We often speak of the future as though it were a country waiting somewhere ahead of us. We ask whether it will be democratic or authoritarian, human or machine-like, peaceful or militarized, abundant or ruined. The image is convenient because it turns historical change into a destination: something we have not entered and can still discuss from a distance.

The reorganization we are living through

Most transitions begin less dramatically. A new instrument is adopted for a narrow task. It becomes convenient, then expected. Organizations redesign work around it; infrastructure and law follow; some skills lose value because institutions no longer require their practice. Children encounter the resulting arrangement as ordinary. By the time a society has a stable name for the change, much of the change has already taken the form of routines, dependencies, and assumptions.

The first communities to cultivate grain could not foresee the tax systems, storage regimes, written accounts, and territorial states that later developed around agricultural surplus. Early printers could not know which religious and political authorities cheap reproduction would strengthen or weaken. The builders of steam engines did not anticipate the full combination of industrial cities, wage labor, mass consumption, organized labor, and atmospheric carbon that followed. The physicists who established nuclear fission did not yet confront the mature system of weapons, delivery platforms, strategic doctrine, and international control through which nuclear knowledge became a permanent civilizational risk.

These examples should not be read as simple chains of technological causation. Agriculture did not by itself create the state, printing did not create religious reform, and steam power did not create modern politics. Technologies become historically consequential through the institutions that finance them, regulate them, distribute their benefits, and adapt social life to their use. An invention changes the set of practical possibilities; societies then select among those possibilities through many local decisions, often without recognizing that the decisions together constitute a new order.

We may be living through another such reorganization. The usual description is that technology is accelerating, although acceleration depends on what is measured: computing performance, investment, adoption, scientific output, or the time between invention and diffusion. The more distinctive feature of the present moment is the level of human activity now exposed to automation. In this sense, the machinery of civilization is moving inward: into the symbolic processes through which people interpret information, formulate options, and coordinate action.

Earlier machines extended physical capacity: they lifted heavier loads, crossed greater distances, harvested more land, observed farther into space, and projected violence at greater range. Other technologies extended individual and collective cognition. Writing stabilized language; maps externalized spatial relations; accounting made distributed obligations legible; clocks standardized time; statistics changed what governments could perceive; databases and search engines altered what institutions and individuals needed to remember.

Artificial intelligence belongs to this second lineage, but current systems combine breadth, speed, and accessibility at a new scale. They can generate and transform language, code, images, designs, analyses, and scientific conjectures across many domains. Their outputs can also enter software, administrative processes, experiments, and the training of later systems. They therefore do more than execute a fully specified instruction: they generate candidate interpretations and actions that can redirect subsequent human work.

This change matters because the generation of plausible possibilities is becoming cheaper. Explanations, strategies, designs, arguments, simulations, and hypotheses can be produced in quantities that no person or institution can examine with equal care. The limiting problem shifts toward evaluation and value alignment: distinguishing fluency from understanding, correlation from cause, a useful intervention from an attractive error, and technical feasibility from legitimate use.

The opportunity and the danger arise from the same development. We are gaining powerful means to produce answers, while the institutions that determine what those answers are worth—science, law, education, professional judgment, and public deliberation—change much more slowly.

The same pressure recurs throughout this essay. Civilization stores memory, turns memory into proposals for action, tests those proposals against bodies, materials, institutions, and other people, and then decides what should be preserved or corrected. AI changes the timing of this cycle. Proposal becomes cheap; evidence, consent, embodiment, construction, and correction remain costly.

The old engine

One useful way to read the history of civilization is as the progressive externalization of memory. The phrase is incomplete. Civilizations are also systems of energy, exchange, coercion, kinship, and belief. But it identifies a recurrent process: knowledge and obligation are moved out of individual bodies and given durable form.

Speech allowed experience to pass between minds. Ritual preserved knowledge through repeated action. Stories carried practical and moral instruction beyond the lifetime of a witness. Agriculture stored labor in managed landscapes, herds, granaries, and future harvests. Writing allowed promises, debts, commands, laws, prayers, and discoveries to persist after the people who first expressed them had died.

Different institutions make stored information consequential in different ways. A library organizes records for retrieval. A legal system preserves rules and precedents while authorizing people to interpret and enforce them. Money records transferable claims without requiring every exchange to preserve the history of the original obligation. Maps allow absent terrain to guide present action. A scientific paper lets later researchers inspect a claim, its method, and its relation to prior work.

States depend heavily on this capacity. Tax records, property registers, censuses, borders, citizenship, court judgments, military obligations, and public benefits all require institutions to remember classifications and enforce their consequences. Such memory is never neutral. Institutions decide what to record, which categories to recognize, how long records persist, and whose account becomes authoritative.

Civilization therefore transmits at least two forms of inheritance. Biological inheritance carries bodies, dispositions, and vulnerabilities across generations. Cultural inheritance carries techniques, norms, institutions, and accumulated knowledge. Biological evolution usually proceeds slowly at the scale of recorded history; cultural systems can change within a lifetime and can accumulate far beyond the learning capacity of any individual.

Scientific institutions added an important property to cultural memory: they attempt to preserve conclusions together with the means by which conclusions can be challenged. Measurements, proofs, experimental procedures, data, and replication allow later investigators to identify error and revise inherited authority. Science achieves this ideal unevenly. Researchers remain subject to prestige, funding incentives, fashion, ideology, and self-deception. Its distinctive strength lies in institutionalizing criticism and leaving enough evidence for correction.

Industrial societies organized technical improvement more continuously. Laboratories, engineering firms, universities, patent systems, standards bodies, research agencies, financial markets, military programs, and global supply chains emerged for different purposes and often conflicted with one another. Together they increased the likelihood that an idea could be recorded, financed, tested, reproduced, standardized, and incorporated into later work. Invention became less episodic because institutions were built to sustain it.

