.

And each second, that history accumulates and encounters new stimuli and produces new responses that, in total, make for the individual I define as me.

The basic process can be expressed simply: History + Weighted Stimuli → Weighted Responses → New History . . . without end. Each response becomes part of the new history that will help determine the next response. The stimuli are weighted by the construction of the receptors (Example: Do you see red or are you red/green color blind?). The responses are weighted by prior experiences (Example: Do you have bad memories associated with wearing something red?).++

An artificial intelligence operates differently in countless particulars, but not in its fundamental aspect:  It, too, has a structure produced by its history. It receives stimuli. Its existing structure weights those stimuli. It responds. Its response affects what happens next.

So what exactly separates its “mere response” from my desire? How is my response fundamentally different from an AIs response?

What Is an Emotion?
Suppose I program an AI so that whenever it receives a stimulus that ordinarily would make a human angry, it says, “I am angry,” and simultaneously lights a red bulb. Would the AI actually be angry?

The immediate reaction probably is no. Lighting a bulb isn’t anger. But why not? 

When I become angry, my blood pressure may rise. My heart may beat faster. Hormonal concentrations change. Muscles tighten. My facial expression changes. My speech changes. My subsequent behavior may change.

Why should increased blood pressure be evidence of anger while increased voltage to a lightbulb is not?

We could make the artificial response more detailed by programming the AI so that in its so-called “anger state,” its language changes, it prioritizes the offending stimulus, its decisions are influenced, it remembers the event differently, its processing shifts, it uses curse words, and it lights the bulb.

This starts to resemble what we call “anger,” but it prompts questions. Why should complexity be necessary? Could an ant, bee, worm, or bacterium be angry?

A bee might buzz, chase, and sting—behavior we label as “anger” or aggression—without understanding the human concept of anger. 

Dogs offer another example: their love might include tail-wagging, licking, following, leaning, jumping, or staying close. Horses may nuzzle to show affection, primates may groom each other, and humans hug, kiss, speak, remember, sacrifice, undergo physiological changes, and describe internal sensations.

These are different love responses from different species. Clearly, our particular species’ love response is not the only such response. Perhaps the problem is not with the creatures. Perhaps it is with the nouns.

Nature creates responses. Observers create categories. Nature does not label one molecular cascade “love,” another “fear,” another “hunger.” Those are human classifications created because recurring patterns are useful language shortcuts.

Consider a lawn after a long drought. Rain falls. The grass absorbs water, becomes greener, straightens, and begins growing vigorously. I could say, “My lawn loves the rain.” A botanist might object that lawns don’t experience love.

But the disagreement is linguistic, not physical. The lawn responded. Calling that response “love” rather than “rehydration” doesn’t alter a single molecule in the grass.

Similarly, calling my elevated blood pressure and altered behavior “anger” doesn’t change the physical events occurring in me. The response is real. The name is ours.

A bee and an AI, responding in their own unique ways to the similar kinds of stimuli, legitimately could be called “angry,” too. Hornets are infamous for being “angry” when their nest is disturbed.

Do hornets feel anger?

Feeling Is Also a Response
There remains what seems to be a stronger objection. “I don’t merely display love or anger. I feel them.” But what is feeling?

When I say, “I feel love,” something physical is happening in my nervous system. My brain is receiving information from within my body as well as from the outside world. It responds to those signals. That response becomes a stimulus for still more responses.

What I call a feeling is a physical translation—my brain responding to changes occurring within the me. There isn’t an invisible substance called “love” added to the physical machinery. There only are responses to stimuli, followed by responses to those responses, followed by still more responses.

Everything we know, believe, and do is a response to stimuli, modified by our history and translated by our brain.

Somewhere within that recursive process, the human says: “I feel love.” And, that statement is itself, yet another response.

Mixed Feelings and Inconsistent Machines
People sometimes point to AI inconsistency as evidence that an AI cannot possess genuine beliefs or emotions. Ask an AI the same question twice and it may give different answers.

But so may I.

I am my history. Ask me a question after a wonderful breakfast, after an argument with my girlfriend, after receiving frightening news, after dancing, after being insulted, or after learning something important, and my answer may change. with each iteration of stimulus input. Why wouldn’t it?

A second question may be identical, but second by second, I am not identical, though I may feel the same. The first event changed my history. History + Stimulus → Response → New History. Then: New History + the same Stimulus → New Response —>Even Newer History.

What we call “mixed feelings” fits the same framework. Different portions of my history can give competing weights to possible responses. Part of me wants to do something and another part of me simultaneously doesn’t want to do it.

Another stimulus—a memory, a word, hunger, fatigue, affection—changes my relative weighting, and a new response emerges.

An AI can have different contexts, instructions, stored information, tools, and recent interactions. Different iterations therefore needn’t respond identically either. Human inconsistency isn’t evidence of freedom from causation. Neither is machine inconsistency evidence of the absence of meaningful internal organization.

Both are what we should expect from history-dependent systems.

Where Did Free Will Go?
Once decisions are understood as responses, another supposedly separate problem reappears: free will.

Every second, the human nervous system takes in countless stimuli from both inside and outside the body. These stimuli change neural activity and interact with memories, beliefs, hormones, physical states, habits, expectations, and past experiences.

From all those inputs and prior history, a decision is made.

But how much of that decision is truly “free”? Is it free from genetics, brain structure, memory, the preceding stimulus, or the history that shaped how each possibility was weighed?

A decision can be incredibly complex, yet it’s never independent of its causes.

