AI Judgment: What Humans Should Do With an Almost Unlimited Machine
September 2, 2026
AI capability is not the same as AI judgment. As machines become capable of producing almost anything on demand, the scarce resource may become the human ability to decide what should actually be done.
Part 2
The first part of this argument ↗ examined a basic distinction: AI capability is not the same as AI judgment.
An AI system can know enormous amounts, reason through complicated problems, generate excellent answers and execute instructions at extraordinary speed. None of these capabilities automatically answers a more important set of questions: What deserves attention? How much effort does a problem deserve? Which possibilities matter? When should reasoning stop? What should be delegated? What should remain under human control? Is the objective itself worth pursuing?
As AI moves from answering questions to conducting research, writing software, operating tools, controlling machines and interacting with the physical world, these questions become more important than raw capability.
The question therefore changes from:
What can AI do?
to:
What should AI do?
And eventually:
What is worth doing?
1. More Thinking Is Not Necessarily Better Thinking
Recent research from Google DeepMind on AI “overthinking” provides a useful example.
Reasoning models can spend enormous inference effort on simple problems. The research found cases where models took several times longer on straightforward tasks without substantial accuracy gains. It also identified patterns such as Explorer, where a model continues exploring alternative paths, and Late Landing, where it has effectively reached the answer but continues reasoning before settling on it.
The problem is not reasoning itself. The problem is failing to recognize when additional reasoning has stopped producing useful returns.
A good engineer does not perform a complete failure analysis before moving a chair. A good mathematician does not treat every calculation as a research problem. A good driver does not continuously recalculate the entire road network while travelling down an empty highway.
The ability to think is different from the ability to decide how much thinking a problem deserves.
That is judgment.
2. Not Every Problem Is P vs NP
Most interactions with AI are nowhere near the frontier of human knowledge.
People ask AI to rewrite emails, explain charts, summarize papers, compare alternatives, translate documents, fix code, calculate numbers and organize information.
Very few prompts are attempts to solve P versus NP.
There is another useful observation here. Human thinking is often more predictable than it appears. Give several intelligent people the same information, tools and problem, and many will explore similar conceptual paths. Differences emerge through experience, depth, persistence and the ability to notice something unusual.
This creates a large amount of intellectual duplication.
Research repeatedly rediscovers intermediate results, approaches problems that have already been explored and spends time reconstructing knowledge that exists somewhere in the literature.
The P versus NP discussions provide a small illustration. Independent reasoning can arrive at results that turn out to have already been established.
AI can potentially act as a pre-discovery filter:
“This line of reasoning has already been explored. Here are the relevant results. Here is what remains unresolved.”
That does not replace thinking. It prevents wasted thinking.
AI may not make research think less. It may make research waste less thought.
3. The 99% and the 1%
Research can be viewed as having two very different components.
The first is the foundation: existing knowledge, literature searches, measurements, calculations, replication, comparison, filtering, classification, refinement, testing, elimination of possibilities and recombination of existing ideas.
Call this 99% as a conceptual model, not a measured scientific ratio.
The remaining 1% represents the disproportionately valuable moment when an unexpected connection appears.
An anomaly suddenly makes sense.
Two apparently unrelated observations turn out to belong together.
An accepted assumption is discarded.
A new explanation appears.
Eureka.
The asymmetry is remarkable. The foundation may require years of work, while the conceptual jump may occur in seconds. Yet that small jump can reorganize an entire field.
This distinction matters when considering what AI will do to research. AI does not need to produce every Eureka moment to transform discovery. It can compress much of the work surrounding the Eureka moment.
4. Induction, Deduction and Abduction
Three forms of reasoning help illustrate the distinction.
Induction extracts patterns and generalizations from observations.
Deduction derives consequences from established premises.
Abduction asks what explanation could account for an observation, particularly when existing explanations fail.
AI is naturally powerful at induction and deduction. It can process vast quantities of information, identify patterns, calculate consequences and explore combinations at extraordinary speed.
Abduction is more interesting.
Scientific breakthroughs often begin with something that does not fit the existing model.
The important question becomes:
Why doesn’t this fit?
