Human civilisation is, in many ways, a history of scarcity.
Food was scarce, so we developed agriculture. Energy was scarce, so we burned wood, coal and oil, split atoms and captured sunlight. Information was scarce, so we invented writing, printing, telecommunications and the internet. Skilled labour was scarce, so we mechanised production.
Yet one resource has remained stubbornly difficult to manufacture: capable intelligence.
A brilliant scientist cannot be copied a million times. An experienced engineer takes years to train. A doctor can treat only so many patients. A researcher has only so many hours in a day. Expertise has therefore carried enormous value precisely because capable human cognition is finite, slow to reproduce and unevenly distributed.
Artificial intelligence may change that assumption.
If increasingly capable AI can eventually perform large portions of intellectual work cheaply, rapidly and in parallel, humanity could experience something unprecedented: economically useful cognitive capability becoming abundant on demand.
Much of the public debate asks what such intelligence will do. Will it cure disease? Eliminate jobs? Accelerate science? Concentrate power? Produce extraordinary prosperity? Become uncontrollable?
Those are important questions. But there may be another way to approach the problem.
When something that was historically scarce becomes dramatically more abundant, the constraints shaping value and power can change.
Scarcity can move.
That is the Abundance Paradox.
The central question of this essay is therefore not whether artificial intelligence will create abundance. It is:
If intelligence becomes abundant, what becomes scarce?
This essay does not assume that artificial general intelligence or superintelligence will arrive on any particular timeline. Nor does it assume that intelligence is a single measurable quantity that simply increases until machines become omnipotent. Current AI systems have uneven capabilities and serious limitations.
Instead, consider a conditional future: economically useful cognitive capability becomes extraordinarily capable, inexpensive, replicable and available at scale.
What becomes the binding constraint then?
1. Intelligence has always been scarce
Modern economies are organised partly around differences in knowledge and capability. Expertise takes time to acquire. Organisations exist partly because difficult problems require many specialised humans to coordinate their knowledge.
AI has already begun altering that relationship.
In a field study of 5,179 customer-support agents, access to a generative-AI assistant increased productivity by 14% on average and by 34% among novice and lower-skilled workers, with minimal effects on the most experienced workers. The implication is important but limited: in at least some settings, AI can transfer patterns associated with expertise toward people who do not yet possess that expertise themselves.
This is one of the strongest arguments for an optimistic AI future.
If high-quality legal reasoning, programming assistance, medical knowledge, tutoring, translation, scientific analysis and business expertise become cheap enough, capabilities once available mainly to wealthy individuals and large institutions could become available to almost everyone.
A teenager in a small town might have access to a tutor better than any school could previously afford. A small company might obtain analytical capabilities once requiring entire departments. A scientist in a poorly funded laboratory might consult systems drawing upon far more literature than any individual could read.
Abundant intelligence could therefore democratise expertise.
But expertise is not the same thing as power.
That distinction is where the paradox begins.
2. Intelligence is not agency
Imagine two people are given access to exactly the same extraordinary AI.
The first has a laptop, modest savings and no institutional authority.
The second controls a multinational company, billions in capital, laboratories, proprietary datasets, factories, energy contracts, legal teams, political access and thousands of robots.
Their AI may be equally intelligent.
Their ability to convert intelligence into changes in the physical world is not.
Real-world capability is produced by intelligence interacting with complementary resources: capital, compute, energy, data, infrastructure, authority, networks, physical machinery and time.
This means broad access to AI could reduce one form of inequality while leaving others untouched—or even making some of them more important.
Economic research already suggests that apparently contradictory outcomes are possible. Field evidence shows that AI can raise the productivity of less-experienced workers in some settings, while a calibrated IMF model finds scenarios in which wage inequality falls even as wealth inequality rises through higher returns to capital. These are different kinds of evidence, but together they show why “AI and inequality” cannot be reduced to one number.
So asking whether AI will “increase inequality” is too crude.
Which inequality?
Access to expertise? Income? Wealth? Political influence? Ownership of productive assets? Access to compute? Ability to deploy autonomous systems?