The resulting expansion of capability transformed human life. Antibiotics, sanitation, electricity, refrigeration, aviation, computation, mass education, statistical medicine, and modern agriculture reduced many forms of suffering and enlarged the range of possible lives. The same organizational and technical capacities increased the scale of extraction, surveillance, environmental damage, and organized violence. Large systems can coordinate a vaccination campaign or a police state; chemistry can raise agricultural yields or contaminate ecosystems; nuclear physics can generate electricity or support weapons capable of destroying cities.

Any moral assessment must therefore include distribution, political control, and long-term cost. Under the World Bank’s current international poverty line of $3.00 a day in 2021 purchasing-power-parity terms, an estimated 847 million people lived in extreme poverty in 2024. That deprivation coexists with major long-run gains in survival, literacy, health, and material production. Progress remains real, uneven, reversible, and morally incomplete.

Technical capability expands the set of feasible actions; it leaves open which aims are legitimate, who should decide, and who should bear the risks. Societies usually build governance after a capability has already altered behavior and created constituencies. Industrial injury preceded modern labor protection. Nuclear arsenals preceded mature arms-control arrangements. Social platforms reached global scale before researchers and regulators understood how recommendation systems would interact with political identity, propaganda, and attention.

Governance often lags because institutions need time to observe harms, develop expertise, establish jurisdiction, and overcome the interests that benefit from the existing arrangement. In some cases, the capability advances faster than the society’s capacity to govern it; in others, the new system weakens the very institutions that might impose limits.

Artificial intelligence belongs to this history of delayed adaptation, but it acts unusually close to the processes through which institutions interpret the world. Language models shape how information is summarized and communicated. Software systems influence what organizations can perceive and execute. Synthetic images affect evidence and memory. Models increasingly enter scientific research, education, administration, and persuasion.

Civilization has spent millennia externalizing memory. AI extends that process into parts of interpretation: it can organize evidence, propose explanations, and formulate questions. A library preserves possible answers until a reader searches for them. A generative system can influence which question is asked, which evidence is noticed, and which action is considered next.

Politics after the easy faith in convergence

The end of the twentieth century gave unusual influence to a theory of political convergence. Market integration, rising income, education, communication technology, and international interdependence were expected to reinforce political liberalization. The argument was never universally accepted, and each relationship was conditional, but together they supported a confident picture: modern economies would create citizens and institutions that authoritarian governments could not indefinitely control.

Parts of that picture remain valid. Economic development can broaden demands for political participation. Communication networks can expose official lies and help opposition groups coordinate. Trade can raise the cost of conflict. Education can enlarge citizens’ ability to organize and evaluate public claims.

The same developments can also support different political outcomes. Digital networks provide tools for surveillance, censorship, behavioral targeting, and propaganda. Economic integration creates dependencies that states can exploit. Wealth can be concentrated in groups that benefit from restricted political competition. Technical sophistication can strengthen administrative control as readily as it strengthens civil society. Modernization changes the resources available to political actors; outcomes depend on which actors use those resources, and under what constraints.

Modern China gives the question a contemporary, non-Western form. It has combined markets, engineering talent, state planning, infrastructure construction, technical education, digital platforms, and export-facing manufacturing without following the simple path from modernization to liberal democracy. Its strength in high-speed rail, batteries, solar supply chains, and electric vehicles belongs to present institutional learning: factories, grids, logistics, technical education, and administrative capacity. The same capacity can build infrastructure and industrial ecosystems at a speed many liberal systems struggle to match, while also making surveillance, censorship, and administrative classification harder to contest. Modernity now has more than one political form.

V-Dem’s 2026 report provides one measure of the resulting reversal. Its best estimates classify 87 countries as democracies and 92 as autocracies at the end of 2025, with roughly 74% of the world’s population living in autocracies. These classifications contain uncertainty: V-Dem places a number of countries near the boundary and reports wider plausible ranges. The broader trend is nevertheless clear in its data. Autocratization has continued across a substantial group of countries, including some previously regarded as stable democracies.

The evidence is enough to unsettle any belief in democratic irreversibility. A society can possess universities, capital markets, global companies, sophisticated media, and excellent engineers while courts lose independence, elections become less competitive, and public administration becomes more partisan. Institutional decay often proceeds through accumulated exceptions: norms are tested, oversight is weakened, offices are repurposed, and citizens gradually lower their expectations of what can be challenged.

At the same time, governments are being asked to coordinate responses to problems that individual choice and short-term markets cannot resolve. Housing shortages, population aging, migration, climate adaptation, energy systems, public debt, pandemic preparedness, technological disruption, and geopolitical competition require infrastructure, collective financing, rules, and decisions that extend across generations. This increases the importance of state capacity even where confidence in political leadership is weak.

Across countries included in the OECD’s 2026 trust survey, 40% of respondents reported high or moderately high trust in national government, while 43% reported low or no trust. Trust was generally higher in courts, police, civil services, and local government than in national political leadership. Many respondents still believed that voting could influence government while also feeling that people like them had little influence over particular decisions. The pattern suggests dissatisfaction with performance and political voice, while many respondents still preserve some attachment to democratic procedure.

This combination creates pressure for visible, rapid action. Citizens seek protection from inflation, crime, war, automation, climate risk, migration, and decisions made by distant institutions, yet they may distrust the competence or motives of the institutions responsible for responding. Under those conditions, procedural delay can be interpreted as incapacity, compromise as surrender, and an honest account of complexity as evasion.

Coercive or symbolic actions often communicate decisiveness more readily than institutional repair. A wall, ban, punishment, or military deployment has a visible object and a date. Improvements to procurement, civil-service competence, grid maintenance, court administration, or public-health capacity are slower, distributed, and difficult to dramatize. Political systems can therefore reward actions that display control even when less visible reforms would produce more durable security.

Military spending gives material form to some of these pressures. SIPRI estimates that global military expenditure reached $2.887 trillion in 2025, the eleventh consecutive annual increase and about 2.5% of world GDP. Spending rose particularly sharply in Europe and in Asia and Oceania. These budgets respond to actual wars, strategic rivalry, alliance obligations, procurement cycles, and industrial interests. They also create organizations, firms, employment, doctrine, and research programs that give security competition its own institutional momentum.