Then comes the more disruptive question:
What if the existing model is wrong?
AI may eventually become capable of such abductive jumps. There is no reason to assume that it cannot. But even before that happens, AI can radically accelerate the preparation that makes such jumps possible.
5. Research Is a Fire
Research resembles building a fire.
The preparation requires gathering material, selecting useful fuel, removing what is unsuitable, arranging it properly and creating the conditions for ignition.
Most of that work is tedious.
Then something changes.
Smoke becomes an ember. The ember becomes a flame. A collection of separate pieces suddenly becomes a functioning whole.
A scientific insight often behaves in the same way. The breakthrough appears suddenly, but it depends on the organization of everything that came before it.
AI can perform much of the intellectual preparation at extraordinary speed:
searching
collecting
filtering
comparing
classifying
connecting
testing
reorganizing
Once an insight appears, AI can also help develop it by testing consequences, running simulations, finding exceptions, comparing explanations and exploring applications.
The future research process does not have to be divided into “human discovery” and “AI execution.”
A more realistic model is:
AI prepares the intellectual material → human and AI explore → an unexpected connection appears → AI rapidly develops the consequences.
AI may not replace the scientist who discovers the fire. It may become the machine that gathers the world’s firewood.
6. Collaborative Generation
The same pattern already appears in AI-assisted intellectual work.
An idea can begin as a small observation. AI can expand it, challenge it, connect it to other ideas, search for evidence, generate alternatives, identify contradictions, reorganize the argument and turn it into an article or research direction.
The AI is not necessarily the origin of the seed. It becomes the mechanism through which the seed is explored and developed.
This is better understood as collaborative generation than simple AI generation.
The human contribution is often the initial direction, intuition or question. The AI contribution is the enormous expansion of the intellectual search around that seed.
The result can be more than either component could produce efficiently alone.
The human provides the seed. AI provides the care and growth.
7. AI Diverges. Human Judgment Converges.
This may be the most useful way to describe the emerging division of labour.
AI is extraordinarily good at divergence.
One idea can become dozens of alternatives, hundreds of connections and thousands of possible paths. AI can keep expanding the possibility space at very low marginal cost.
Human judgment performs a different function: convergence.
Which possibility matters?
Which one is useful?
Which one fits the larger objective?
Which one has acceptable consequences?
Which one deserves further investment?
AI can explore almost indefinitely. Human attention cannot.
As AI capability grows, the value of selection may therefore increase rather than decrease.
AI expands the search space. Human judgment defines the value space.
8. The Laser-Focus Principle
AI can be thought of as an extraordinarily powerful energy source. Human judgment is the focusing mechanism.
A laser becomes powerful because energy that could otherwise spread in every direction is concentrated on a very small target.
AI has a similar problem. It can generate thousands of possibilities. The important question is not whether those possibilities are possible. It is whether they are useful.
AI can produce ten thousand solutions. Human judgment asks: Which one?
AI can identify a thousand problems. Human judgment asks: Which problem is worth solving?
AI can generate a hundred strategies. Human judgment asks: Which strategy serves the objective?
More capability does not automatically produce more value.
Possibility is not usefulness.
AI provides the energy. Human judgment provides the focus.
9. The Ferrari Principle
Imagine a self-driving Ferrari.
On a straight, well-marked highway, constant human intervention makes little sense. The machine can drive.
The interesting moments are the sharp turns, narrow streets, unexpected obstacles, contradictory signals and unfamiliar situations.
The human does not need to control every kilometre. The human needs to become especially important when the situation stops being ordinary.
This suggests a better model of human oversight.
Instead of:
AI → human → AI → human → AI → human
the future may look more like:
AI → AI → AI → AI → unusual situation → human judgment → AI
Human involvement becomes event-driven rather than continuous.
The human does not need to supervise every operation. The human needs to recognize which operations deserve supervision.
Human-in-the-loop does not have to mean human-at-every-step.
And this is why not every AI task requires enormous reasoning. Most roads are straight. Not every problem is P versus NP.
10. Verification May Become Like Grammar Checking
The same principle applies to verification.
If AI can generate thousands of claims, documents or analyses, requiring humans to verify every sentence will not scale.