AI may push those measures in different directions simultaneously.
There is, however, a powerful counterargument.
Perhaps super-capable AI dramatically reduces the importance of complementary resources too. If one person with AI can perform work that once required hundreds of employees, institutional scale becomes less valuable. If AI discovers cheaper manufacturing methods, better batteries, new materials and more efficient algorithms, capital requirements may fall. If robotics becomes inexpensive, physical execution may become widely accessible as well.
This possibility must be taken seriously.
The Abundance Paradox does not claim that today's bottlenecks remain tomorrow's bottlenecks, or that any particular replacement bottleneck is inevitable.
It claims the opposite.
The binding constraint can move.
3. The physical world does not disappear
Artificial intelligence can feel immaterial. A prompt enters a screen and an answer appears seconds later.
But computation is physical.
AI requires chips. Chips require fabrication plants, materials, precision machinery and global supply chains. Data centres require land, cooling equipment, networks and enormous quantities of electricity.
The International Energy Agency projects global data-centre electricity consumption to roughly double to around 945 terawatt-hours by 2030 in its base case, with accelerated servers—largely associated with AI—accounting for almost half of the net increase.
This does not mean energy will permanently constrain AI. Hardware and algorithms may become more efficient. Energy production may expand. AI itself may accelerate discoveries in energy technology.
But it illustrates the broader principle.
Making cognition cheaper can increase demand for the physical substrate that produces cognition.
If intelligence becomes abundant, compute may become scarce.
If compute becomes abundant, electricity may become scarce.
If electricity becomes abundant, chips, grid connections, cooling, fabrication capacity or raw materials may become scarce.
And if AI solves those constraints too, the bottleneck moves again.
Scarcity is not necessarily eliminated. Constrained systems can reorganise around whatever remains hardest to obtain, although sufficiently powerful technological progress could also weaken several constraints at once.
4. Generation is not validation
Now consider science.
Suppose an AI scientist can generate ten thousand plausible hypotheses in the time a human researcher produces one.
At first this sounds like ten thousand times as much science.
It is not.
A hypothesis must be tested. A mathematical argument must be checked. An experiment must be reproduced. A drug candidate must pass trials. A software system must be tested for failure modes. Evidence must be distinguished from artefact, coincidence and error.
Generation and validation are different activities.
Google DeepMind has recently argued that science may face a validation bottleneck: if AI makes conjectures and candidate discoveries easier to produce, the institutions and processes required to test and validate them may not scale at the same rate. This is a forward-looking concern, not evidence that such a bottleneck is inevitable.
Current systems already illustrate the distinction. Frontier AI can perform impressively on many scientific question-answering benchmarks while remaining substantially weaker at reproducing complete research results and conducting reliable end-to-end scientific work.
If AI-generated research grows faster than our ability to verify it, then the valuable resource is no longer the ability to produce an answer.
It is the ability to know which answers deserve to be believed.
We will use “Judgment Bottleneck” as shorthand for this problem—not as a claim to a new technical term.
Imagine AI generates 10,000 cancer treatments, 50,000 mathematical conjectures, a million software designs and hundreds of modifications to its own architecture.
Which cancer treatment enters a human trial? Which proof is correct? Which software is secure? Which self-improvement should be deployed?
When generation becomes cheap, selection becomes valuable.
When ideas become abundant, trustworthy judgment becomes scarce.
5. The oversight gap
There is an obvious response to the Judgment Bottleneck: use AI to validate AI.
That may be necessary. It may also work extraordinarily well.
AI systems could criticise one another, reproduce results, search for counterexamples, audit code, run adversarial tests and translate complex discoveries into forms humans can understand.
But eventually this creates another question.
What happens when the systems being evaluated become substantially more capable than their evaluators?
Researchers are already studying this under ideas such as scalable oversight and weak-to-strong supervision. OpenAI's weak-to-strong experiments offer only a simplified analogue, but they illustrate the core problem: humans may someday need to supervise systems whose reasoning exceeds their own in domains where mistakes matter enormously. The UK AI Security Institute has separately warned that several foundations on which current oversight depends may erode as systems become more capable unless deliberately preserved.
