Over the next 100 years, the most defensible expectation is not a single \u201cAI event,\u201d but a prolonged civilizational reorganization around machine cognition. Artificial intelligence is already becoming cheaper, more capable, and more widely deployed: frontier systems improved sharply on difficult benchmarks in 2024, open-weight models nearly closed some performance gaps with closed models, and the cost of querying a GPT-3.5-class model fell by more than 280-fold between late 2022 and late 2024. At the same time, organizational use of AI rose sharply, and OECD data show both firm adoption and individual use accelerating in 2025. \ue200cite\ue202turn30search0\ue202turn29search0\ue202turn29search1\ue202turn28view0\ue202turn28view3\ue201<\/p>\n\n\n\n\n
Robotics is also advancing, but on a slower, more uneven timetable than software AI. As of 2024, the world installed about 542,000 industrial robots, bringing the global operational stock to 4.664 million, while medical and professional service robots grew quickly in some niches. Yet robotics remains constrained by embodiment, safety, real-time control, limited multimodal data, edge compute limits, and return-on-investment hurdles. In other words, software intelligence is diffusing faster than physical autonomy. \ue200cite\ue202turn22view0\ue202turn22view1\ue202turn31search11\ue202turn31search5\ue202turn31search14\ue202turn32search6\ue201<\/p>\n\n\n\n\n
The central lesson from previous technological revolutions is that civilization changes most when a technology becomes general-purpose infrastructure. Agriculture created food surplus, settlement, hierarchy, and specialization. Writing and then printing expanded storage and transmission of knowledge. Industrialization and electrification changed not only tools, but factory design, household life, and state capacity. Digital networks lowered the cost of coordination and information distribution, while also enabling platform concentration, network effects, and new forms of misinformation. AI most closely resembles a new general-purpose technology: pervasive, improvable, and innovation-spawning. But like electrification and computing before it, its largest effects are likely to arrive through complements such as organizational redesign, new legal regimes, energy build-out, and institutional adaptation rather than from model capability alone. \ue200cite\ue202turn19search1\ue202turn19search8\ue202turn19search3\ue202turn34search7\ue202turn34search1\ue202turn34search10\ue202turn41search4\ue202turn41search5\ue201<\/p>\n\n\n\n\n
The next century is therefore likely to be shaped less by whether AI becomes \u201csmart,\u201d and more by who controls compute, energy, data, robotics supply chains, and the rules of deployment. Electricity for data centers is already a binding question: global data-center consumption was about 415 TWh in 2024, roughly 1.5% of world electricity use, and had been growing around 12% annually. The International Energy Agency also argues that countries able to provide electricity at speed and scale will be best positioned to benefit from AI. Recent U.N. analysis adds that AI\u2019s footprint is also about water, land, and cooling, not just compute. \ue200cite\ue202turn11search6\ue202turn11search13\ue202turn11search2\ue202turn36search0\ue201<\/p>\n\n\n\n\n
The labor story will be mixed. The IMF estimates that nearly 40% of global employment is exposed to AI, rising to about 60% in advanced economies, yet current real-world use still leans more toward augmentation than full automation. Field evidence shows productivity gains in some domains, especially for novices, but also shows that AI performance is uneven and task-dependent rather than universally substitutive. The most likely future, therefore, is neither \u201cmass unemployment tomorrow\u201d nor \u201cno real impact,\u201d but decades of occupational churn, re-bundling of jobs, new entry barriers, and growing pressure to redesign education, social insurance, and ownership structures. \ue200cite\ue202turn22view3\ue202turn40search1\ue202turn20search0\ue202turn20search9\ue202turn20search4\ue201<\/p>\n\n\n\n\n