But verification itself can become automated.
Grammar software already checks enormous amounts of text while humans focus on unusual or questionable cases.
AI verification could follow a similar model. A system generates an output. Another system checks it. High-risk or uncertain claims receive additional scrutiny. A statistical sample is independently checked. The verifier’s historical error rate is measured. Humans investigate exceptions.
The goal is not to verify everything manually. The goal is to allocate verification effort intelligently.
The scarce resource then becomes neither generation nor routine checking. It becomes the ability to determine what deserves checking.
11. Collaborative Verification
AI-assisted writing provides a useful example.
A collaborative process does not require every generated sentence to be independently checked. There are several layers of validation.
The discussion establishes the underlying argument.
AI converts that argument into prose.
Selected passages are examined closely.
The output is compared with the reasoning that produced it.
The important questions become: Did the AI preserve the idea? Did it distort the conclusion? Did it introduce an unsupported claim? Does the argument remain coherent? Does the final article actually represent the intended thought?
This is different from simple fact-checking. It is semantic verification.
The conversation itself becomes part of the validation layer. That may become increasingly important in collaborative AI systems.
12. AI Changes the Meaning of Job Displacement
The obvious AI story is that machines perform tasks previously performed by people.
But the deeper transformation may involve several different levels.
First: AI performs the task.
Second: AI replaces the worker performing the task.
Third: AI redesigns the product so the task is required less often.
Fourth: AI changes the experience so the physical activity itself becomes optional.
The third and fourth categories are easy to miss when employment statistics are used as the primary measure of AI impact.
13. The Job Can Disappear Without Being Automated
Consider an air conditioner.
A future system might contain better components, self-diagnosis, predictive maintenance and increasingly effective self-cleaning.
Fewer things break. Fewer service visits are required. Less maintenance is necessary.
AI has not replaced the technician directly.
AI has helped make the technician less necessary.
The same principle could apply to refrigerators, washing machines, vehicles, industrial equipment and many other products.
AI does not have to automate a job to reduce demand for the job. It can engineer the job out of existence.
14. The Self-Driving Car That Services Itself
Consider the service chain around a vehicle.
A fault occurs. The owner notices it. The owner contacts a service center. A technician diagnoses the problem. The vehicle is transported or driven there. The repair is performed. The owner collects the vehicle.
A sufficiently autonomous vehicle could eventually detect the problem, diagnose it, arrange service, travel to an appropriate facility, undergo robotic maintenance and return.
The driver disappears from the process.
Other human roles may gradually disappear as well.
This is second-order automation.
AI is not simply doing the existing job. It is redesigning the system around the job.
15. Planned Obsolescence May Plan Its Own Obsolescence
For decades, a familiar product cycle has been:
make → sell → wear out → repair or replace → sell again
Better AI-assisted engineering could push products in another direction:
make → use for a long time → diagnose automatically → maintain automatically → upgrade
More durable components, predictive maintenance, self-diagnosis, self-cleaning and longer service intervals could reduce the need for repeated human intervention.
Commercial incentives will still matter. AI does not automatically eliminate planned obsolescence. But increasingly capable design systems can make durability technically easier.
There is an irony here:
Planned obsolescence may eventually have to plan its own obsolescence.
Technological progress has another interesting effect.
The latest generation is expensive. Those who can afford it upgrade. The previous generation becomes cheaper. It moves into a wider market. Then another generation arrives at the top. The process repeats.
Those who can pay get the newest version. Those who cannot may eventually receive yesterday’s breakthrough at a much lower cost.
Yesterday’s luxury becomes tomorrow’s ordinary.
This can produce an unusual combination: greater inequality at the technological frontier while technological capability spreads across society.
The frontier may become more unequal while the technological baseline continues to rise.
17. From Physical Automation to Physical Substitution
AI combined with robotics, autonomous vehicles, augmented reality, virtual reality, spatial computing and increasingly sophisticated displays could produce another transformation.
Instead of merely automating physical work, technology can reduce the need to physically move people and objects for certain experiences.
Imagine a room becoming Paris.
Not literally Paris, but an increasingly convincing representation of Paris generated through immersive technology.
