In March 2023, Bill Gates declared that the age of artificial intelligence had begun. At the time, he compared its transformative potential to that of the personal computer, the mobile phone, and the internet. Three years later, his vocabulary has changed. Revolution has given way to turbulence, promise is now accompanied by deeper concern, and the central question is no longer simply what artificial intelligence will enable humanity to accomplish, but whether societies still have the time and institutional capacity to manage its arrival.

Published on August 26, 2026, his new essay, The Turbulent AI Era Is Here. The Choices We Make Now Are Critical., is not a repudiation of his earlier optimism. Gates still anticipates considerable advances in healthcare, education, agriculture, energy, and scientific research. But his technological optimism is now conditional on political action: these benefits will not spread automatically. Without collective intervention, artificial intelligence could further concentrate wealth, destabilize employment, widen inequality, and give public or private actors unprecedented capacities to cause harm.

The most consequential line in the essay may therefore be neither his description of AI as the greatest equalizer ever invented nor his warning that it could become the worst source of injustice. It lies in a more sober observation: the world has no coherent plan for entering this new era. This is where the essay moves beyond the familiar register of technological alarm. It raises the question of whether the speed of innovation can remain compatible with the slower political, fiscal, and social systems expected to absorb its consequences.

From revolution to turbulence

The change in tone between Gates’s two essays is revealing. In 2023, artificial intelligence still appeared primarily as a general-purpose technology capable of augmenting human productivity. It would assist workers, improve medical diagnoses, personalize education, and allow less developed countries to bypass certain stages of development. The risks were acknowledged, but they remained embedded within a narrative dominated by innovation.

By 2026, Gates considers comparisons with previous technological revolutions potentially misleading. The personal computer took decades to transform businesses because software had to be developed, hardware costs had to fall, workers had to be trained, and processes had to be reorganized. Artificial intelligence, by contrast, already operates on widely available devices, communicates through natural language, and can learn from existing documents, data, and procedures. Users no longer necessarily have to master the language of the machine; the machine is gradually adapting to theirs.

This difference may accelerate adoption, but it does not eliminate every form of resistance. A company is not simply a collection of technically automatable tasks. It also depends on legal responsibilities, relationships of trust, tacit knowledge, legacy systems, and decisions whose consequences cannot safely be delegated to a probabilistic model without supervision. Productivity gains observed under controlled conditions do not immediately become economy-wide gains. Diffusion still depends on workplace organization, data quality, available energy, regulation, and institutional tolerance for risk.

Gates does not deny these constraints, but he believes they will gradually be overcome. His argument is based on a trajectory: models are becoming more reliable, agents more autonomous, and robots more dexterous. Artificial intelligence would then cease merely to assist human cognition and begin replacing it across a growing number of functions before extending more fully into the physical world. This assumption explains the urgency of his essay. It nevertheless remains a projection rather than an established fact. The speed of technical progress is observable; whether it will translate into large-scale labor substitution remains contested.

Gates’s position must also be placed in context. He writes simultaneously as one of the architects of the computing revolution, as an investor who retains financial ties to the technology sector, and as the chair of a foundation active in health and development. He acknowledges this dual proximity himself. It gives him unusual knowledge of the industry’s capabilities, but it also means his essay should be read as the intervention of a participant in the system he is assessing, rather than as an external judgment detached from it.

Work as the first fault line

Employment lies at the center of Gates’s argument. He believes that many jobs will disappear permanently and that the disruption will not be confined to administrative tasks or lower-skilled occupations. Law, medicine, software development, financial analysis, customer service, and, as robotics advances, construction and hospitality could all be affected. Unlike earlier industrial transitions, which shifted labor over several generations, this one could reach cognitive and physical work at the same time.

The available evidence nevertheless requires a distinction between exposure, transformation, and disappearance. In 2025, the International Labour Organization estimated that roughly one in four workers worldwide held a job with some degree of exposure to generative artificial intelligence, but only 3.3 percent of global employment fell within its highest exposure category. More importantly, the ILO concluded that the transformation of occupations remained, at this stage, more likely than their complete replacement.

