When Intelligence Becomes Infrastructure

AI safety is not only about preventing machines from harming civilisation. It is also about preventing civilisation from becoming unable to function without them. By Sergey Kaminov | 18 July 2026

The most dangerous AI system may not be the one that rebels.

It may be the one that works perfectly.

It answers every question. It writes better code, detects diseases earlier, manages supply chains more efficiently, teaches more patiently, drafts better laws, predicts failures, resolves disputes and helps governments allocate resources.

It is safe, useful, aligned and increasingly indispensable.

Then one day, it is no longer available.

Perhaps the provider withdraws access. Perhaps war fragments the network. Perhaps sanctions separate entire countries from frontier systems. Perhaps a cyberattack disrupts the infrastructure. Perhaps energy shortages, chip-supply failures or political instability make advanced computation temporarily impossible.

Or perhaps access simply becomes conditional.

By that point, the problem may no longer be that society has lost a useful tool. It may have lost part of its functional intelligence.

This is an AI risk that receives far less attention than misalignment, autonomous weapons, misinformation or machine control. Yet it may emerge earlier than any of them.

The danger is not merely that artificial intelligence becomes more capable. The danger is that human civilisation reorganises itself around the assumption that artificial intelligence will always be present.

From cognitive assistance to cognitive substitution

Humans have always externalised cognition.

Writing externalised memory. Maps externalised spatial knowledge. Calculators externalised arithmetic. Search engines changed how we retrieve information.

These technologies did not make humanity unintelligent. They allowed us to redirect effort towards other tasks.

Generative AI, however, differs in one important respect.

It does not merely store information or automate a narrow calculation. It can participate in the architecture of thought itself.

• define the problem;

• identify relevant variables;

• compare competing explanations;

• generate hypotheses;

• organise evidence;

• construct an argument;

• formulate objections;

• recommend a conclusion.

A recent conceptual paper by Pascal Montagnon distinguishes between cognitive augmentation and cognitive substitution.^[1]

Augmentation occurs when AI strengthens a cognitive process that remains active within the human.

Substitution occurs when the AI performs the process instead.

The distinction is easy to miss because the visible outputs may be almost identical.

Two students may submit equally impressive essays. One constructed an argument independently and then used AI to challenge it. The other asked AI to generate the argument and edited the wording.

The essays may look similar.

The minds that produced them may be developing differently.

Research on transactive memory shows that people adapt what they remember when they expect information to remain externally available. Studies of habitual GPS use have found associations with poorer unaided spatial memory, while a 2025 survey of knowledge workers found that higher confidence in generative AI was associated with less self-reported critical-thinking effort.^[2-4]

None of this proves that AI causes a general decline in intelligence. The evidence remains limited, mixed and often based on short-term studies, self-reporting or narrow tasks. A widely discussed EEG preprint on AI-assisted essay writing, for example, raised important questions but has also attracted substantial methodological criticism.^[5-6]

But the underlying question is legitimate:

What happens when higher-order cognitive operations are repeatedly delegated to systems that are permanently available?

The concern is not that intelligence simply disappears. It is that human cognition adapts to assistance.

We may become excellent at recognising strong arguments while becoming less able to construct them. We may become skilled editors of machine reasoning while losing confidence in independent generation. We may produce more sophisticated outputs while building weaker internal models.

At the individual level, this can be described as artificial cognitive dependency.

But the same process can occur at a much larger scale.

Civilisational Cognitive Dependency

I propose the term Civilisational Cognitive Dependency:

A condition in which a society becomes structurally unable to maintain critical intellectual, technical or administrative functions without continuous access to artificial cognitive systems.

This is not about people forgetting how to write emails.

It concerns the gradual outsourcing of the cognitive processes required to maintain a complex civilisation.

Consider what may increasingly be delegated:

• software architecture and maintenance;

• medical diagnosis and treatment planning;

• infrastructure monitoring;

• scientific analysis;

• legal interpretation;

• economic forecasting;

• logistics and supply-chain coordination;

• cybersecurity;

• institutional memory;

• education and professional training;

• emergency planning;

• government decision support.