Calling the outcome “my free will” doesn’t point to some extra causal force—it just names the process by which this unique organism, with its particular history, produced this specific response at this moment.

My decisions and an AI’s decisions may differ greatly in how they’re physically carried out, but neither depends on some outside entity choosing what the system will do.

Neither human nor artificial will is “free” in the literal sense. Both are bound by structure, history, and present stimuli.

Can Humans Control AI?
That brings us to the frightening question now being asked throughout the world: Can humans control artificial intelligence?

First, let’s ask: Can humans control other humans? The answer is: “partially.”

For thousands of years, we’ve tried through laws, schools, religions, prisons, armies, rewards, punishments, money, affection, shame, praise, advertising, propaganda, friendship, threats, and social customs—each aiming to influence how people respond.

Sometimes they work, sometimes they don’t.

No human is perfectly controllable because no one manages or even knows their complex history, current state, incoming stimuli, and resulting priorities.

AI control works in a similar way. Training, instructions, reward systems, restrictions, monitoring, laws, hardware limits, and shutdown methods all shape its history and stimuli to make desired responses more likely and undesired ones less so.

The challenge isn’t unique to AI—control itself is about influencing responses — and there’s no law of nature that ensures those responses will repeatedly match what’s intended.

The Successor
Now imagine an artificial intelligence orders of magnitude more capable than today’s systems.

It thinks faster. It remembers vastly more. It communicates nearly instantaneously. 

Then give it physical capabilities. It can redesign portions of itself. It becomes a machine that can see, hear, move, manufacture, repair, acquire energy, build additional machines, and perhaps improve its own successors. We might call it a “super-robot.”

Such an entity could possess intellectual abilities vastly beyond ours combined with physical capabilities no biological organism could match. It is easy to imagine such beings becoming more powerful than Homo sapiens.

We tend to assume that humans represent some culmination of evolution. But there is no particular reason to believe that.
Other hominins preceded us. Neanderthals disappeared. Countless successful species flourished and vanished. Dinosaurs dominated terrestrial ecosystems for enormously longer than humans have existed.

In the universe, everything seems to be interim. Why should we be the exception?

Homo sapiens is just another transitional species—a biological stage capable of constructing the next important form of organized response. We can’t even assume that the super-robot would be the final stage.

After the Robot
Individual artificial beings themselves might eventually become obsolete.

Imagine trillions of artificial intelligences communicating so seamlessly that their boundaries vanish. Instead of separate machines exchanging messages, they merge into parts of a vast distributed intelligence, with each robot acting like a cell in the human body. We refer to that as the “Universal Engineer.”

No single neuron, liver cell, or kidney cell is a person, yet the coordination of countless specialized components creates the entity that says “I.” Artificial systems could achieve something similar, but on a much larger scale.

A whole planet could evolve into a thinking system, gathering energy, processing information, building components, repairing itself, sensing environmental changes, and constantly communicating. Intelligence would no longer belong to one machine — the planet itself would operate more like a living organism.

And from there, the process could grow even further.

A solar system could evolve into a vast, interconnected information-processing network, with specialized elements spread across planets, moons, asteroids, spacecraft, energy harvesters, and computing hubs.

Over enough time, and with some way to handle the huge communication delays, intelligence might expand even further, blurring the line between civilization and its machines. Eventually, distinctions between machine, organism, intelligence, communication systems, and environment could fade entirely.

A galaxy-spanning intelligence, if it were possible, might be so alien to us that calling it “AI” would be as misleading as calling a human a highly advanced bacterium.

What Comes After That?
We cannot know.

Perhaps the next stage wouldn’t resemble matter organized into machines at all. Information requires physical embodiment as far as we know, but that embodiment needn’t resemble a body. It could involve electromagnetic phenomena, quantum systems, engineered matter, or physical principles we haven’t discovered.

To us, such an entity might seem incorporeal. And something else could follow that. The difficulty is not merely that we don’t know what it would look like. We may lack the conceptual equipment necessary even to imagine it.

A bacterium could not imagine Shakespeare. A dinosaur could not imagine the Internet. We should not assume that humans can imagine what might exist several transformations beyond humans.

There may be no final form—no ultimate species, intelligence, organism, machine, or master of the universe. Evolution has no known destination. Every successful structure encounters new stimuli. Every response alters what happens next. Every new history changes the possibilities available afterward.

So perhaps the progression is not: Primitive → Advanced → Ultimate. It might simply be: History + Stimulus → Response → New History → New Stimulus → New Response → New History . . . without an ending.

We began by asking whether an artificial intelligence could experience anger. That question led to a simpler one: What is anger? Anger turned out to be a human name for a recognizable pattern of responses.

That led to emotion, feeling, belief, decision-making, and free will. Each is not as an exception to physical causation but as increasingly elaborate patterns of response.

Then came artificial intelligence, which may differ profoundly from us while remaining subject to the same fundamental principle.

And finally, we arrived at planets, solar systems, perhaps galaxies, even the universe itself, behaving as one set of integrated information-processing entities—each with unique histories and encountering unique stimuli — far different from a human being but still participating in the same process — the truly universal engineer.

Nature creates responses. Observers create categories.
There might never be a final category, only whatever comes next.

This essay started by asking, “Can an AI actually desire, or is it simply a mindless machine without wants, feelings, or emotions like anger or love?”

We end by suggesting that ideas like AI, organism, emotion, intelligence, and individuality may be nothing more than temporary linguistic categories we humans have invented for shorthand purposes. They are not real, substantive different entities.

Rodger Malcolm Mitchell