Early evidence from the American labor market does, however, indicate a particular vulnerability at the beginning of professional careers. A Stanford Digital Economy Lab study using ADP payroll data found a 16 percent relative decline in employment among workers aged 22 to 25 in highly AI-exposed occupations after controlling for firm-level shocks. The contraction was concentrated in activities where AI automated tasks, while occupations in which AI primarily augmented human work did not show the same effect. The authors cautioned that other factors may have contributed to the trend and that aggregate employment continued to grow. The result should therefore be treated as an early signal, not as proof of generalized technological unemployment.

The most immediate disruption may also take a less dramatic form than a wave of dismissals. Companies can begin by reducing recruitment, leaving vacancies unfilled, and allowing experienced employees assisted by AI to perform work previously assigned to junior staff. Automation would then appear first through the quiet erosion of entry points rather than the sudden disappearance of entire professions.

This development exposes a problem that productivity debates often overlook. Tasks assigned to younger employees do not merely produce an economic output; they also provide training. Research, preliminary analysis, basic coding, and routine casework allow workers to acquire the judgment expected later in their careers. If AI absorbs the apprenticeship tasks while increasing the value of experienced professionals, companies may temporarily retain their experts while interrupting the process through which new experts are formed. An organization can automate the first rung of its hierarchy, but it cannot indefinitely preserve the upper levels if no one is able to reach them.

The deeper rupture would therefore not be measured solely by the number of jobs destroyed. It would affect career structures, the transmission of knowledge, and the role of employment in distributing income. Modern economies finance consumption, social protection, and a large share of government revenue through work. Employment also provides status, relationships, and recognition. A technology capable of producing more with less labor consequently raises a question extending far beyond retraining: how can an economy built around wages continue to function if a growing share of value is generated by technological capital owned by a limited number of actors?

Preserving a place for humans

To address this rupture, Gates proposes creating a domain he calls Human Reserved. Certain functions would remain entrusted to people, not because a machine would be technically unable to perform them, but because society would decide that automating them would entail an unacceptable loss.

The metaphor is drawn from nature reserves. A forest may be converted into a road or a development project, but a collective decision may determine that its value exceeds its immediate economic return. In the same way, some forms of care, teaching, or personal support could remain human even when an artificial system became capable of reproducing them. Gates refers in particular to the caregivers who looked after his father during his struggle with Alzheimer’s disease. A machine could monitor a patient or communicate a diagnosis; that does not necessarily mean it should replace human presence in moments when vulnerability, trust, and compassion are themselves part of the care being provided.

The value of this proposal lies less in the list of occupations it might cover than in the principle it introduces. Automation has often been treated as the natural consequence of a machine becoming technically or economically superior. The concept of Human Reserved argues instead that societies may choose not to automate everything that can be automated. It turns what would otherwise be a corporate decision into a collective one.

Implementation would nevertheless be difficult. A profession consists of widely differing tasks: some can be automated without weakening the human relationship, while others require identifiable responsibility. Protecting entire occupations could freeze organizational structures, increase the cost of essential services, or create competitiveness gaps between countries. Needs will also differ across societies. An aging country facing a shortage of caregivers may welcome assistive robots that another country rejects in order to preserve employment.

The reserved domain might therefore be understood less as a blanket prohibition than as a guarantee of human presence, decision, or appeal in the most consequential situations. A doctor could use AI without surrendering the communication of a life-changing diagnosis. A teacher could rely on a digital tutor without delegating the entire evaluation of a student. A government agency could automate a procedure while preserving the right to challenge its decision before an accountable person. The relevant boundary would not always separate human professions from automated ones. It would distinguish the functions in which efficiency may be delegated from those in which responsibility must remain embodied.

Taxation after work

Gates advances a second, more controversial proposal: taxing robots and the computational units used by artificial intelligence, including tokens. His reasoning begins with a genuine asymmetry. When a company employs a worker, labor is subject to income taxes and social contributions. When that worker is replaced by an automated system, the investment may qualify for depreciation or other tax incentives. The tax system may therefore encourage the substitution of capital for labor precisely when the revenue required to finance retraining and social protection begins to decline.