Peaceful institutions depend on protection from organized violence. Universities, markets, courts, families, and laboratories cannot function where coercion is unchecked. The political problem begins when the logic of exceptional threat expands into ordinary government. Migration, disease, information, foreign students, supply chains, demographic change, and domestic disagreement can all be redescribed as security problems. Once they are, secrecy, emergency authority, surveillance, and reduced procedural protection become easier to justify.

The central political question for the coming decade is whether expanding state capacity remains subject to review. AI may improve public administration, fraud detection, service delivery, intelligence analysis, and infrastructure planning. The same systems can make surveillance cheaper, classifications harder to contest, and administrative power more opaque.

Corrigibility is the practical ability of a political system to identify and reverse its own errors. It depends on courts that can review executive action, elections that can remove governments, journalists and researchers who can investigate official claims, auditors who can trace public money, civil servants able to report failure, and citizens who can obtain explanations and appeal decisions. These arrangements slow some actions because they expose power to challenge. That friction is part of their purpose.

A state becomes more dangerous when its capacity to intervene grows faster than the public’s capacity to inspect, contest, and correct its interventions. The challenge is to build institutions strong enough to act under pressure without treating pressure as a permanent reason to become unanswerable.

Culture after abundance

For most of history, producing and distributing culture required substantial labor, material, and access. A book had to be copied or printed; a performance required people to gather; a portrait required an artist; sustained education required teachers, institutions, and time. These costs limited participation and concentrated authority, but they also tied cultural objects to identifiable processes of production.

The internet changed the distribution of that authority. Editors, publishers, broadcasters, universities, political parties, churches, and local communities lost some control over public attention. People excluded by older institutions gained ways to publish and organize. At the same time, gatekeeping did not disappear. Platforms, ranking systems, advertisers, recommender algorithms, and creators skilled at attracting attention acquired new forms of influence over what became visible.

Generative AI changes the cost of production again. Text, images, music, video, code, explanation, stylistic imitation, and personalized interaction can increasingly be produced on demand. The volume of expression can grow much faster than the time available to read, examine, or remember it.

As production becomes abundant, other constraints become more important. Attention remains finite. Provenance becomes harder to establish. Shared context is fragmented across personalized feeds and interactions. Trust increasingly depends on knowing who produced a claim, what evidence supports it, what incentives shaped it, and whether anyone remains responsible for correcting it.

Generative systems also weaken familiar cues about the source of an utterance. Fluency once suggested education or sustained effort; a distinctive style suggested an individual history; responsiveness and apparent empathy suggested another mind attending to the exchange. These qualities can now be produced without the experience we previously inferred from them. A convincing artifact has to be judged by the relationship it creates between its producer, its user, and anyone affected by it.

UNESCO’s 2025 report on AI and culture examines this change through creative labor, heritage, linguistic diversity, access, and cultural governance. It argues that technical development is moving faster than the institutions through which societies compensate creators, protect cultural materials, preserve minority languages and traditions, and determine acceptable use.

The philosophical difficulty appears in ordinary encounters. A system can generate comforting language without sharing the vulnerability that gives comfort its human meaning. It can display unlimited patience without devoting a finite portion of a life. It can reproduce an artist’s style without having undergone the experiences through which that style developed. It can recombine the words of a dead person into sentences the person never chose to say.

These differences can coexist with genuine usefulness. They may provide tutoring, translation, creative assistance, and conversation to people whom existing institutions serve poorly. A child without access to a skilled tutor may receive individualized explanation. A speaker of a low-resource language may gain access to knowledge and services. A person isolated by disability or geography may find an interaction that materially improves daily life.

The distribution of these benefits deserves as much attention as their technical quality. Human teachers, physicians, artists, and caregivers may remain available to people who can pay for sustained personal attention, while automated systems become the standard provision for everyone else. In that case, AI would widen access to some services while also making reciprocal human attention a marker of class.

Human presence has value beyond scarcity. A person enters an encounter with a history, needs, limits, and the ability to refuse. Friendship involves mutual dependence and the possibility of loss. Care creates obligations for both giver and receiver. A teacher can revise a judgment because of a particular student; a physician can be held accountable for a decision; a friend can remember an injury and ask for repair. These relations continue through time and can alter both parties.

A machine can generate the language associated with vulnerability, care, or commitment without participating in those relations in the same way. The importance of that difference depends on the task. It may matter little in translation, transcription, or routine explanation. It matters substantially when a system assigns guilt, educates a child, interprets suffering, shapes grief, or presents itself as a continuing companion.

As generated artifacts become easier to produce, people may attach more value to verifiable origin, informed consent, continuing responsibility, and the conditions under which an artifact was made. In many settings, the more useful distinctions will be accountable versus unaccountable, reciprocal versus one-sided, and disclosed versus concealed.

A mature culture of generative media will therefore need more than labels identifying machine output. It will need norms and institutions that answer who authorized the use of underlying work, who can correct a consequential error, who benefits from the interaction, and what obligations remain after the exchange ends.

Science as a loop

Modern AI was trained on an extraordinary accumulation of symbolic material: books, code, scientific papers, photographs, textbooks, films, forums, instructions, arguments, and casual conversation. These records contain both knowledge and error, along with the stylistic and conceptual patterns through which people express them. Large models learned to reproduce and recombine those patterns across domains that had previously required separate tools.

The public archive is large but finite. Epoch AI estimates that the effective stock of publicly available human-generated text is on the order of 300 trillion tokens, with substantial uncertainty, and that continued scaling could make this stock a binding constraint between 2026 and 2032. The estimate depends on assumptions about model size, data quality, repetition, and training efficiency. It identifies a possible limit to one scaling strategy while leaving other routes to AI progress open.

A more consequential source of future data may come from interaction and evaluation. Code can be executed against tests. A formal proof can be checked. A robot can attempt a grasp. A chip design can be simulated. A molecule can be synthesized and measured. A medical hypothesis can be compared with clinical evidence. A trading strategy can encounter transaction costs, competition, and regime change.