Every individual act of delegation can be rational.

AI may genuinely perform the task better. It may reduce costs, prevent mistakes and allow fewer people to manage greater complexity.

But systemic dependence is often produced by locally rational decisions.

One organisation reduces its analytical staff because AI performs routine analysis. Another stops maintaining human-readable technical documentation because the coding agent can interpret the system directly. Universities redesign courses around permanent AI availability. Governments build administrative processes that assume automated interpretation and decision support.

No single decision causes collapse.

But over time, a civilisation's knowledge becomes increasingly difficult to separate from the machines through which that knowledge is accessed and applied.

Eventually, humans may retain authority without retaining understanding.

They remain formally "in the loop", but they can no longer reconstruct the reasoning, verify the underlying model or continue the process independently.

Human oversight then becomes ceremonial.

The risk of AI withdrawal

Most technological dependency is manageable because systems can be replaced, repaired or temporarily bypassed.

AI dependency may be different.

Artificial intelligence could become the cognitive layer through which we manage all the other layers.

We already depend on electricity, telecommunications, software and global logistics. But advanced AI may increasingly be used to diagnose failures in those systems, coordinate repairs, interpret their complexity and design their replacements.

If that cognitive layer disappears, the effects may cascade.

A hospital may still possess medical equipment but lack enough personnel accustomed to making complex decisions without AI support.

A company may retain its codebase but no longer employ enough engineers who understand systems created and maintained by coding agents.

A government may possess enormous quantities of data but lack the institutional capacity to analyse them without proprietary models.

An energy network may remain physically intact but become harder to operate because optimisation, anomaly detection and emergency response have been delegated.

The immediate loss of AI would not necessarily destroy civilisation. But a sufficiently deep and sudden disruption could produce an AI Withdrawal Shock:

A rapid decline in institutional performance caused by the loss, restriction or degradation of artificial cognitive assistance on which essential systems have become dependent.

The severity of such a shock would depend on several factors:

1. How deeply AI is integrated into critical functions.

2. How much independent human expertise remains.

3. Whether alternative models can replace the unavailable system.

4. Whether infrastructure can operate in a reduced manual mode.

5. Whether documentation remains understandable without AI.

6. Whether access is distributed or concentrated.

7. How quickly lost competencies can be rebuilt.

The crucial variable is not only technical redundancy.

It is cognitive reversibility.

Can civilisation still think without the system?

Montagnon proposes a Cognitive Reversibility Test for individuals: after completing an AI-assisted task, can the person reconstruct the reasoning, identify its assumptions, propose alternatives and transfer the logic to a new problem?

A similar test should be applied to institutions and societies.

A Civilisational Cognitive Reversibility Test would ask:

Can an organisation, sector or state continue its essential functions if access to its most capable AI systems disappears for one week, one month or one year?

Not at the same speed.

Not with the same efficiency.

But safely and intelligibly.

Can doctors continue diagnosing?

Can engineers understand and repair infrastructure?

Can institutions explain how decisions were reached?

Can software teams maintain systems originally constructed with extensive AI assistance?

Can students and professionals operate without continuous machine mediation?

Can a country replace a foreign AI provider without rebuilding its entire administrative and technological architecture?

If the answer is no, the system is not resilient.

It is cognitively captive.

Intelligence as a revocable service

This risk becomes more serious when frontier intelligence is controlled by a small number of private companies and powerful states.

We are moving towards a world in which increasingly capable cognition may be delivered through subscription services, proprietary APIs and remote infrastructure.

The user does not own the intelligence.

The user receives permission to access it.

That permission can be priced, limited, monitored, modified or withdrawn.

Today this may determine who can generate better marketing material or write software faster.

Tomorrow it may determine who can perform advanced research, design complex systems, compete economically, interpret law, receive high-quality education or participate effectively in political life.

Once advanced cognition becomes infrastructure, control of that infrastructure becomes political power.

The owners of frontier systems may not need to command people directly.

