10 Years Later · The Age of AI
Seven areas · Select a tile to explore its charts.
Counts include dedicated and shared charts; a shared chart appears in each relevant area. Editorial assessments · reviewed 17 October 2026. Colours combine direction and consequences; they are not automatic risk scores or labels for how fast a curve grows.
01 / ACCELERATING WORLD - MONITORING SEVEN AREAS
Sixteen indicator cards with 20 main graph panels, plus recovery details and thematic and scenario explorers. Each panel has its own dates, time window and trend context.
10 years, 50 years and All available count back from each panel’s latest historical year. The urbanisation panel adds a labelled 2026 projection to its historical window; the waste panel counts back from 2026, including labelled projections. Trend tags and historical CAGR tiles keep their stated reference periods. Open “Explore” for further analysis, source notes and data downloads.
02 / PEOPLE & DISTRIBUTION
Average progress can hide different experiences. Distribution, ageing and the geography of growth belong beside the headline totals.
DISTRIBUTION
Compare the top tenth with the bottom half. Start with wealth, or switch to income for a different view of inequality. The world view and the distribution within a country answer different questions.
WID estimates, equal-split adults. Income is pre-tax national income after pensions; wealth is net personal wealth. Current downloaded histories end in 2023. They are not the separate 2025 snapshots in WIR 2026.
World Inequality Database ↗URBAN CHANGE
Not necessarily. A city can grow outward while its population grows. Compare population, people per square kilometre and concentration in a country's cities.
Six illustrative cities, not a representative global sample. UN WUP 2025 uses harmonised DEGURBA settlement extents, not administrative boundaries. Density is not a direct measure of crowding or service quality.
UN World Urbanization Prospects 2025 ↗UN World Population Prospects 2024, processed by Our World in Data. Estimates end in 2023; 2024 onward, including 2026, is a medium-variant projection. Dashed lines are projections, not observed outcomes. Source ↗
03 / THE COMING OF AI
AI connects the trends above: computation and data become tools that can help people analyse, create and act. The question is how to turn expanding capability into useful, widely shared progress.
MEASURING CAPABILITY
Selected tests show rapid gains in language, scientific questions, coding and computer use. Several cross their human reference level; others remain below it. These are different tasks, not one measure of intelligence.
A higher benchmark score opens possibilities. Dependable performance outside the test requires separate evidence.
Benchmarks can saturate, contain errors or reward familiarity with a test. Read this as evidence of changing task capability, not proof of general judgment or safe autonomy. Source and interpretation ↗
WHY NOW?
Useful examples, careful selection and checked demonstrations shape what a model can learn. Quality and coverage matter alongside volume.
Architectures, training objectives and feedback change how a model uses its examples. Pre-training builds broad patterns; post-training shapes behaviour and task performance.
Accelerators, memory and parallel processing make larger training runs and more experiments feasible. Efficient software helps turn hardware into useful work.
Better examples improve learning; better methods use resources more effectively; more usable computation expands what can be tried.
Conceptual explanation, not a measured allocation of progress between three causes.
The transistor curve spans more than half a century. Moore’s roughly two-year doubling became an industry reference, not a law guaranteeing performance. Modern GPUs add parallel computation and advanced packaging; some combine multiple compute dies. Transistor totals therefore do not measure speed, efficiency or useful work on their own.
Memory, data movement, numerical precision and software also matter. The same hardware can deliver different results depending on how it is used.
Explore the transistor and computing charts in the Atlas ↑FROM POSSIBILITY TO DIFFUSION
Stanford’s 2025 AI Index reports that the price of querying a model at a fixed GPT-3.5-level benchmark threshold fell from $20 to $0.07 per million tokens between November 2022 and October 2024—more than a 280-fold reduction. This compares a particular performance threshold, not every model or the full cost of deploying AI.
Lower prices can widen access. Useful adoption still depends on skills, reliable infrastructure, suitable workflows and who controls the tools.
Stanford AI Index 2025 · inference prices ↗POSSIBLE DIRECTIONS · NOT A FORECAST
Agents could coordinate longer workflows across software. The test is successful completion with bounded permissions, clear accountability and a way to stop or undo actions.
AI could help propose molecules, materials and experiments. The test is independently validated results, useful treatments and wider access—not simply more generated candidates.
Robotics could extend assistance into factories, care and daily life. The test is reliable operation in varied environments, at an acceptable cost and level of risk.
These are possible paths and proposed tests of progress. Their speed and reach remain uncertain; no date for general intelligence is implied.
Illustrative opportunities and pressures to examine, not measured net effects.
| Area | Potential contribution | Question of control |
|---|---|---|
| Population | Care assistance and planning for changing needs | Does support preserve autonomy and reach underserved communities? |
| Economy | Productivity and new services | Who gains, who is displaced, and how concentrated are the returns? |
| Technology | Faster design, coding and experimentation | Can testing and oversight keep pace with deployment? |
| Data | Better analysis and cyber defence | How do we limit surveillance, misuse and automated attacks? |
| Environment | Resource optimisation and environmental monitoring | Do verified savings exceed added energy and material demands? |
| Health | Research and clinical assistance | Are benefits validated, affordable and reliable across populations? |
| Freedom & society | Access to knowledge and participation | Can people challenge decisions and retain independent judgment? |
AI could increase the pace of change and improve our ability to respond. Which effect dominates depends on deployment, institutions and the distribution of power. The aim is a sustainable pace of useful progress, with the capacity to detect harm, correct mistakes and share the gains.