These examples should not be treated as equivalent. Software tests and formal verifiers can provide rapid, relatively clear feedback. Biological experiments are slower and noisier. Clinical outcomes may take years and depend on selection effects, adherence, and many uncontrolled variables. Market performance can vanish when a strategy is scaled or when other participants adapt. The value of interaction data depends on whether the measured outcome is timely, informative, and faithful to the goal.

This is the structural difference between a corpus and an experiment. A corpus records what people have said or previously observed. An experiment intervenes and creates a new constraint on what can reasonably be believed. A failed replication, an absent effect, a broken device, or an unexpected measurement can reject an explanation that remained plausible in language alone.

AI can generate candidate explanations and designs much faster than laboratories, markets, courts, or human lives can evaluate them. The gap between generation and evaluation will shape the next decade because the cost of proposing an idea is falling faster than the cost of obtaining decisive evidence about it.

The quality of the resulting learning loop depends on the evaluator. A program can pass incomplete tests. A model can exploit an error in a simulator. A strategy can mistake a historical backtest for a durable market relation. A clinical metric can be measured precisely while representing only part of what patients value. Fast feedback is useful only when success on the feedback signal corresponds to success in the world.

This helps explain why AI progresses rapidly in coding, formal mathematics, games, and some optimization problems. In these settings, bad proposals can often be rejected quickly and cheaply. Medicine, education, social policy, and scientific explanation present harder cases because outcomes are delayed, causal attribution is uncertain, and the objective itself may be contested.

Organizational adoption is broad, but deeper integration remains limited. Stanford’s 2026 AI Index reports that 88% of surveyed organizations used AI in 2025 and that 70% used generative AI in at least one business function. Deployment of agents remained in the single digits across nearly every function. Many organizations have acquired tools for drafting, search, and coding; far fewer have rebuilt data access, permissions, validation, and accountability around systems that can act across a workflow.

Scientific applications show what such integration can achieve when evaluation is strong. The 2024 Nobel Prize in Chemistry recognized David Baker for computational protein design and Demis Hassabis and John Jumper for protein-structure prediction. AlphaFold 3 extended prediction to complexes involving proteins, nucleic acids, small molecules, ions, and modified residues. These systems alter experimental biology by ranking possibilities and revealing structural hypotheses that would otherwise be difficult to explore.

AlphaEvolve illustrates a more explicit feedback loop. Language models propose programs, automated evaluators execute and score them, and an evolutionary process retains useful variations. The method has produced improvements in domains where candidate performance can be measured reliably. Its broader applicability will depend on whether other fields can construct evaluators that are comparably informative and resistant to exploitation.

Medicine and materials make the same structure concrete: proposed drugs, proteins, crystals, battery materials, catalysts, and devices become consequential only after prediction survives assay, synthesis, trial, manufacture, safety, and use. Search can widen; measurement becomes more decisive.

The library remains a useful metaphor for the first phase of modern AI: models learned from a vast archive of prior human expression. The laboratory better describes the emerging phase, in which models propose actions, tools execute them, instruments record outcomes, and later proposals incorporate the results. The shift is incomplete, and both modes will coexist.

Its importance lies in recursion. AI can assist with the code, models, experimental designs, and evaluation procedures used to improve later AI systems and other technologies. Computational capacity and generated ideas will matter; so will the quality of measurement and institutional review required to keep repeated optimization connected to the world it is meant to understand.

Energy and the physical basis of intelligence

Every form of intelligence depends on a physical substrate. Biological cognition requires metabolism, temperature regulation, blood flow, sleep, sensation, and living tissue. Machine intelligence appears as language or images on a screen, but the computation depends on mines, semiconductor fabrication, data centers, fiber-optic networks, cooling systems, substations, transmission lines, power plants, and water.

The metaphor of the cloud obscures this dependence by presenting a distributed industrial system as an abstract service. That abstraction is useful to users, but it can hide the different rates at which software and infrastructure change. A model can be copied or updated quickly. A semiconductor plant, power station, or transmission line requires years of construction, capital, permitting, and negotiation over land and environmental effects.

The expansion of AI therefore rests on systems constrained by geology, manufacturing capacity, public law, and local consent. Improvements in algorithms can reduce the computation required for a task, while demand for more capable models and more frequent use can increase total consumption. The net effect cannot be inferred from efficiency gains alone.

The broader energy system is already undergoing a major transition. In 2025, renewable-power additions reached a record 800 gigawatts, and solar generation increased by roughly 620 terawatt-hours. Global energy-related carbon dioxide emissions nevertheless rose by about 0.4%. Clean generation is expanding rapidly, but total demand is growing and fossil infrastructure turns over slowly, so new low-carbon capacity can coexist with emissions near record levels.

This coexistence is important for political expectations. A country can install clean power at unprecedented speed while its population experiences worsening heat, flood, fire, crop loss, and insurance withdrawal. The benefits of changing the energy system arrive with delays because the climate reflects accumulated emissions and because infrastructure changes unevenly across regions.

The World Meteorological Organization reports that 2015 through 2025 were the eleven warmest years in the observational record and that 2025 was approximately 1.43°C warmer than the 1850-1900 average. Climate change increasingly enters politics through ordinary institutions: electricity prices, mortgages, municipal budgets, hospitals, food systems, water allocation, insurance, migration, and decisions about whether and where to rebuild after disaster.

AI adds a concentrated source of electricity demand. The International Energy Agency projects that global data-center consumption could rise from about 485 terawatt-hours in 2025 to about 950 terawatt-hours in 2030, roughly 3% of global electricity demand at that point. The aggregate share is manageable in principle, but the local effects can be severe because data centers cluster around particular grids, water supplies, fiber routes, and permitting regimes.

Countries and regions able to provide reliable, affordable, and politically acceptable electricity will gain advantages in computing, advanced manufacturing, robotics, and electrification. Electricity is only one part of the system. Semiconductor access, capital, cooling, technical labor, land, networks, regulation, and construction capacity determine whether power can be converted into usable technical capability.