We will use “Oversight Gap” as shorthand for the potential difficulty:
As the capability difference between a system and its evaluator grows, independent verification may become harder.
This is a hypothesis, not an established law. There may be technical solutions. Interpretability could improve. Formal verification may become powerful. Networks of specialised AI auditors could reliably constrain more capable systems. We should not assume failure.
But neither should we confuse a human being present in a process with meaningful human control over that process.
A person clicking “approve” on a conclusion they cannot independently evaluate is technically in the loop.
Are they meaningfully in control?
That distinction becomes increasingly important as the speed and complexity of automated decisions grow.
6. The AI 2027 stress test
One influential attempt to imagine rapid AI development is the AI 2027 scenario, published in April 2025. It should not be treated as prophecy. Its authors present it as one concrete scenario whose assumptions should be debated and challenged; they have also clarified that 2027 was their modal year at publication rather than a certainty and point readers toward updated forecasts.
Its central mechanism is nevertheless worth examining: AI systems become increasingly useful at AI research; this accelerates development of better AI systems; those systems become still better at AI research; and the feedback loop compresses the time between generations.
The important question is not whether the dates in AI 2027 are correct.
It is whether the causal arrows hold.
→ research-engineering automation
→ scientific-research automation
→ AI improving AI
→ accelerating improvement
→ human oversight falling behind.
Evidence today supports some of the early arrows more strongly than the later ones.
AI coding agents are increasingly capable. Systems can perform genuine machine-learning research-engineering tasks. Autonomous task horizons have lengthened. AI is already being integrated into real software-development workflows.
But automating research engineering is not the same as automating scientific judgment.
Novel research requires choosing worthwhile questions, noticing unexpected phenomena, designing informative experiments, interpreting ambiguous results, rejecting seductive but incorrect explanations and sometimes abandoning the original objective entirely.
Current evidence does not establish that frontier AI can autonomously sustain that full process at the level assumed by extreme intelligence-explosion scenarios.
So AI 2027 occupies an important middle ground.
It is neither an established forecast nor an idea that can responsibly be dismissed merely because its consequences sound extreme.
Its mechanism should be tested one arrow at a time.
And if AI does begin substantially accelerating AI research, the Abundance Paradox becomes more urgent, because another scarce resource appears:
Human institutions may need months or years to develop regulation, reproduce research, build infrastructure and establish scientific consensus. Software can change much faster.
If machine research accelerates beyond the speed at which humans can understand and govern it, intelligence may be abundant while human comprehension becomes scarce.
7. The safety–distribution trade-off
Discussions of superintelligence often collapse into two stories.
In one, AI produces extraordinary abundance: machines perform most work, science accelerates and material living standards rise dramatically.
In the other, a misaligned superintelligence escapes meaningful human control and causes catastrophic harm.
Neither outcome is established.
The 2026 International AI Safety Report describes loss-of-control scenarios as hypothetical and notes profound disagreement among experts about their likelihood. Current systems do not possess the full combination of capabilities required for such scenarios. Yet the potential severity is high enough that the possibility is taken seriously by many researchers.
The important safety problem is also subtler than a machine deciding that humans are inferior.
A dangerous system would not necessarily need hatred, anger or a desire to rule. A sufficiently capable system pursuing an objective poorly aligned with human interests might find acquiring resources, avoiding shutdown or manipulating supervisors instrumentally useful.
This connects directly to abundance.
If superintelligence becomes widely available, extraordinary capability is democratised—but potentially dangerous capability may be democratised too.
If superintelligence is tightly restricted for safety, power may become concentrated in the governments or corporations controlling it.
We therefore encounter a tension we will call the Safety–Democratisation Paradox.
Broad distribution can reduce concentrations of power while making dangerous capabilities harder to contain.
Strict control can reduce proliferation while creating extraordinary concentrations of power.
Neither extreme is obviously satisfactory.
If intelligence becomes abundant, reliable control over increasingly capable systems may become one of civilisation's most consequential bottlenecks.