VR says: Put the user somewhere else.
AR says: Put something else here.
Holographic and spatial systems can potentially make digital objects appear increasingly present in physical space.
The important question is not whether virtual reality will literally become better than physical reality in every respect.
The question is:
At what level of fidelity does physical presence become unnecessary for a particular experience?
That threshold matters economically.
Travel today involves aircraft, hotels, taxis, restaurants, attractions, guides and shopping. Some future experiences may require much less physical infrastructure.
Physical reality will not disappear. But some experiences that currently require physical movement may become optional.
18. The Restaurant Is Different
A restaurant illustrates the limits of simple automation.
Food can be automated. Recipes can be optimized. Cooking can be robotic. Menus can be personalized.
But a restaurant is more than food.
atmosphere
service
social interaction
smell
sound
timing
location
the experience of being there
AI may reproduce the food long before it reproduces the entire reason people visit a restaurant.
Yet even here, AI could transform the experience. A sufficiently capable system might know that a customer would probably love a dish they have never tried, or could continuously personalize the menu, environment and service.
That is not necessarily replacement.
It is experience engineering.
19. Second-Order Automation
The AI transformation therefore goes beyond the question:
Which jobs can AI perform?
A more useful sequence is:
AI performs the task.
AI changes how the task is performed.
AI changes the product so the task is needed less.
AI changes the experience so the physical activity becomes optional.
This is why occupational automation statistics may capture only part of the eventual transformation.
A repair technician can disappear because an AI system performs repairs. But the technician can also disappear because the product no longer needs frequent repairs.
A travel agent can disappear because AI plans travel. But tourism itself can also change if sufficiently convincing experiences no longer require physical travel.
20. The Control Problem Becomes a Containment Problem
A superintelligent system inside an isolated environment is very different from the same system connected to financial systems, laboratories, industrial machinery, autonomous vehicles, communication networks and critical infrastructure.
The important transition is not simply:
intelligence → danger
It is:
intelligence → access → authority → action
A system can have extraordinary intelligence while having limited real-world impact if its access and authority are tightly constrained.
The question therefore becomes:
How much ability to act does the system actually have?
This suggests that part of the traditional AI control problem may become a containment problem.
The objective is not necessarily to control every thought. It is to control the channels through which thought becomes power.
21. Control the Domain, Not the Function
Instead of attempting to specify every possible action that an intelligent system must never take, first define its domain.
Then define its range of action.
Then define its goal.
Domain
Where may the AI operate? Which information, systems, machines, networks, people and resources are within its reach?
Range
What can the AI do within that domain? Can it read, analyze, simulate, recommend, modify, execute, purchase, communicate or deploy?
Goal
Only then comes the objective: What should it accomplish?
The hierarchy becomes:
First domain. Then range. Then goal.
Or even more simply:
Domain defines where. Range defines how much. Goal defines why.
This approach does not require perfect prediction of every strategy a highly intelligent system might invent. If a system cannot access something, its ability to affect that thing is constrained regardless of how clever it becomes.
22. Focused Containment
A powerful laser does not have to become weak to be made manageable. Its energy can be contained and its output controlled through carefully designed boundaries and apertures.
AI can be viewed similarly.
The objective is not necessarily to make intelligence weak.
It can instead be:
contain intelligence → control its interfaces → constrain its permissions → focus its capabilities → monitor its consequences
This is focused containment.
A powerful AI does not need unrestricted access to the world to perform a highly valuable task. It can have enormous intelligence while operating through a narrow and controlled aperture.
23. The New Human–AI Division of Labour
AI
Human judgment
Generate
Select
Explore
Focus
Search
Decide
Calculate
Judge
Rearrange
Interpret
Refine
Set purpose
Simulate
Evaluate consequences
Execute
Authorize
Verify routine cases
Handle exceptions
Expand possibilities
Decide what matters
These categories are not permanent laws. AI capabilities will continue to expand.
The table identifies where human judgment remains particularly valuable.
AI can increasingly answer:
How?
The harder question is:
Why?
And harder still:
Why this rather than something else?