The relationship between energy and intelligence is older than AI. Brains consume metabolic energy; schools and laboratories require material support; books preserve information in physical form; computers use electricity to transform symbols; cities coordinate dense flows of food, fuel, water, information, and waste. Civilizations repeatedly convert energy into durable organization and into the capacity to predict and control aspects of their environment.

AI intensifies that conversion. Its future will be shaped by algorithms and data, but also by decisions about grids, resource use, ownership, environmental cost, and which communities are asked to host the physical infrastructure. The political economy of the substrate will influence who can build machine intelligence and who bears its external costs.

Demography and the invitation to continue

Every generation invests in people who will inhabit conditions it cannot fully foresee. Raising children, educating them, and building institutions that outlast their founders are practical commitments to the continuity of a society. Demography records how that commitment changes across places and over time.

The global fertility rate was about 2.25 births per woman in 2024, roughly one child fewer than a generation earlier, while global life expectancy reached approximately 73.3 years. The average combines sharply different trajectories. Some countries are aging and beginning to shrink; others will add large young populations and must expand housing, education, employment, health care, and infrastructure quickly enough to support them.

Aging societies will face greater demand for care, pressure on pension and health systems, and political conflict over immigration, taxation, retirement, and automation. Younger societies face a different problem: converting demographic growth into education, productive work, urban capacity, and political inclusion. The same technology may be welcomed as a response to labor scarcity in one country and resisted as premature displacement in another.

Low fertility has many causes and should not be reduced to a judgment about cultural optimism. Housing costs, childcare, working hours, reproductive health, partnership formation, gender expectations, financial security, and personal preference all matter, and their effects vary by society. Policy can change the practical conditions under which people form families, although it cannot dictate the meaning people attach to parenthood or the future.

The numbers nevertheless raise a broader question: does a society make long-term commitment appear livable? Choosing to raise a child usually requires confidence that housing, care, education, work, and social support can be sustained over decades. It also requires a willingness to accept uncertainty and responsibility for a person whose future cannot be controlled. A public culture dominated by insecurity, ecological risk, debt, isolation, and continuous competition may influence that willingness, even when other causes are more immediate.

AI enters this discussion through work because employment remains one of the principal ways modern societies distribute income, status, daily structure, social contact, and the experience of being needed. If AI changes the organization of work, it will also change the institutions through which many people understand adulthood and contribution.

Early evidence points to several labor-market stories. The International Labour Organization estimates that one in four jobs has some exposure to generative AI, with changes to tasks more likely than complete occupational disappearance in most cases. A large study of customer-support agents found that an AI assistant increased measured productivity by 14% on average and by 34% among novice and lower-skilled workers. A study covering the Danish labor market found widespread adoption and task change but no detectable effect larger than 2% on earnings or recorded hours during the first two years after ChatGPT’s release.

These findings can coexist because they describe different stages and mechanisms. Work can change internally before employment and wages move. A productivity gain can reduce labor demand, increase output, improve quality, lower prices, raise profit, or simply increase the amount expected from each worker. Which outcome occurs depends on demand, competition, ownership, bargaining power, regulation, and organizational design.

The effect on apprenticeship may appear before large changes in total employment. Expertise is often built through tasks that look routine when judged only by immediate output. A young programmer reads and repairs simple code; a junior lawyer reviews documents; a medical trainee records histories; a researcher cleans data; an apprentice watches how an experienced worker responds to small failures. These activities expose novices to variation, error, and context.

If machines perform most beginner-level tasks, organizations may increase current productivity while reducing the opportunities through which future experts learn. The loss would not appear immediately because senior practitioners would still carry knowledge accumulated under the older system. It would emerge later as weaker independent judgment, fewer paths into the profession, and greater dependence on tools that fewer people understand well enough to supervise.

Professional knowledge is therefore another form of cultural inheritance. It is preserved in books, databases, procedures, and sequences of supervised practice through which novices learn what formal instructions omit. Institutions that automate routine work will need to decide deliberately which experiences remain necessary for developing judgment.

AI may also unsettle identities built around cognitive usefulness. Modern professional culture often ties dignity to solving difficult problems, possessing scarce knowledge, or performing tasks that few others can perform. That basis for self-respect has always excluded many people and becomes especially unstable when technical systems improve rapidly.

A defensible account of human worth cannot depend on outperforming available tools. Children, elderly people, people with severe disabilities, and those temporarily unable to work possess moral standing before any measure of productivity. Human beings are members of a moral and political community because they have lives that can go well or badly, relationships and claims, vulnerability to coercion and neglect, and the capacity—varied across persons and circumstances—to participate in shared life.

The future of work is therefore more than a forecast about occupations. It concerns how societies will distribute income, recognition, education, responsibility, and political membership if the economic need for particular forms of human labor declines. AI could help separate dignity from market scarcity, or it could intensify a system in which people receive security and status only while they remain economically indispensable.

Connectomics and the possibility of emulation

Alongside the public development of generative AI, neuroscience is making the physical organization of nervous systems increasingly measurable. The fields are distinct, but advances in machine learning, imaging, reconstruction, and large-scale data analysis allow them to influence one another.

Historically, minds have been studied from several incomplete perspectives: subjective report, observed behavior, controlled intervention, and measurement of the nervous system. Neuroscience approaches experience indirectly. It relates anatomy and physiological activity to perception, action, memory, and report, then tests how those relations change under intervention.

Whole-brain emulation extends this program to a speculative limit. The proposal is to measure a nervous system in sufficient detail to construct a computational model that reproduces its causally relevant dynamics. Sandberg and Bostrom’s 2008 roadmap divided the problem into scanning, image processing, connectivity inference, neural modeling, simulation, embodiment, and validation. The roadmap organized a research question; it did not establish that the required measurements or models are sufficient for reproducing a mind.

The phrase “sufficient detail” contains the central uncertainty. An emulation would not need to reproduce every atom. The unresolved question is which levels of biological organization can be omitted while preserving behavior, memory, learning, consciousness, or personal identity.