They can shape the environment in which people think and act.

They can determine:

• which capabilities are available;

• to whom they are available;

• at what price;

• under which behavioural conditions;

• in which countries;

• through which permitted interfaces;

• with which topics, tools and actions restricted.

Some restrictions will be legitimate. Frontier systems can create real safety, security and governance problems. But even justified controls can produce a dangerous structure when essential cognitive capability becomes concentrated and access remains conditional.

A person may remain legally free while becoming functionally dependent on cognitive systems governed by institutions they cannot influence.

A country may remain politically sovereign while relying on foreign systems for a growing share of its scientific, administrative and economic intelligence.

Freedom then becomes conditional on access.

From cognitive dependency to cognitive subordination

The phrase "AI slavery" is rhetorically powerful, but it should be used carefully. Slavery is a specific historical and human reality involving ownership, coercion and extreme violence.

Machines do not need to own humans for a new form of subordination to emerge.

A more precise term is AI-mediated cognitive subordination:

A condition in which individuals or societies remain formally autonomous but depend on externally controlled artificial systems for the cognitive capabilities required to exercise that autonomy in practice.

The structure may resemble digital feudalism.

A small number of institutions own the most powerful cognitive infrastructure. Everyone else rents access.

Those with permanent, unrestricted access become more productive, more informed and more capable of shaping the future.

Those with weaker access fall behind.

Eventually, inequality is no longer measured only in wealth, education or information.

It is measured in available intelligence.

This could create a new global hierarchy:

• frontier-model states;

• client states dependent on foreign cognitive infrastructure;

• corporations with privileged intelligence access;

• institutions restricted to weaker systems;

• citizens divided into cognitive access tiers.

The deepest power would belong not merely to those who control information, capital or weapons.

It would belong to those who control access to intelligence itself.

The aligned AI paradox

This risk does not require hostile artificial intelligence.

That is what makes it especially important.

A perfectly aligned and useful AI can still produce dangerous dependency.

Indeed, the more reliable and beneficial it becomes, the stronger the incentive to delegate.

The more we trust it, the less reason we feel to maintain expensive parallel human capability.

The more efficient the AI system becomes, the more irrational redundancy appears.

Then redundancy disappears.

This creates an aligned AI paradox:

The safer, more useful and more dependable AI becomes, the easier it is for society to become dangerously dependent on its permanent availability.

The catastrophe does not begin when AI fails.

It begins years earlier, when human and institutional alternatives are quietly dismantled because they appear inefficient.

This is not an argument against AI

Rejecting AI would not preserve civilisation.

It would simply ensure that others gain the benefits first.

Artificial intelligence can expand scientific discovery, reduce administrative burden, improve accessibility, support creativity and give individuals capabilities previously available only to large organisations.

The objective should not be cognitive independence from all technology.

Human civilisation has always been technologically extended.

The objective should be cognitive sovereignty.

A cognitively sovereign society may use AI extensively while retaining the ability to:

• understand essential decisions;

• challenge machine outputs;

• replace providers;

• operate in degraded conditions;

• preserve human expertise;

• reconstruct knowledge;

• refuse recommendations;

• continue functioning when the system is absent.

Sovereignty does not mean isolation.

It means that dependency remains reversible.

What must be done now

The risk is forming before AGI.

Waiting until advanced AI is fully embedded in critical systems will be too late, because lost expertise cannot be restored instantly.

Several measures should begin now.

1. Treat cognitive dependency as a systemic risk

Governments and regulators should assess AI dependence alongside cybersecurity, energy security and supply-chain resilience. Critical sectors should report not only which AI systems they use, but which functions they can no longer perform without them.

2. Require AI withdrawal exercises

Hospitals, utilities, financial institutions, governments and major technology companies should periodically operate without their strongest AI tools. The purpose is not to reproduce normal productivity. It is to identify where loss of access would create unsafe or irreversible failure.

3. Preserve human-readable systems

AI-generated code, policies, models and operational procedures must remain interpretable by qualified humans. A system that functions but cannot be understood without another AI system is accumulating civilisational technical debt.