The most likely 2126 is a deeply hybrid civilization: routine cognitive work heavily automated, significant parts of logistics and care robotized, medicine increasingly predictive and personalized, science accelerated by autonomous systems, and governance partially mediated by AI. But it is not likely to be frictionless, uniform, or post-political. Demographic aging, climate stress, wars, supply shocks, and regulatory fragmentation will interact with AI at every stage. Global population is projected by the U.N. to peak in the mid-2080s, while the share of older populations rises sharply and urbanization continues. Those secular facts matter because they make labor scarcity, care demand, infrastructure adaptation, and public-finance stress at least as important as algorithmic progress. \ue200cite\ue202turn12search12\ue202turn12search5\ue202turn12search3\ue202turn13search0\ue201<\/p>\n\n\n\n\n
Key findings<\/strong><\/p>\n\n\n\n\n This assessment uses strategic foresight rather than prediction. It follows the standard futures-studies logic described by OECD and the U.K. Government Office for Science: systematically scan evidence, distinguish trends from weak signals, construct multiple plausible futures, and stress-test conclusions against uncertainty rather than pretending to forecast a single path. That approach is especially important for a 100-year horizon, where institutional responses, wars, climate change, and demographic shifts can dominate raw technology curves. \ue200cite\ue202turn35search0\ue202turn35search8\ue202turn35search2\ue202turn35search7\ue201<\/p>\n\n\n\n\n The report uses the following working categories.<\/p>\n\n\n\n\n As of June 2026, the baseline state of AI is unusually strong in text, code, multimodal perception, search, and constrained agentic workflows. Stanford\u2019s 2025 AI Index showed that difficult benchmarks introduced in 2023 improved dramatically by 2024; the best and tenth-best frontier models also compressed toward each other, meaning high-end capability is becoming less exclusive. Smaller models are getting much better, and new reasoning-heavy paradigms can sharply improve performance on some classes of tasks. At the same time, current frontier agents still have limited autonomous reliability on long, messy tasks: METR estimates a roughly 50-minute 50%-success task horizon for leading frontier systems in early 2025, though that horizon had been doubling at about a seven-month pace. \ue200cite\ue202turn30search0\ue202turn29search1\ue202turn39search0\ue202turn39search3\ue201<\/p>\n\n\n\n\n The present limits are just as important as the present gains. Stanford\u2019s 2025 technical review found that complex reasoning and planning remain unreliable, especially on tasks requiring formal correctness beyond the model\u2019s training distribution. The 2026 International AI Safety Report also emphasizes that general-purpose systems remain difficult to interpret, difficult to assure across open-ended use cases, and hard to evaluate robustly before deployment. In practice, this means current AI is often strongest as a fast pattern engine and workflow component, not a universally trustworthy autonomous actor. \ue200cite\ue202turn30search0\ue202turn27view4\ue202turn37search0\ue202turn37search12\ue201<\/p>\n\n\n\n\n The robotics baseline is more mature in factories and logistics than in homes. Industrial robots continue to scale, especially in Asia; collaborative robots are becoming more common; and service robots are growing in logistics, cleaning, labs, and medical settings. But robotization is still highly concentrated: recent economic work finds that robots remain a small share of aggregate equipment spending and far less economy-wide than past information technologies. The likely implication is diffusion through specific high-value environments first, rather than a sudden universal humanoid takeover. \ue200cite\ue202turn22view0\ue202turn22view1\ue202turn29search2\ue202turn32search6\ue201<\/p>\n\n\n\n\n The hard constraints are fivefold.<\/p>\n\n\n\n\n Hardware and compute.<\/strong> Machine-learning hardware keeps improving, with annual gains in speed, price-performance, and energy efficiency, but frontier-model training compute is still rising much faster than hardware efficiency. Stanford reports hardware performance doubling about every 1.9 years, while notable-model training compute has been doubling around every five months. This means capability gains remain tied to large capital expenditures and concentrated supply chains, not purely to clever software. ASML\u2019s 2025 reporting also directly tied strong demand in advanced logic and memory equipment to AI-related investment. \ue200cite\ue202turn29search0\ue202turn29search1\ue202turn6view2\ue201<\/p>\n\n\n\n\n Energy, cooling, water, and land.