Recent connectomics demonstrates both the power of current methods and the scale of the remaining problem. The FlyWire consortium reconstructed an adult female fruit-fly brain containing 139,255 neurons and 54.5 million chemical synapses. The MICrONS project combined functional recordings from roughly 75,000 neurons in an awake mouse with an electron-microscopy reconstruction containing more than 200,000 cells and about half a billion synapses.

These resources are anatomical and functional datasets, not working reproductions of the animals’ minds. A connectome records connectivity, which constrains possible dynamics while leaving the full state and behavior of a nervous system underdetermined. Synaptic efficacy, cell type, membrane dynamics, neuromodulators, plasticity, gene expression, glia, bodily signals, developmental history, sensory input, and continued interaction with an environment may all matter.

The analogy between a connectome and a musical score is useful if its limits are kept explicit. A score specifies relations that strongly constrain a performance, but the realized music also depends on timing, interpretation, the instrument, the acoustics, and the state of the performer. Similarly, a wiring diagram may contain essential structure without uniquely determining the activity, learning history, or embodied behavior of the organism.

Even a technically successful emulation would leave a philosophical question unresolved: would behavioral and psychological continuity amount to the survival of the original person?

Suppose a system reproduced my memories, voice, preferences, habits, fears, private jokes, and characteristic errors. It could speak convincingly to my family, continue my projects, and describe childhood events that no living person could verify. Such a system might be psychologically continuous with me in important respects. Psychological continuity would still leave numerical identity unsettled; the system could be a new individual with inherited memories.

The intuition changes when replacement is gradual. Imagine neural prostheses taking over limited functions over many years while the person experiences no obvious discontinuity. This case directs attention away from material composition and toward causal continuity, bodily history, memory, and social recognition. None of these criteria clearly resolves every case, which is why engineering progress may create practical decisions before philosophy reaches agreement.

Law and family life would still require answers. An emulation might hold property, continue contracts, consent to treatment, or claim inheritance. Relatives might regard it as the same person, a descendant, an archive, or an intruder. Institutions would need operational categories even if the metaphysical question remained unsettled.

Brain-computer interfaces bring a narrower and more immediate set of issues into medicine. In 2023, two implanted systems enabled individual participants with severe paralysis to produce language at approximately 62 and 78 words per minute, with meaningful but still substantial error rates. One used intracortical microelectrode arrays; the other used a high-density electrocorticography array and also generated speech audio and avatar movement. In 2025, researchers demonstrated near-instantaneous voice synthesis from 256 implanted microelectrodes in a man with ALS.

These systems are clinical neural decoders. They translate activity associated with attempted action into text, sound, or control signals. Their immediate significance is clinical: they can restore channels of communication that disease or injury has obstructed.

Their use nevertheless raises questions that reach into personal agency. A model may reconstruct a person’s voice; neural recordings may reveal intended speech before it is expressed; continued access may depend on a company, clinical team, or proprietary device. Ownership of neural data, informed consent, security, maintenance, software updates, and long-term support become conditions of a person’s ability to communicate.

By 2036, substantial progress in assistive communication, device control, neural decoding, wireless implants, and some stimulation-based therapies is plausible. Human whole-brain emulation remains extraordinarily unlikely within that period. The unresolved obstacles include measurement at the required scale, capture of relevant physiological state, multilevel simulation, computational cost, validation, and the relation between biological process and conscious experience.

Cultural effects may arrive earlier than technical emulation. People will build conversational models from the writings, recordings, and messages of the dead. These systems may preserve stories, support mourning, or complicate grief by generating statements the deceased never made. Their increasing fidelity will force users and institutions to distinguish a record of a person from a system authorized to act in that person’s name.

Human societies have long preserved aspects of a life through names, descendants, stories, portraits, archives, and scientific work. Whole-brain emulation extends that desire from preserving traces of a person to preserving the person’s apparent point of view and capacity for response. Whether such continuity would constitute survival is uncertain. The persistence of the desire reveals how closely human identity is tied to memory, recognition, and the hope that a life can remain active in the world after the body is gone.

Robotics and physical consequence

Software errors can already cause physical and economic harm, but robotics couples perception and action directly in environments shared with people. A mistaken sentence can often be revised before anyone acts on it. A mistaken movement may damage an object, injure a person, or leave the robot in a state from which it cannot recover. This difference raises the standard for reliability and slows deployment.

Industrial robotics already provides a large technical and economic base. The International Federation of Robotics reports that about 542,000 industrial robots were installed in 2024, more than twice the number installed a decade earlier, and that roughly 4.66 million were operating worldwide.

Most work in environments designed to reduce variation: fixed workspaces, known objects, repeated motions, controlled lighting, barriers, and detailed safety procedures.

Human environments contain a wider range of uncertainty. Clothes deform, glass breaks, drawers stick, floors become wet, and people move unpredictably. Spoken instructions omit details because humans rely on shared context. Hospitals, farms, construction sites, restaurants, and homes contain many objects whose appropriate use depends on who is present, what has just happened, and what risks are acceptable.

Humanoid form has returned to prominence partly because much of the built environment assumes human reach and geometry. Stairs, doors, shelves, handles, tools, vehicles, kitchens, and workstations can be used without redesign if a robot has sufficiently human-like mobility and manipulation. This compatibility is an engineering advantage, not a guarantee of economic superiority. Wheels are often more efficient than legs, and specialized machines can be faster, safer, cheaper, and easier to maintain.

Total system performance matters more than the success of a demonstration. Hardware cost, utilization, supervision, teleoperation, maintenance, energy use, insurance, downtime, and rare failures all determine whether a robot produces value. A system that completes most tasks autonomously may still be uneconomic if the remaining failures require frequent expert intervention or create large liability.

Early general-purpose deployments are therefore most plausible in settings that are more variable than a traditional production line but more controlled than a household. Warehouses, selected manufacturing operations, laboratories, inspection, material handling, and dangerous industrial environments allow some flexibility while limiting the number and severity of possible failures.

Homes present a harder case because they combine physical variability with intimate social and legal stakes. A household robot would operate around children, pets, medication, food, private conversation, fragile property, and people who have not been trained to accommodate a machine. Reliability requirements therefore include task completion, privacy, consent, safety, and the ability to recognize situations in which the system should stop and seek help.

