Bill Gates envisages two decades of upheaval before an economy of abundance emerges. That prospect opens a debate extending far beyond the capabilities of machines. If artificial intelligence makes it possible to produce more with less human labour, who will own the means of production? How will income be distributed? And which institutions can turn greater technological power into shared progress?

Atlas Limits — Reports & Analysis | October 25, 2026

A society can become more productive without becoming fairer. It can manufacture more, automate faster and accumulate extraordinary technical capabilities while leaving part of its population economically insecure. That possibility, too often overshadowed by new model announcements and promises of growth, should be central to the debate over artificial intelligence.

In an interview with The Verge’s Decoder podcast, published on October 22, Bill Gates envisaged a deeply disruptive transition lasting roughly twenty years towards a possible economy of abundance. He discussed shorter working hours and activities reserved for humans. This is a personal scenario, not a scientifically established timetable.

The significance of this warning lies less in the number of years suggested than in the problem it raises. Even if these technologies fulfil a substantial share of their promise, their spread could weaken the mechanisms through which societies currently distribute income, organise careers and finance social protection.

Public debate tends to focus on a dramatic question: how far can machines go? Another deserves equal attention: how will we live together if they perform a growing share of the activities around which our economic systems are organised?

The first concerns research and engineering. The second involves ownership, power, institutions and democracy. AI

A transformation to take seriously, without treating it as inevitable

Any discussion of the future of artificial intelligence should begin with a simple discipline: distinguish what has been observed, what can reasonably be anticipated and what remains speculative.

Systems can already improve the performance of clearly defined tasks. That does not demonstrate that they will be able to replace entire professions within the same timeframes and at the same cost. Success in a controlled environment does not guarantee equivalent reliability when facing unfamiliar situations, contradictory information or consequences that are difficult to reverse.

Between a technical capability and an economic transformation stand investments, procedures, responsibilities, organisational constraints and human decisions.

A company must be able to integrate a tool into its systems, protect its information, verify its outputs and determine who is accountable for its errors. It must also establish that the benefits exceed the costs of deployment, supervision and correction.

Robotics adds another layer of difficulty. Operating in the physical world requires dealing with changing environments, wear, maintenance and human safety. Automating a digital operation and acting autonomously on a construction site, inside a home or in a care facility involve different requirements.

These limitations do not make the transformation insignificant. They do prevent us from describing it as a uniform process whose outcome is already known.

Uncertainty should lead us to prepare for several possible paths, rather than turn one of them into a collective destiny.

Exposure to AI does not mean job disappearance

Employment figures are among the most frequently cited—and most frequently misinterpreted—elements of the debate.

In an analysis published in 2024, the International Monetary Fund estimated that almost 40% of global employment was exposed to artificial intelligence, rising to around 60% in advanced economies. However, it distinguished between situations in which AI could complement human work and those in which it could reduce demand for labour. These percentages were therefore not forecasts of job losses.

A study published in May 2025 by the International Labour Organization and Poland’s NASK research institute, focusing on generative AI, estimated that one in four workers worldwide held an occupation with some degree of exposure. It concluded that job transformation was, overall, more likely than complete replacement.

These assessments use different methodologies and cover different technological scopes. They should neither be added together nor compared as though they measured exactly the same thing. Their shared lesson is nevertheless essential: technical exposure describes the possibility of change, not its social outcome.

An occupation consists of several tasks. Some are repetitive; others depend on judgement, negotiation, contextual understanding or trust. Automating one component may allow a person to perform the others better. It may also lead an organisation to redistribute responsibilities among fewer employees.

Both outcomes are possible.

The ultimate effect will depend, among other factors, on demand. If lower costs make it possible to serve substantially more customers, activity may expand enough to support employment. If the market remains limited, the same improvement may primarily reduce staffing requirements.

Two opposing certainties should therefore be avoided: that work will disappear across the board, and that every lost job will automatically be replaced.

Productivity does not tell us who benefits

Productivity gains are more than an abstract promise.

A study by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, published in 2025 in the Quarterly Journal of Economics, examines the introduction of an AI assistant among 5,172 customer-support agents. It finds an average increase of 15% in the number of issues resolved per hour, with particularly substantial benefits for less experienced workers. This finding concerns a specific system and activity; it does not measure AI’s effect on the entire economy.

Nevertheless, it provides a basis for asking the central question of distribution.

When a team accomplishes more in the same amount of time, several choices become possible. The company can improve service, raise wages, shorten working hours, expand its business, increase its profit margin or reduce its workforce.

The software does not choose among these options.