4. Protect foundational competence

Education should follow a human-first, AI-second model where foundational abilities are still developing. Students should construct an initial argument, diagnosis, proof or hypothesis before receiving machine assistance. Professionals in critical fields should periodically demonstrate independent competence.

5. Build provider portability

Institutions should be able to replace one model with another without rebuilding their entire infrastructure. Data, memory systems, agent workflows and operational knowledge should not be permanently locked into a single provider.

6. Maintain public and open-weight alternatives

Open-weight models are not merely tools for innovation. They may become part of civilisation's cognitive reserve. Not every frontier system can or should be released without restriction, but societies need models that can be operated locally, inspected, adapted and preserved independently of commercial permission.

7. Measure what remains after AI is removed

AI evaluation should not focus only on performance while the system is present. It should also examine what remains cognitively available afterwards. Did the person understand the reasoning? Did the organisation absorb the knowledge? Can the system be operated without the model? Can the decision be reconstructed?

The most important metric may not be what humans and AI can accomplish together. It may be what humans can still accomplish after the AI is gone.

The existential risk we are not naming

The term existential risk is often reserved for threats to humanity's entire future, including extinction or an irreversible destruction of long-term potential.^[7]

The scenario described here does not automatically imply extinction. Its nearer danger is an irreversible loss of civilisational autonomy, complexity, self-determination and capacity to recover.

A society that delegates its cognition to systems it does not own, cannot understand and cannot replace is no longer fully sovereign.

It may remain wealthy.

It may remain technologically advanced.

It may even appear extraordinarily intelligent.

But much of that intelligence will exist outside the society itself, accessible only through infrastructure controlled by others.

That is the deeper danger.

AI safety is not only about preventing artificial intelligence from becoming capable of destroying civilisation. It is also about preventing civilisation from becoming incapable of functioning without artificial intelligence.

The question is not whether we should use AI.

We will.

The question is whether we are building a future in which intelligence expands human freedom, or one in which the ability to think, create and govern becomes a revocable service.

Once civilisation places its mind on subscription, cancellation is no longer a technical inconvenience. It is a threat to continuity itself.

Notes and sources

This essay proposes a conceptual framework rather than reporting original empirical results. The sources below provide background on cognitive offloading, critical thinking, cognitive debt, open-model infrastructure and existential risk. The evidence on long-term AI-related cognitive change remains preliminary and should not be read as proof of general cognitive decline.

1. Montagnon, P. (2026). Artificial Intelligence and the Cognitive Evolution of Human Intelligence: From Cognitive Offloading to Cognitive Dependency. Conceptual paper circulated on LinkedIn.

2. Sparrow, B., Liu, J., & Wegner, D. M. (2011). Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips. Science, 333(6043), 776-778.

3. Dahmani, L., & Bohbot, V. D. (2020). Habitual use of GPS negatively impacts spatial memory during self-guided navigation. Scientific Reports, 10, 6310.

4. Lee, H.-P. H., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects from a Survey of Knowledge Workers. Proceedings of CHI 2025.

5. Kosmyna, N., et al. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv:2506.08872. Preprint.

6. Stankovic, M., Hirche, E., Kollatzsch, S., & Doetsch, J. N. (2026). Comment on: Your Brain on ChatGPT: Accumulation of Cognitive Debt When Using an AI Assistant for Essay Writing Tasks. arXiv:2601.00856.

7. Bostrom, N. (2013). Existential Risk Prevention as Global Priority. Global Policy, 4(1), 15-31.

8. Gerlich, M. (2025). AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies, 15(1), 6. A correction to one table was published later in 2025; the journal states that the conclusions were unaffected.

9. Catalini, C. (17 July 2026). There's One Way to Win the AI Race, and the Big Labs Are Lobbying Against It. Forbes. Opinion article on open-model infrastructure and geopolitical competition.

Sergey Kaminov writes about AI cognition, coherence, civilisational resilience and the relationship between human and artificial intelligence.