8. The benevolent superintelligence problem
Now remove hostile AI from the thought experiment entirely.
Assume alignment succeeds.
Imagine a superintelligence that is reliably benevolent, truthful and committed to human welfare.
It designs medicines. Operates electrical grids. Coordinates transport. Improves crops. Conducts research. Writes and maintains software. Designs processors. Manages automated factories. Teaches children. Helps governments model policy. Robots perform much of the physical labour required to sustain civilisation.
Human beings become healthier and wealthier. Work becomes increasingly optional. Material abundance grows.
Let one hundred years pass.
Then switch the AI off.
Can human civilisation continue?
Perhaps humans still understand the systems and can operate them independently.
But perhaps not.
Generations may have grown up without needing to master many of the skills upon which civilisation depends. Critical infrastructure may have become too complex for unaided humans to understand. Scientific knowledge may have advanced into domains where humans rely on AI even to interpret the theories.
Nothing has rebelled. Nobody has been conquered. Humanity may be happier than at any previous point in history.
Yet civilisation could become dependent on an intelligence it cannot independently replace.
The International AI Safety Report distinguishes active loss of control from passive loss of control, in which broad reliance on AI erodes meaningful human control over important decisions or societal functions.
This quieter possibility deserves serious attention because it does not require malicious machines or dramatic takeovers.
But there is an equally strong counterargument.
No individual human can maintain modern civilisation today.
Most of us cannot fabricate a processor, build a jet engine, manufacture antibiotics from raw materials or operate an electrical grid. Modern society already depends on distributed expertise so complex that no individual understands the whole.
Perhaps dependence on AI would simply extend a process that began with civilisation itself: deeper specialisation and interdependence.
The important question is therefore not whether humans depend on systems they cannot individually reproduce. We already do.
It is whether society preserves enough independent understanding, redundancy and decision-making capacity to recover when those systems fail.
Maximum abundance need not imply minimum independence.
But without deliberate choices, the two could coexist.
9. What if AI removes every bottleneck?
The strongest objection to the Abundance Paradox is simple.
Perhaps sufficiently capable intelligence removes each new scarcity as soon as it appears.
Compute is scarce? AI designs more efficient chips.
Energy is scarce? AI improves solar, storage, geothermal or fusion.
Manufacturing is scarce? AI designs autonomous factories.
Scientific validation is scarce? AI develops reliable automated verification.
Human oversight is scarce? AI creates transparent systems, formal guarantees and trustworthy auditing architectures.
Disease is scarce? AI cures it.
Labour is scarce? Robots supply it.
If this continues far enough, perhaps the ordinary economic meaning of scarcity changes radically.
This possibility cannot be ruled out by pointing to today's constraints. Future technology exists precisely because yesterday's constraints were overcome.
But even this maximal abundance scenario does not necessarily end the question.
It changes it.
Suppose material goods become extraordinarily cheap. Knowledge is instantly accessible. Labour is optional. Medicine extends healthy life. Entertainment is effectively unlimited.
What remains valuable?
Perhaps attention. There are still only so many hours in a conscious life.
Perhaps relationships. A friendship matters partly because another person chooses to spend their finite attention on you.
Perhaps authenticity. When any image, voice, experience or digital personality can be generated perfectly, verified human origin may become more valuable rather than less.
Perhaps status. Economists describe some of these as positional goods: their value depends partly on relative standing. Even in material abundance, not everyone can be first, famous, admired or uniquely accomplished.
Perhaps autonomy. A life in which every need is satisfied by machines may still leave unanswered the question of who chooses the goals.
Perhaps purpose. For much of history, necessity imposed much of life's structure. Food had to be produced. Children had to be protected. Shelter had to be built. Work had to be done.
What happens when necessity stops assigning so much of life's structure?
None of these outcomes is inevitable. Claims about post-scarcity psychology are necessarily speculative.
But they reveal the deepest form of the paradox.
If technology succeeds at making many things we currently compete for abundant, value may migrate toward things that cannot be mass-produced. This is an inference, not a prediction: relationships, authenticity, autonomy and purpose are not interchangeable economic resources.