What employers are beginning to value
The shift is not only theoretical. McKinsey’s 2026 HR Monitor shows a similar movement in the skills employers expect to matter in the future. Problem-solving rose by 9 percentage points, creativity by 6, ability to reason by 3, and digital literacy by 15. At the same time, several specialized execution-oriented skills declined in their ranking.
Source: McKinsey & Company, HR Monitor 2026, Exhibit 1, “Future skills will shift from routine task-based capabilities to skills focused on analyzing and interpreting AI outputs.” McKinsey HR Monitor Survey, January 2026, n = 1,303 HR professionals across Belgium, China, France, Germany, Italy, the Netherlands, Poland, Spain, the UK, and the US. Year-over-year differences are directional because country coverage changed slightly from 2025.
The chart does not measure “judgment” directly. But it points toward something important: as AI takes over more routine execution, the relative value of knowing how to interpret, evaluate, question and use its outputs increases.
This is precisely where the human role begins to move upward—from producing every answer to deciding which answer matters, whether it should be trusted, and what should happen next.
24. The Human Role May Become Smaller and More Important
If AI succeeds, humans may perform fewer routine intellectual tasks. Machines may perform more physical tasks. Products may require less maintenance. AI may perform much of the routine preparation involved in research. AI may generate and verify enormous amounts of information. Robots may operate more of the physical economy. Virtual environments may replace some physical experiences.
At first glance, this looks like human displacement.
There is another possible interpretation.
Perhaps the purpose of automation is not to preserve every existing human task. Perhaps it is to eliminate tasks that machines can perform better and leave humans with the responsibility for deciding what deserves to be done.
That is a very different role.
The scarce resource may no longer be execution.
It may be attention, purpose and judgment.
25. The New Scarcity: Judgment
AI makes generation cheap. It makes exploration cheap. It makes iteration cheap. It makes many forms of knowledge retrieval cheap. It makes routine reasoning cheap. It may make much physical execution cheap.
But objectives are not automatically supplied by capability.
A system can generate a million possibilities without knowing which possibility matters.
It can solve a thousand problems without knowing which problem deserves solving.
It can reason for an hour without knowing whether the first five seconds were sufficient.
It can execute an objective perfectly while the objective itself is wrong.
This is the central asymmetry.
AI can expand what is possible. Human judgment decides what is worthwhile.
26. From Human-in-the-Loop to Human-at-the-Goalpost
The human role may therefore move upward.
The human does not need to sit inside every operation.
Define the domain.
Define the permissions.
Define the objective.
Set the constraints.
Monitor exceptional situations.
Review high-consequence decisions.
Reconsider the objective itself when circumstances change.
The human becomes less of a manual operator and more of a goalpost.
This is where the earlier metaphors come together.
AI is the firewood collector.
AI is the fire-management system.
AI is the Ferrari.
AI is the laser source.
AI is the ocean of knowledge and possibilities.
Human judgment provides the focus, destination and boundary.
27. The Larger Principle
The future advantage may not belong to the person who can make AI do the greatest number of things.
It may belong to the person—or institution—that can focus enormous AI capability on the few things that actually matter.
That requires a different kind of intelligence.
Not merely knowing more.
Not merely reasoning longer.
Not merely generating more.
Not merely acting faster.
It requires knowing what to pursue, what to ignore, what to verify, when to stop, where to intervene, what to delegate and what must remain outside the machine’s domain.
AI can make the production of answers abundant. It can make the production of possibilities almost unlimited. It can make much of the intellectual and physical work surrounding discovery dramatically cheaper.
The scarce resource could therefore shift from generating possibilities to choosing among them.
That is the territory of judgment.
AI can explore the ocean. Humans decide which shore matters.
The future of AI does not have to be a contest between human intelligence and machine intelligence.
It can be a multiplication problem:
Human judgment × AI capability.
A powerful machine without direction produces enormous amounts of activity.
A powerful machine with clear human purpose can produce extraordinary leverage.
The challenge of the AI age may therefore not be learning how to make AI do everything.
It may be learning how to focus an almost unlimited machine on the things that are actually worth doing.
AI supplies the power. Human judgment supplies the direction.