The outcome depends on competition, corporate strategy, collective agreements, workers’ bargaining power and public policy. The same technology can therefore produce different social effects depending on the organisation into which it is introduced.

This distinction is fundamental. Productivity describes the relationship between the resources used and the output obtained. It contains no automatic rule for sharing the benefits.

An economy can deliver stronger results while some workers suffer declining incomes or deteriorating prospects. Conversely, well-integrated technology can simultaneously improve service quality, working conditions and a company’s viability.

The decisive issue becomes how technological progress is organised: for what purposes, by whom and under which conditions?

The risk of a quiet erosion of work

The upheaval could be less visible than a succession of mass layoffs.

A company can reduce recruitment without immediately eliminating positions. It can leave departing employees unreplaced, use fewer contractors, shorten assignments or ask an unchanged team to absorb a growing workload.

In such a scenario, aggregate employment statistics could conceal profound changes in the conditions governing entry into the labour market and career progression.

The early stages of professional life are particularly important.

Many professions depend on gradual learning: preparing documents, conducting research, checking information, handling straightforward cases and eventually taking on greater responsibilities. These tasks have productive value, but they also build experience.

If some of them are automated, an organisation may gain immediate efficiency while weakening its own training pipeline.

This difficulty is not insurmountable. An assistant can also accelerate learning, make reasoning more explicit and improve access to expertise. But that requires designing its use around those objectives, rather than treating every entry-level task as a cost to eliminate.

A society cannot indefinitely demand experience from young workers while removing the opportunities through which they acquire it.

The quality of the transition will therefore have to be assessed through recruitment, career paths, pay and autonomy at work, as well as total employment.

Ownership at the centre of the automated economy

If a growing share of production depends on automated systems, ownership of those systems becomes a major determinant of income distribution.

The issue extends beyond the models themselves. It encompasses computing infrastructure, specialised software, distribution networks, data that can lawfully be used, industrial equipment and commercial relationships.

A company that automates an activity may create considerable value. But the share of that value reaching employees, suppliers, consumers and capital owners depends on market structure and the rules governing it.

One possible scenario involves AI spreading productive capabilities widely. A small business can access services previously available only to large organisations. An independent professional can handle more cases. An institution with limited resources can improve certain services.

Another involves greater dependence: those same actors become reliant on a small number of suppliers able to change their prices, access conditions or features.

These two dynamics can coexist. Access to tools can broaden while infrastructure ownership and revenue capture remain concentrated.

Democratising access to a technology does not necessarily democratise control over it.

That is why competition, the ability to switch suppliers, data portability and system interoperability are major economic questions. They help determine whether AI increases users’ autonomy or establishes new dependencies.

Abundance does not eliminate every form of scarcity

The promise of abundance itself deserves scrutiny.

A substantial reduction in the cost of certain intellectual operations can transform access to services. It does not mean that all the goods required for a decent life will become abundant or inexpensive.

Housing illustrates the point. Reducing design or construction costs does not automatically create well-located land, transport links, water and electricity networks, or homes affordable to lower-income households. Part of the gain may be absorbed by land values or other costs.

In healthcare, improving the analysis of a patient’s records is not enough to ensure comprehensive treatment. Care also depends on equipment, medicines, staff, capacity and financing arrangements.

In agriculture, better forecasts do not replace water, soil, logistics infrastructure or the stability necessary for production.

Even a highly automated economy would retain scarce resources: space, certain materials, natural environments, human time, attention and trust.

Scarcity can also shift. When content becomes exceptionally easy to produce, verification and the ability to identify reliable sources become more valuable. As digital services multiply, the physical infrastructure enabling them can become a focus of economic concentration.

The issue is therefore not whether AI will abolish all scarcity. It is which constraints it will reduce, which will remain and who will control the resources that become strategically important.

Artificial intelligence has a physical geography

AI depends on physical infrastructure. Its capabilities are tied to semiconductors, networks, electricity, cooling and the equipment that keeps data centres running.

In its updated 2026 projections, the International Energy Agency estimates that global electricity consumption by data centres could rise from around 485 terawatt-hours in 2025 to 950 terawatt-hours in 2030, equivalent to approximately 3% of global electricity demand. These figures cover all data centres, not AI alone.

This physical dimension creates tension between different timescales. Software can evolve rapidly; extending an electricity grid or building industrial infrastructure generally takes longer.

It also requires scrutiny of how costs are distributed. Who finances grid connections? How are competing demands for electricity and water managed when local constraints emerge? What economic benefits compensate for the resources committed?