<\/strong> There is no AI without physical infrastructure. Data centers consumed about 415 TWh in 2024, and the IEA identifies electricity availability as a decisive factor in AI development. Cooling can account for a substantial share of data-center power use, and recent U.N. work warns that the AI boom also brings substantial water and land footprints. A world of trillion-parameter models, always-on agents, sensor networks, and widespread robotics is therefore not merely a software future; it is an infrastructure future. \ue200cite\ue202turn11search6\ue202turn11search13\ue202turn36search0\ue202turn36search4\ue201<\/p>\n\n\n\n\n Data and legal access.<\/strong> Epoch AI argues that, on then-current trends, high-quality public human-generated text could become scarce for frontier pretraining between 2026 and 2032, while research on \u201cmodel collapse\u201d shows risks from recursively training on synthetic outputs. The U.S. Copyright Office has also described substantial legal uncertainty around the use of copyrighted material in generative-AI systems. The likely short-run response is more synthetic data, licensed corpora, multimodal data collection, simulation, and reinforcement learning rather than endless free scaling of web text. \ue200cite\ue202turn7search2\ue202turn7search4\ue202turn7search5\ue201<\/p>\n\n\n\n\n Economic realism.<\/strong> AI adoption is rising rapidly, but measured business value is still often modest and localized. Stanford reports that most companies seeing financial gains from AI still report small cost savings or revenue uplifts in many functions. Field evidence is also mixed: call-center workers with AI support saw strong gains, especially novices; consultants did better on some tasks but could also fail outside the \u201cjagged frontier\u201d; and current real-world usage still leans more toward augmentation than full automation. This suggests a long diffusion curve rather than immediate economy-wide replacement. \ue200cite\ue202turn28view0\ue202turn20search0\ue202turn20search4\ue202turn40search1\ue201<\/p>\n\n\n\n\n Human factors, trust, regulation, and institutional capacity.<\/strong> Public optimism about AI has risen in some countries, yet trust in fairness and data stewardship has declined, and support for regulation is broad. Governments themselves face legacy IT, skills gaps, thin evaluation capacity, and weak implementation guidance. The political baseline, therefore, is neither laissez-faire nor coherent global governance; it is fragmented acceleration under uneven guardrails. \ue200cite\ue202turn28view2\ue202turn24search0\ue202turn24search2\ue202turn27view2\ue202turn27view3\ue201<\/p>\n\n\n\n\n The most realistic near-term trajectories are therefore these: AI copilots and agents embedded in enterprise software; more autonomous but still bounded systems in logistics, labs, warehouses, mines, and transport corridors; medical AI that continues to grow faster in imaging, triage, and decision support than in fully autonomous care; public-sector AI pilots that expand where data and incentives are aligned; and education systems increasingly shaped by AI tutors, assessment tools, and curriculum support. Less realistic in the next two decades are mass-market general household humanoids, full post-scarcity economics, and robustly aligned autonomous military decision chains accepted worldwide. \ue200cite\ue202turn22view0\ue202turn22view1\ue202turn28view1\ue202turn14search1\ue202turn27view2\ue202turn10search2\ue202turn10search4\ue201<\/p>\n\n\n\n\n Uncertainty rises steeply with time. The timeline below therefore emphasizes direction of change, major bottlenecks, and confidence level rather than pretending to offer precise century-scale sequencing.<\/p>\n\n\n\n\n Two timeline judgments matter most. First, the 2026\u20132040 window is about diffusion and reorganization<\/strong>, not science-fiction discontinuity. Second, the period after 2040 is increasingly a governance question<\/strong>: who owns the systems, who gets the gains, and who decides what remains human by law, culture, and design. Those are not secondary issues; they are the main path-dependent variables. \ue200cite\ue202turn28view0\ue202turn24search0\ue202turn24search2\ue202turn27view2\ue201<\/p>\n\n\n\n\n The direct effects of AI and robotics matter, but the larger civilizational question is how they cascade.<\/p>\n\n\n\n\n Several higher-order consequences follow from this matrix.