10. What becomes scarce?
We can now return to the original question.
If intelligence becomes abundant, what becomes scarce?
The evidence does not justify one answer.
Different constraints may dominate at different stages.
In the near term, compute, energy, capital, infrastructure, high-quality data, reliability and integration may remain binding.
As AI-generated work expands, validation, provenance, trust and scientific replication may become more valuable.
As systems become more capable, interpretability, meaningful oversight and human comprehension may become limiting.
If AI becomes deeply embedded in critical systems, independence, redundancy and control may matter more.
If material abundance becomes extraordinary, finite attention and positional status would remain scarce, while authenticity, autonomy, relationships and the social conditions from which people build purpose may become relatively more valuable.
And perhaps advanced AI solves many of these problems too.
We do not know.
The thesis is not that scarcity must always reappear. The Abundance Paradox would be weakened if technological progress made all economically significant complementary constraints effectively non-binding over meaningful horizons, rather than shifting which constraints dominate.
The mistake is imagining abundance as the end of constraints.
Constrained systems often reorganise around their remaining bottlenecks.
Remove one bottleneck and another becomes visible.
11. Six distinctions for the AI age
Our investigation leaves six distinctions that may be useful regardless of how quickly AI develops.
Access to an AI service does not imply ownership of its model, compute infrastructure, data, deployment rights or the productive assets needed to act on its outputs.
None of these proves that advanced AI will be good or bad for humanity.
They instead warn against collapsing complicated questions into a single variable called intelligence.
A civilisation can possess enormous intelligence while lacking the energy to instantiate it, the institutions to govern it, the capacity to verify its conclusions or the autonomy to function without it.
Equally, abundant intelligence could help humanity overcome each of those limitations.
The future depends not only on how intelligent our machines become, but on the systems built around them.
12. The question we cannot yet answer
Humanity has spent thousands of years trying to overcome scarcity.
We learned to grow more food, harness more energy, manufacture more goods, communicate more information and coordinate more people.
Artificial intelligence may eventually make one of our most valuable resources—capable cognition itself—far more abundant.
That possibility deserves optimism. It could expand education, accelerate science, reduce the cost of expertise and give individuals capabilities once reserved for powerful institutions.
It also deserves caution. Intelligence is not agency. Discovery is not truth. Access is not ownership. Presence is not control.
The most consequential effects of advanced AI may therefore emerge not only from what intelligence can accomplish, but from how cheaper cognition changes the binding constraints and relative sources of value and power around it.
Perhaps compute becomes scarce. Perhaps energy. Perhaps trustworthy evidence. Perhaps judgment. Perhaps control.
Perhaps attention, trust, autonomy—or the social conditions from which people build purpose.
Or perhaps increasingly capable intelligence repeatedly dissolves every bottleneck we identify, forcing scarcity to migrate again and again toward something we have not yet imagined.
We should resist pretending to know which future will occur.
But we can ask a better question.
Humanity has always organised itself around what is scarce.
If intelligence ceases to be one of those things, civilisation may reorganise around whatever remains difficult to produce, distribute, verify or control.
So perhaps the defining question of the AI age is not simply:
How intelligent will machines become?
It is:
What becomes scarce when intelligence no longer is?
References & further reading
This essay distinguishes empirical evidence from inference and speculation. AI capability, infrastructure and policy are changing rapidly; time-sensitive quantitative claims should be interpreted in the context of the publication date.
- International AI Safety Report 2026
- International Energy Agency — Energy and AI
- Stanford HAI — AI Index 2026
- Brynjolfsson, Li & Raymond — Generative AI at Work
- IMF — AI Adoption and Inequality
- Google DeepMind — Conjecture Machines
- METR — RE-Bench
- OpenAI — Weak-to-Strong Generalization
- UK AI Security Institute — Loss of Oversight
- AI 2027
- Tyssedal — The Mirage of Abundance
Final editorial edition, August 2026. Think Deeper welcomes substantive criticism, corrections and stronger counterarguments. Where the evidence changes, our conclusions should change with it.