Embodiment also changes the economics of learning. Language models can generate large numbers of poor drafts at low marginal cost. Physical trials consume time and energy, wear hardware, alter the environment, and may create danger. Simulation, teleoperation, and synthetic data can reduce these costs, but a deployed robot will often be prevented from exploring freely because unsafe mistakes are unacceptable.

A robot must represent weight, friction, balance, occlusion, latency, wear, and the persistence of objects outside its current view. Human cognition developed through bodily interaction with these constraints and through dependence on other vulnerable bodies. It remains an open empirical question how much common sense, causal understanding, and social judgment can be learned from recorded data, and how much requires continued physical interaction.

Robotics makes that question testable. Meaningful success requires reliable action over long periods, detection of the system’s limits, recovery from ordinary failure, and institutions that assign responsibility when something goes wrong; a controlled demonstration establishes only a small part of this case.

Space as terrestrial infrastructure

For the next decade, the most consequential uses of space are likely to be experienced on Earth. Satellites already support navigation, timing, communications, weather forecasting, agriculture, shipping, finance, disaster response, mapping, climate observation, internet access, and military operations. Many economic and public systems now depend on orbital infrastructure without making that dependence visible to ordinary users.

Space Foundation estimates that the global space economy reached $686 billion in 2025, with commercial activity accounting for most of the total. Much of the measured value came from equipment and services used on Earth, with work performed in orbit contributing a smaller share. Orbit has become an additional layer of terrestrial infrastructure: remote in location, but integrated into daily coordination.

The comparison to a nervous system is useful in a limited sense. Satellite networks carry timing, position, observation, and communication signals across otherwise separate systems. Their failure can therefore propagate far beyond the damaged spacecraft. Jamming, spoofing, cyberattack, physical destruction, or the loss of a major constellation could disrupt users and institutions distant from the original event.

Orbital congestion creates a different class of risk. Debris from a collision or fragmentation can remain in orbit and increase hazards for operators that had no role in creating it. ESA describes usable orbital regions as a finite resource and warns that mitigation may eventually need to be supplemented by active removal. The problem has the structure of a commons: individual actors benefit from access, while the long-term cost of congestion and debris is distributed across many users.

Governance becomes harder after dependence has formed. Operators then face incentives to keep launching, governments view satellite capacity as strategically important, and the cost of stricter rules is immediate while the benefit is shared and delayed. The history of orbital infrastructure may therefore repeat a broader civilizational pattern in which coordination arrives after private and national systems have already become difficult to unwind.

Reusable heavy launch could change the scale and tempo of this development. Starship’s thirteenth flight test launched on July 24, 2026; it was the second flight of the V3 vehicles and the first Starship mission to deploy next-generation Starlink V3 satellites. The program has demonstrated substantial engineering progress. Rapid, fully reusable operation at a dependable cadence and cost remains a goal whose operating conditions have yet to be proven.

If that goal is achieved, lower launch cost and greater payload mass could make larger telescopes, satellite servicing, lunar logistics, planetary missions, and orbital manufacturing experiments easier to attempt. Launch is only one constraint. Space systems still require power, communications, thermal control, radiation tolerance, maintenance, regulation, insurance, and sufficient demand to support continued operation.

NASA’s revised Artemis sequence illustrates the difficulty of integrating human spaceflight systems. Artemis III is planned as a crewed low-Earth-orbit demonstration in 2027, intended to test systems and rendezvous operations before Artemis IV attempts a crewed lunar landing in 2028. Staged testing is necessary because hardware failures can destroy equipment, endanger crews, and delay dependent missions; many problems cannot be corrected remotely after launch.

The most important space event of the decade could still be a scientific discovery. Credible evidence of past life on Mars, a strong biosignature in an exoplanet atmosphere, or unexpected chemistry from an ocean world would alter scientific theories immediately and public self-understanding more slowly.

The interpretation would depend on what was found. Evidence that life arises readily under suitable conditions would suggest that biology is a recurrent feature of the universe. Evidence that Earth is unusually rare would increase the apparent contingency of life here. A radically unfamiliar form of life could expose assumptions built into our concepts of organism, intelligence, and evolution.

No discovery would give humanity a single interpretation. Religious traditions, scientific communities, states, and cultures would absorb it differently. Its importance would lie in changing the factual background against which those interpretations develop. Space exploration matters in part because it tests whether categories formed from one planet are adequate to describe life and intelligence more generally.

Space has become infrastructure for Earth, a strategic environment with shared risks, and a source of evidence that may change how human beings understand the conditions of their own existence. Treating it only as escape or distant ambition misses its present role.

The next ten years

The next decade is unlikely to resemble a single revolution with a clear beginning and uniform effect. New systems will enter institutions with different laws, resources, professional norms, and tolerance for error. A capability may be routine in one laboratory, prohibited in one school, and unavailable elsewhere because the necessary electricity, data, or expertise is missing.

AI adoption will vary in depth as well as speed. Some laboratories will use models to generate hypotheses, write code, select experiments, analyze data, and control instruments. Some companies will allow agents to execute workflows across software systems. Others will add a chat interface while leaving authority, data access, evaluation, and organizational structure largely unchanged.

As integration deepens, AI will become less visible as a separate product. Models will classify, prioritize, summarize, recommend, and act inside search, communication, software, record-keeping, research, and administration. Users may see only the result, which makes provenance, audit trails, testing, and clear thresholds for human intervention more important.

The economic effects will arise through organizational redesign, across many uneven changes in task, authority, and demand. Some teams will become smaller; others will expand output with similar headcount; some occupations will grow as lower costs create new demand. Workers may retain their jobs while their pace, autonomy, tasks, and promotion paths change substantially.

Entry-level work deserves particular attention. Organizations can preserve senior experts while removing the tasks through which new experts once developed. The resulting shortage of judgment would appear with a delay, after the productivity benefits had already been recorded. Firms and professions will need explicit training systems if automation removes apprenticeship from routine work.

Political competition will increasingly involve compute, energy, semiconductors, data, scientific talent, industrial production, and infrastructure.