<\/p>\n\n\n\n\n First, the occupation<\/strong> is more fragile than the job bundle<\/strong>. Historical automation usually reallocated tasks rather than eliminating all work at once, and current evidence still shows strong augmentation. But AI targets many of the specific tasks that once justified large entry-level knowledge-work cohorts. That makes career ladders vulnerable even when \u201cemployment\u201d in aggregate remains resilient. The result may be a period in which apprenticeship, credentialing, and early-career signaling are more disrupted than total employment. \ue200cite\ue202turn40search1\ue202turn20search0\ue202turn20search4\ue202turn22view3\ue201<\/p>\n\n\n\n\n Second, the biggest redistribution may be from labor to infrastructure ownership<\/strong>. Because advanced AI depends on chips, power, cooling, cloud capacity, proprietary models, and regulation, the relevant asset is not just software. It is the whole stack. That favors already-capitalized firms and state-backed ecosystems unless policy deliberately broadens access. This is one reason AI may widen inequality within and between countries even while raising aggregate productivity. \ue200cite\ue202turn29search1\ue202turn11search13\ue202turn15search5\ue202turn15search7\ue201<\/p>\n\n\n\n\n Third, AI will likely alter the meaning of expertise<\/strong>. When machine systems can draft, explain, summarize, simulate, and diagnose at high speed, raw information recall matters less, while judgment, trustworthiness, negotiation, taste, and accountability matter more. Universities, professions, and licensing systems will therefore face a double challenge: proving they still certify something scarce, and integrating AI without collapsing their own value proposition. This resembles earlier general-purpose technologies, which changed institutions most when they forced organizational redesign rather than simple tool substitution. \ue200cite\ue202turn19search1\ue202turn19search8\ue202turn19search3\ue202turn14search0\ue202turn27view2\ue201<\/p>\n\n\n\n\n Fourth, several institutions are likely to change at different speeds.<\/p>\n\n\n\n\n Two cross-sector conclusions stand out. First, AI excels earliest where the world is already digitized, instrumented, and high-volume<\/strong>. Second, robotics excels earliest where environments are structured, repetitive, and economically legible<\/strong>. Homes, politics, childcare, open-world retail, and messy field environments are therefore later and harder than software demos suggest. \ue200cite\ue202turn28view0\ue202turn22view0\ue202turn31search11\ue202turn31search14\ue201<\/p>\n\n\n\n\n The wild cards below are not forecasts. They are probability judgments anchored to current evidence, expert disagreement, and visible bottlenecks. Because long-range technological forecasting is noisy, the ranges are intentionally broad.<\/p>\n\n\n\n\n The four scenarios below are internally coherent futures, not ranked predictions.<\/p>\n\n\n\n\n Managed Abundance.<\/strong> AI becomes a widely regulated utility layer. Governments and firms build public-interest guardrails, invest in power and data infrastructure, and distribute gains through productivity-sharing, public services, and education reform. Daily life includes ubiquitous personal tutors, automated paperwork, precision medicine, robotic logistics, and high trust in machine-assisted systems because appeals, transparency, and accountability improve alongside deployment. Human purpose shifts away from routine labor toward care, craft, science, local leadership, and self-directed creation. This scenario requires strong state capacity, broad access to compute, and successful management of concentration risks. \ue200cite\ue202turn27view2\ue202turn24search2\ue202turn15search9\ue202turn14search0\ue202turn11search13\ue201<\/p>\n\n\n\n\n Corporate Feudalism.<\/strong> A handful of firms and their partner states control the critical stack: compute, cloud, identity, models, and agent platforms. Most people access civilization through leased AI services rather than owning productive systems directly. The economy is productive, but bargaining power weakens for many workers and small firms. Education is personalized but platform-dependent; culture is abundant but heavily intermediated; politics becomes more about regulating private infrastructures than directing public ones. Family and community remain emotionally central, but daily life is embedded in subscription hierarchies and proprietary ecosystems. This scenario is supported by current concentration dynamics in investment, infrastructure, and public-sector dependence on private vendors. \ue200cite\ue202turn28view0\ue202turn29search1\ue202turn11search13\ue202turn27view2\ue201<\/p>\n\n\n\n\n Fragmented AI Worlds.