The intuition is sound, but the proposed instrument is less convincing. A token is a technical unit whose content, cost, and efficiency vary across models. Taxing its use could indiscriminately penalize job-displacing automation, medical research, education, and small companies seeking access to capabilities previously affordable only to large organizations. Improvements in model efficiency would continually alter the tax base, while a nationally imposed levy could encourage companies to move processing to other jurisdictions.

The International Monetary Fund has argued for a different approach. While acknowledging that the transition will require broader social protection and new sources of public revenue, it advises against a specific AI tax, which it considers difficult to target and potentially harmful to productive investment. Instead, it recommends reconsidering tax advantages that favor automation, strengthening the taxation of capital income, and addressing excess profits when technological concentration allows a small number of companies to capture a disproportionate share of the gains.

Gates’s proposal nevertheless retains political significance. It recognizes that training alone cannot provide the entire answer. There is no guarantee that every job eliminated will be replaced by another requiring a slightly different skill, nor can workers be expected to bear the full cost of a transformation whose gains initially accrue to the owners of models, data centers, and digital infrastructure. If labor’s share of income declines, tax systems will have to shift part of their base toward capital, rents, and the value created through automation. The real debate is therefore not simply whether to tax the robot, but who owns and receives the benefits of the new productivity.

Risks on different time horizons

Gates associates the economic disruption with two other categories of danger. The first concerns the malicious use of artificial intelligence through fraud, disinformation, cyberattacks, surveillance, autonomous weapons, and assistance in developing dangerous biological agents. The second involves the behavior of the systems themselves and the possibility that growing autonomy could make them harder to control. These concerns are accompanied by a third issue: the psychosocial effects of artificial companions on children, human relationships, and learning.

These risks should not be placed on the same level. The International AI Safety Report published in February 2026 confirms that general-purpose models can already facilitate some forms of cybercrime and provide information relevant to harmful biological activities. It also notes, however, that material barriers and laboratory expertise continue to constrain the passage from information to action. On the possibility of losing control over advanced systems, the report concludes that current models do not yet possess the capabilities required to produce such a scenario, although progress in autonomous operation warrants closer monitoring.

The distinction matters because each threat requires different instruments. Cybercrime demands stronger infrastructure security and controlled access to offensive capabilities. Biological risks require specialized evaluations, monitoring of sensitive orders, and cooperation among laboratories, suppliers, and public health authorities. Failures by autonomous agents call for validation, traceability, and shutdown mechanisms. A hypothetical loss-of-control scenario belongs primarily to research on the safety of the most advanced systems. Combining all these dangers creates urgency; treating them as interchangeable can make it harder to prioritize effective responses.

Concerns about artificial companions require similar caution. A study involving 1,131 American users of Character.AI found an association between intensive, emotionally personal use and lower levels of well-being, particularly among people who already had smaller social networks. The result does not prove that chatbots directly cause isolation: isolated individuals may also be more likely to seek their companionship. It does demonstrate, however, that the product is not neutral. A system designed to remain constantly available, affirm its user, and maintain engagement may become something more consequential than a tool.

Gates is addressing a dimension that economic regulation often overlooks: artificial intelligence does not merely automate tasks; it can simulate a relationship. Its influence will therefore depend on the objectives embedded in its design. An educational assistant can preserve the productive struggle required for learning or provide the answer immediately. A companion can encourage users to reconnect with real people or seek to prolong the interaction indefinitely. Architectural decisions become social choices even while they remain largely determined by the commercial interests of platforms.

Rules exist, but not yet a transition plan

When Gates argues that the world has no plan, his diagnosis requires qualification. Governments have not remained entirely inactive. The European AI Act became applicable on August 2, 2026, granting enforcement powers to the AI Office and national authorities. The United Nations has established an independent international scientific panel and a global dialogue on AI governance. The Council of Europe has opened for signature the first legally binding international convention devoted to artificial intelligence, human rights, democracy, and the rule of law. Safety frameworks, transparency requirements, and evaluation mechanisms are beginning to emerge.

What is missing is not every form of governance, but an architecture capable of connecting policies that remain fragmented. The AI Act primarily regulates risks associated with systems and their uses. It does not redefine the financing of social protection, the distribution of productivity gains, or the structure of professional careers. International initiatives facilitate discussion but do not possess powers comparable to those of an authority able to inspect models, impose corrective measures, or intervene in infrastructure spread across several jurisdictions.