The assessment must remain balanced. AI can help optimise energy systems or reduce certain forms of consumption. But greater efficiency per operation does not guarantee lower overall consumption if usage increases sharply.

Digital abundance therefore rests on choices about planning, investment and access to resources. It does not exist independently of the real economy.

A new international divide is possible

AI’s effects will also be geographical.

For an advanced economy, the challenge may be to transform high-value services while preserving incomes and career mobility. For a country whose development depends on exporting administrative or digital services, the risk may concern the durability of a competitive advantage.

If some operations become less dependent on labour costs, location decisions may change. This does not mean that all outsourced activities will disappear. Language skills, customer relationships, market knowledge and the ability to manage complex situations may remain decisive. But development strategies will have to account for the possibility.

Conversely, less expensive tools could help businesses and public administrations overcome skills shortages. The potential gains are substantial where access to certain services remains limited.

The IMF nevertheless stresses that lower exposure to disruption does not guarantee a stronger position: less well-equipped economies may also struggle to benefit from AI because of infrastructure and skills constraints.

For African and Mediterranean countries, the strategic challenge is therefore to connect technological adoption with local value creation.

This can involve skills, applications adapted to business needs, better public services and control over sensitive data. It also means evaluating infrastructure projects according to their actual effects on the wider economy.

Hosting computing capacity is not, by itself, enough to build an innovation economy.

Purchasing power remains essential to a productive economy

An economy that produces more must still enable people to access what it produces.

If automation lowers costs, consumers may benefit from lower prices. This is an important mechanism: it can release purchasing power and support new activities.

But its effects depend on several conditions. Savings must be passed through to prices. Competition must function. And household incomes must not deteriorate faster than the cost of essential spending.

Someone whose income falls does not necessarily benefit meaningfully from cheaper digital services if rent, food or transport remain difficult to afford.

It is therefore possible to imagine an economy in which some products become abundant while material insecurity persists. This is not a forecast; it illustrates why productive capacity cannot be examined independently of income distribution.

Wages remain one of the principal means of accessing consumption. If their role were to diminish across a substantial part of the economy, other mechanisms would have to become more important: more widely distributed capital income, transfers, public services or new forms of participation in productivity gains.

This debate cannot be postponed until the hypothetical arrival of mass unemployment. Institutions governing distribution take longer to build than the tools capable of destabilising them.

Social protection faces a changing world of work

Depending on the country, social protection systems combine contributions on labour income, taxes and other public resources.

A change in the distribution between wages and profits could therefore alter their financing. It would not automatically cause them to collapse: a more productive economy can also generate greater revenues. Much depends on where value emerges and governments’ ability to collect a share of it.

Taxing robots is an intuitive proposal, but defining the tax raises difficulties. Software that assists an employee and a system that replaces a position may use similar technologies. Substitution can be diffuse and difficult to isolate.

Taxation aimed too directly at a particular piece of equipment could also discourage useful investment, including investment that improves safety or enables a small company to remain competitive.

Another approach is to examine the taxation of labour, capital, profits and economic rents more broadly, without making the entire system depend on a particular definition of AI.

The principle under discussion is straightforward: if the economic base supporting collective financing shifts, public revenues must be able to evolve with it.

This debate must account for tax competition and the international mobility of economic activity. A national reform can lose some of its effectiveness if the income it targets can easily be moved to other jurisdictions.

Basic income, public services and shorter working hours

No single response addresses the whole problem.

Depending on its design, a universal basic income could provide a foundation of security independent of employment status. But its level, financing and relationship with existing benefits would determine its effects.

An income set too low would not provide meaningful security. A scheme replacing assistance tailored to specific needs could leave some recipients worse off. And cash transfers alone cannot solve a shortage of housing or healthcare services.

Public services therefore represent another dimension of sharing the gains: reducing unavoidable household expenses can improve economic security as much as increasing disposable income.

Shorter working hours offer another possibility. If an activity produces more per hour, some of the gain can be converted into free time. But this requires actual productivity improvements, suitable organisation and agreement on how costs will be shared.

In some sectors, needs could instead increase. A technology that makes it possible to meet previously unmet demand may justify more professionals and greater resources.

The objective should therefore not be to impose the same formula everywhere. It should be to enable several forms of progress: better pay, shorter hours, more accessible services, improved working conditions and new activities.

Training is essential, but cannot carry the transition alone

Retraining is often presented as the central response to automation. It is necessary, but its effectiveness depends on the jobs actually available.

Training a person does not automatically create an employment opportunity. Asking workers to adapt continuously can even transfer most of the risk of a transformation beyond their control onto them.