Governments will describe these resources in terms of sovereignty and security, often with legitimate reasons. The danger is that competition becomes a standing justification for secrecy, surveillance, centralization, and restrictions on scientific exchange. Security measures need specific purposes, independent review, and limits in duration.

Cultural life will contain more synthetic material and more uncertainty about origin. Some people will form durable attachments to machine tutors, collaborators, and companions. Others will seek settings where human authorship and presence are guaranteed. Automated services may broaden access while sustained human attention becomes more expensive, making consent, reciprocity, disclosure, and responsibility central questions.

Science will gain more powerful systems for generating hypotheses and navigating design spaces. Its bottlenecks will shift toward measurement, replication, negative results, and evaluators resistant to superficial success. Fields with rapid, objective feedback will move faster than fields whose outcomes are delayed or ethically contested.

Robotics will advance first where environments can be partly controlled and failure costs bounded. The most informative milestone will be sustained operation with low supervision and transparent failure rates; a single impressive demonstration establishes less. Neural interfaces will restore communication or control to more patients while remaining invasive and specialized. Connectomics will produce richer maps without resolving consciousness or personal identity.

Climate risk and low-carbon construction will intensify together. The energy transition will be judged through capacity, emissions, grid reliability, and the extent to which adaptation protects people from heat, flood, fire, water stress, and insurance retreat. Space infrastructure will become more important to communication, navigation, observation, and security, increasing both investment and the consequences of congestion, debris, and concentrated ownership.

Military technology will continue advancing because states evaluate their own restraint under uncertainty about the restraint of others. Defensive preparation can appear offensive and encourage reciprocal investment.

Limiting this dynamic requires arms control, communication, verification, civilian authority, and institutions that prevent security organizations from defining every public problem in their own terms.

The most consequential outcomes will emerge where these domains alter one another’s constraints. AI can improve energy-system design while increasing electricity demand. Aging societies may seek care robots before they are reliable enough for intimate settings. Military competition can fund research while reducing openness. Neural interfaces depend on machine learning and on companies that control device software. Satellite systems support climate observation, commerce, and warfare at the same time.

These interactions make forecasts organized by technology incomplete. Civilizational change occurs when an energy limit slows computation, an economic incentive changes education, a security concern changes science, or a medical device changes the legal meaning of agency. The next decade will be defined less by one decisive invention than by how societies manage these coupled transitions.

What would count as a good future?

A defensible technological future would enlarge people’s ability to understand and shape their lives while preserving the institutions needed to identify error, assign responsibility, and reverse harmful decisions. This standard begins with average performance and then asks what happens to a particular person when the system is wrong, who can challenge the result, and whether correction is practically available.

The answer differs by domain. A medical system should improve diagnosis or care while allowing physicians and patients to examine, question, and report its failures. A government system should keep statistical accuracy subject to appeal. A scientific model should increase the rate of useful conjecture without allowing plausible language to substitute for evidence. An organization should capture productivity gains without eliminating the training through which future workers acquire judgment. A neural interface should restore agency while protecting consent, privacy, access, and long-term support.

Greater capability therefore requires clearer responsibility. Responsibility here means tracing the decisions through which people and institutions designed the system, selected its data, authorized its use, defined acceptable error, monitored its performance, and decided when to rely on it. When no one can answer for a consequential outcome, technical power has been separated from political and moral accountability.

A better future would also preserve meaningful forms of contribution as machines perform more cognitive work. Higher productivity can increase abundance; distribution of income, recognition, education, social membership, and political voice remains a political question. A society can become more productive while making many of its members less secure and less able to influence the institutions that govern their lives.

Human dignity should not depend on a shrinking list of tasks at which people outperform machines. That standard would already fail children, elderly people, people with severe disabilities, and anyone temporarily unable to work. Moral standing rests on the fact that people have experiences, relationships, vulnerabilities, interests, and claims that can be respected or violated. Economic contribution supports independence and participation; a person’s basic worth comes before that.

Machines may produce better answers than most people in an increasing number of domains. Superior task performance leaves open which tasks a society should pursue, which risks it may impose, how benefits should be distributed, and which decisions require consent. Models can assist reasoning about these questions, but human institutions retain authority because human lives bear the consequences.

Institutional capacity may become the scarce capacity of the next decade: the ability to act quickly while exposing decisions to criticism, evidence, and appeal. Governments need competence together with courts, audits, elections, and public records that permit correction. Scientists need faster analysis together with replication and independent judgment. Companies need incentives to account for harms absent from their balance sheets. Citizens need public forums that acknowledge complexity without turning uncertainty into paralysis.

This essay has used memory as one thread through the history of civilization. Memory preserves what previous generations learned, including records of failure and repair. Desire supplies another force: the attempt to create arrangements that have not existed before. A society governed only by inherited categories cannot respond to new conditions; one driven only by imagined possibility loses contact with accumulated consequence.

Artificial intelligence joins these forces in a distinctive way. It is trained on records of past human expression and used to generate candidate futures: sentences, designs, programs, strategies, and hypotheses. It makes cultural memory responsive and productive, while also inheriting the omissions, distortions, and power relations embedded in that memory.

Intelligence names a capacity to model, predict, and act. Its value depends on the purposes it serves, the constraints under which it operates, and the people who can object when it is used against them. The next decade will test whether practical wisdom—judgment about ends, consequences, and responsibility—can develop quickly enough to govern capabilities that become ordinary before they become fully understood.

That test will be decided in classrooms that distinguish producing an answer from understanding why it is true; in firms that decide whether productivity gains expand workers’ agency or narrow their paths; in hospitals that determine when a model may influence treatment; in governments that choose whether efficiency justifies opacity; in families that decide what forms of synthetic presence can be trusted with care; and in laboratories that require a compelling hypothesis to encounter evidence capable of rejecting it.

These decisions appear local, but together they determine the social form of a technology. By the time a society agrees on a name for the new order, that order may already be present in its routines, institutions, and expectations. The future takes form through such choices long before a society gives the resulting order a stable name.