Continue to Acceleration & Absorption ↓04 / ACCELERATION AND ABSORPTION
The shape of a curve changes the demands a system must absorb. Compounding, connectedness and cycles help explain those demands; our ability to observe, act and correct determines how well we respond.
A rising level is not enough to establish an increasing percentage growth rate. Read the shape, the time horizon and what is growing.
Equal absolute additions. The level rises, while percentage growth declines.
A constant percentage applied to a growing base. Absolute additions become larger.
Growth first gathers pace, then slows towards a limit. Stabilisation is a possibility, not a universal outcome.
Illustrative shapes on a shared linear index scale, starting at 1. Time is schematic; these are not fitted histories.
Each increase expands the base for the next. Several drivers can also reinforce one another. A small sustained rate can create a large change in scale.
A change in one area can alter demands in another: more data can require more energy and stronger security. Interactions may amplify or dampen pressure.
Feedback, constraints and recurring cycles can interrupt expansion or bring reversal. Neither endless growth nor inevitable collapse follows from a rising curve alone.
Every line starts at 1. The annual percentage stays constant; the amount added each year grows.
Linear axes in both panels, with different vertical ranges. Level = (1 + annual rate)years. Constant rates, no shocks or saturation: mathematical scenarios, not forecasts. Multipliers show the level reached, not the sum of annual flows.
If the rising quantity is an annual physical burden, each year can demand more collection, treatment or infrastructure. Unresolved flows can accumulate into a stock even when their growth slows. Capacity, prevention and ecological limits determine the consequences.
FROM SCALE TO RESPONSE
Absorption means the practical ability to handle change: understand it, act on it and correct its consequences. A one-time improvement can create breathing room, but sustained growth may use it up. The following scenario compares pressure and response on one assumed scale; it does not claim a measured global absorption limit.
Common workload index: manual capacity = 100. All future lines are assumptions, not observed global cyber threats or a forecast of attacks.
This example uses a Microsoft October 2025 randomised study of 167 professional analysts. Its workflow calculations imply roughly 3.1–6.5× correctly identified malicious emails per analyst-minute under particular prevalence, precision and recall assumptions. The lower 3.1× case assumes 20% malicious mail and 80% precision and recall, with automated benign-mail handling.
The 20% annual workload increase is illustrative. Staffing, task mix and effectiveness are held constant; implementation overhead, errors and automation bias are not modelled. A task-specific vendor study does not establish a 3.1× reduction in breaches or a universal productivity gain.
Microsoft research · study and assumptions ↗Measure detection delay, error rates and the human ability to check an answer. Faster output alone does not establish better judgment.
Measure the time from a warning to action, trained staff and available infrastructure. Capacity is specific to the task.
Measure appeal times, fallback readiness and successful recovery. A response system needs a practical way to change course.
Control can enable useful speed. Explore how prevention, greater response capacity and shorter delays change the gap in Pace and Control Simulator ↓.
05 / PACE AND CONTROL SIMULATOR
Explore the conceptual relationship between pressure, adaptive capacity and exposure. These controls express assumptions; they do not measure the world's current state.
Dimensionless conceptual ratios, not calibrated indicators, probabilities or safety thresholds. The time paths are an illustrative extension: pressure = 100 × A × E × (1 + p)t; capacity = 100 × C × (1 + c)max(0, t − delay). Equal values only mean equality on this assumed common scale. No empirical global “absorption capacity” is claimed.
Electronic waste generation and documented formal collection and recycling, in million tonnes per year.
Formal recovery is a measurable response, not total societal absorption. The remainder includes informal recycling, storage and other destinations; it is not all landfill or proven harm.
ITU / UNITAR · Global E-waste Monitor 2024 ↗06 / GOVERNANCE FOR CONTROL
A model can reveal a gap. Institutions need the authority, knowledge and accountability to act on it. These possible structures are proposals for discussion, not a claim that one design fits every society.
AI can help detect risks, compare options and evaluate results. Decisions affecting rights, public resources and acceptable risks need accountable human authority. No single operator should both make consequential decisions and be the sole judge of their effects.
These structures can be combined. Each needs a visible route from evidence to action, and from a contested decision to a remedy.
Appeals need published response deadlines and protection against retaliation. Emergency intervention must be proportionate: stopping a critical service can itself cause harm. Review findings should feed back into the next decision.
Track time from warning to action, access to appeals, correction and recovery times, recurrence of failures, distribution of impacts and concentration of decision power. Assess these together: a fast decision is not necessarily a good one, and a low appeal count may reflect barriers to challenge.
Return to Pace and Control Simulator ↑