A credible policy must bring together several elements: time to learn, income during the transition, high-quality training and identifiable prospects. It must also account for family responsibilities, age, mobility and regional disparities.

Initial education raises an even deeper question. If tools evolve rapidly, preparing people only for their immediate use may produce skills that quickly become outdated.

Durable capabilities become essential: understanding a problem, verifying information, reasoning with data, explaining a decision, cooperating and recognising the limits of a result.

Knowing how to use AI also means knowing when its use is inappropriate and how to check what it produces.

Education must preserve this intellectual autonomy. Without it, access to powerful tools could coexist with growing dependence on their answers.

Work provides more than income

Work organises part of social life. It creates relationships, structures time, develops skills and can provide a sense of making a useful contribution.

It would nevertheless be misleading to idealise all forms of employment. Some jobs are exhausting, dangerous, poorly paid or offer little autonomy. Transforming them can represent substantial progress.

The challenge is to preserve the positive dimensions of activity without treating economic necessity as their only possible source.

A society in which paid work occupied less space would need to give greater recognition to other contributions: caring for relatives, participating in local life, passing on knowledge, creating, learning or taking part in collective projects.

But this recognition cannot be merely symbolic. People need time, resources and institutions to pursue these activities under decent conditions.

A reduced role for market-based work could expand human freedom. It could also create a divide between those with the resources to choose their activities and those who remain dependent on decisions taken elsewhere.

Progress would then have to be measured through people’s actual autonomy, not simply the number of hours freed from employment.

Governing uses and their consequences

AI also raises a question of institutional power.

When a system participates in recruitment, access to a service, the assessment of an individual or the allocation of a resource, technical performance does not settle the issue. It is necessary to determine which decisions may be automated, who can challenge them and who bears responsibility.

A decision that is correct in most cases can remain unacceptable if its errors are severe, systematically concentrated on particular groups or impossible to correct.

Governance must therefore address the conditions of use as well as the characteristics of models.

It must also avoid two pitfalls. Insufficient requirements can shift risks onto users and citizens. Excessively expensive or poorly designed obligations can favour the most powerful companies and reduce competition.

The objective is to establish safeguards proportionate to the consequences of each use: documentation, evaluation, effective human oversight and avenues of appeal when fundamental rights or essential interests are at stake.

The technology’s international character complicates this task. Infrastructure, suppliers and users may fall under different jurisdictions. But the absence of a comprehensive global agreement does not leave governments powerless over their public services, markets and the uses deployed within their territories.

Three paths, rather than a single future

These tensions suggest three possible paths, useful as analytical frameworks without assigning any of them a certain probability.

The first would involve widely shared gains. AI improves services, strengthens workers’ capabilities, supports new activities and enables a better distribution of time. Competition and institutions limit the concentration of benefits.

The second would involve concentrated growth alongside greater social insecurity. Productive capacity increases, but income is distributed unevenly. Some occupations lose economic value, careers become more precarious and part of the population struggles to access essential goods.

The third would involve a slower and more expensive transformation than anticipated. Important advances occur in some activities, but constraints involving reliability, integration and infrastructure limit their spread. Some investments disappoint, without preventing lasting disruption in particular sectors.

These paths can overlap across countries, professions and periods. A technology can simultaneously improve services, displace jobs and fail to meet certain financial expectations.

That is why assessment must track several dimensions: actual productivity, service quality, wages, access to a first job, working time, market concentration, resource consumption and institutions’ ability to correct imbalances.

Investment totals and model performance alone cannot describe a social transformation.

Prosperity will not be automatic

Bill Gates’s warning offers a starting point. It neither demonstrates the end of work nor guarantees future abundance.

The more serious question concerns a society possessing new technical capabilities but institutions insufficiently prepared to distribute their benefits and limit their harms.

Decisions taken today can influence this transition: preserving professional learning, encouraging competition, strengthening public-sector capabilities, supporting workers through career changes and enabling them to participate in the transformations affecting them.

It is also necessary to recognise that progress has several dimensions. An economy may seek to produce more, but also to provide better care, reduce arduous work, protect resources and give people greater control over their time.

These objectives do not always coincide. They must be debated and balanced.

The promise of future prosperity cannot, on its own, justify present losses without compensation, just as fear of disruption should not lead societies to abandon possible improvements.

Artificial intelligence can expand our capacity to produce. Prosperity will depend on our ability to share the power that comes with it.

Much of the coming decades will be decided in that gap between technological capability and collective organisation.