<\/strong> The world divides into competing AI civilizational blocs with divergent values, stacks, legal systems, censorship norms, and military doctrines. International governance exists, but mainly at the margins. Cross-border research becomes harder; standards splinter; supply chains regionalize; autonomous systems are common in border control, cyber operations, and military logistics. Daily life varies sharply by regime type: some societies emphasize rights and due process, others state surveillance and industrial policy, others firm-led governance. This scenario becomes more likely if energy, chips, models, and military AI remain strategic sovereignty assets. \ue200cite\ue202turn24search0\ue202turn24search2\ue202turn10search1\ue202turn10search4\ue201<\/p>\n\n\n\n\n Radical Transformation.<\/strong> Capability breakthroughs push machine autonomy, human-machine interfaces, and synthetic agents far beyond today\u2019s baseline. Some jurisdictions recognize novel forms of digital personhood or semi-autonomous institutional actors; others ban them. Work as a universal social organizer weakens sharply. Entire sectors run with human oversight only at exception points. Education becomes continuous cognitive augmentation. Culture splits between biological-authentic and synthetic-native forms. This future is not impossible, but it depends on breakthroughs that remain scientifically unresolved today, including the nature of general intelligence, consciousness, and controllability. \ue200cite\ue202turn23search8\ue202turn33search0\ue202turn17search3\ue202turn27view4\ue201<\/p>\n\n\n\n\n The best overall characterization is this: AI is very likely another general-purpose technology, and quite possibly a civilization-level transition, but it is not yet clearly a new evolutionary phase.<\/strong> The evidence strongly supports pervasiveness, continuous improvement, and innovation-spawning complementarities, which are the classic hallmarks of a general-purpose technology. What remains uncertain is whether machine systems will cross thresholds of autonomy, strategic agency, and person-like status that justify speaking about a deeper evolutionary rupture. \ue200cite\ue202turn19search1\ue202turn19search3\ue202turn19search8\ue202turn29search1\ue201<\/p>\n\n\n\n\n A concise confidence map is useful.<\/p>\n\n\n\n\n Open questions and limitations<\/strong><\/p>\n\n\n\n\n Several questions remain fundamentally unresolved. The strongest are whether current scaling trends translate into robust long-horizon autonomy outside software-like environments; whether AI\u2019s net effect on labor is ultimately productivity-complementing or more displacement-heavy; whether offensive or defensive military and cyber applications benefit more from future capability gains; and whether consciousness science will mature enough to ground any serious claims about machine moral status. The post-2060 portions of this report are therefore best read as structured scenario space, not a single expected path. \ue200cite\ue202turn39search0\ue202turn37search7\ue202turn33search0\ue202turn17search3\ue201<\/p>\n\n\n\n\n Core measurement, trends, and foresight<\/strong><\/p>\n\n\n\n\n Economy, labor, and institutions<\/strong><\/p>\n\n\n\n\n Robotics, infrastructure, and energy<\/strong><\/p>\n\n\n\n\n Medicine, science, and human augmentation<\/strong><\/p>\n\n\n\n\n Security, governance, and global order<\/strong><\/p>\n\n\n\n\n Historical and conceptual framing<\/strong><\/p>\n\n\n\n\n Further reading<\/strong><\/p>\n\n\n\n\n Executive Summary Over the next 100 years, the most defensible expectation is not a single \u201cAI event,\u201d but a prolonged civilizational reorganization around machine cognition. Artificial intelligence is already becoming cheaper, more capable, and more widely deployed: frontier systems improved sharply on difficult benchmarks in 2024, open-weight models nearly closed some performance gaps with closed … <\/p>\n\n\t
<\/li>\n<\/ul>\n\n\n\nForecast Methodology and Baseline<\/h2>\n\n\n\n
Century Timeline<\/h2>\n\n\n\n
Cascading Effects on Civilization<\/h2>\n\n\n\n
Sector Deep Dives<\/h2>\n\n\n\n
Wild Cards and Scenario Futures<\/h2>\n\n\n\n
Strategic Conclusions and Confidence Ratings<\/h2>\n\n\n\n
Bibliography and Further Reading<\/h2>\n\n\n\n
\n\t
<\/li>\n<\/ul>\n\n\n\n\n\t
<\/li>\n<\/ul>\n\n\n\n\n\t
<\/li>\n<\/ul>\n\n\n\n\n\t
<\/li>\n<\/ul>\n\n\n\n\n\t
<\/li>\n<\/ul>\n\n\n\n\n\t
<\/li>\n<\/ul>\n\n\n\n\n\t