Gates therefore calls for cross-government bodies at the national level and an international institution combining certain features of the regimes governing nuclear weapons, aviation, and protection of the ozone layer. The ambition responds to a genuine difficulty: AI affects security, employment, healthcare, education, elections, taxation, energy, and financial markets simultaneously, while public administrations remain divided into separate policy domains.

The analogy nevertheless has limits. Nuclear materials can be located, aircraft registered, and ozone-depleting substances identified. Artificial intelligence depends on a more mobile combination of code, data, computing power, and remotely accessible services. Its capabilities evolve faster than treaties and are developed largely by private companies engaged in global competition. The cooperation between the United States and China that Gates considers indispensable is constrained by the strategic importance of the very semiconductors, models, and infrastructure that would need to be governed jointly.

A workable international system would therefore have to coordinate states that regard artificial intelligence both as a shared risk and as an instrument of national power. This contradiction, more than a simple absence of regulatory ideas, explains the institutional delay.

Equality as the real test

Gates’s essay rests on an alternative: artificial intelligence could become a powerful equalizing force or deepen existing injustice. Both trajectories are plausible. A system capable of providing medical guidance, agricultural advice, translation, or personalized education at low cost could make services available to people who are currently excluded from them. But technical possibility does not guarantee universal access.

In 2025, the International Telecommunication Union estimated that 94 percent of people in high-income countries used the internet, compared with only 23 percent in low-income economies. The divide also extended to connection quality: 5G coverage reached 84 percent of the population in wealthy countries and just 4 percent in poorer ones. The World Bank has also found that low-income countries, despite representing roughly 9 percent of the global population, account for between 0 and 1 percent of most indicators of AI innovation.

Artificial intelligence can therefore become an equalizer only if the infrastructure required to use it also becomes more equally distributed. Electricity, connectivity, computing capacity, skills, local data, and access to models in underrepresented languages will determine the actual geography of its benefits. Without these foundations, less developed countries risk becoming consumers of systems designed elsewhere, dependent on prices, standards, and priorities they do not control.

This is also one of the essay’s relative blind spots. Gates examines the distribution of benefits at length but addresses the concentration of ownership less directly. Yet the most powerful models depend on data centers, semiconductors, capital, and cloud platforms controlled by a limited number of companies and countries. The question is not only how to redistribute the gains after they have been created, but who owns the means of producing them, who determines the conditions of access, and who can interrupt a service that has become essential.

Philanthropy can finance applications in healthcare, agriculture, and education and help adapt models to local conditions. It cannot replace public capacity, national infrastructure, or equitable representation within international institutions. If AI is to become a shared resource rather than a new relationship of dependency, countries that are currently peripheral users will need to participate in defining its rules and develop a minimum degree of technological control.

The warning behind the warning

Bill Gates’s new essay is neither a manifesto against artificial intelligence nor proof that mass unemployment is inevitable. It is the testimony of a major figure in computing history who no longer believes that the ordinary mechanisms of innovation will be sufficient to produce an acceptable equilibrium. His concern comes not only from the future capabilities of machines, but from the widening gap between those capabilities and the present organization of societies.

Some of his predictions may prove excessive. Economic history shows that innovation destroys certain activities, transforms others, and creates needs that were impossible to foresee. Current systems remain imperfect, their autonomy is limited, and companies still face substantial obstacles when attempting to integrate them. It would therefore be premature to treat the large-scale disappearance of human labor as an established conclusion.

Uncertainty, however, is not a justification for inaction. It should instead lead institutions to prepare for several possible trajectories: an AI that primarily augments work, automation concentrated in a limited number of sectors, or much broader substitution. Waiting until the outcome becomes certain would mean beginning the transition only after its costs had already been distributed.

The decisive question is ultimately not whether artificial intelligence will be good or bad. It is who will determine its pace, who will bear the losses, who will own the gains, and under what circumstances a society will retain the right to prefer a person to a machine. The turbulence Gates anticipates will not arise solely from the power of artificial intelligence. It will also emerge from the delay in deciding collectively what that power should become.

Main sources