Bill Gates: Tax tokens and robots. The turbulent AI era will be shaped by three institutional choices over the next decade.

Gates shifts the debate from "how powerful is AI" to "is society ready"

The PC took two decades to transform work. Software needed to be developed, hardware had to get cheaper, and people had to learn new ways of operating. Generative AI spreads under very different conditions: it runs on existing computers and phones, accepts instructions in natural language, and can learn from existing documents, data, and training videos. Companies don't need to rebuild their entire infrastructure before slotting it into customer service, programming, legal, medical, and manufacturing workflows.

There's real uncertainty about whether capabilities will advance as quickly as Gates predicts. He plots vision, speech, reasoning, and robotics along a steep upward curve and expects the shock to hit across industries within a decade. That's a forecast. What's already true: the barrier to adoption is falling, cost competition between companies will drive uptake, and there's no credible global mechanism to synchronize a slowdown.

The problem becomes a speed race. On one side: models updating monthly, products spreading quarterly, and cost pressure following close behind. On the other: public institutions that need legislation, budgets, inter-agency coordination, and international negotiation.

As long as that time gap exists, risks won't fall evenly. People with lower incomes, less access to retraining, and less political influence tend to get the benefits later and absorb the adjustment costs earlier. That ordering is what actually worries Gates.

Three categories of risk hit the same people: those without time and power

The first risk is to jobs. Entry-level and mid-level positions are easier to standardize, and young people—who lack experience and networks—feel the hiring contraction first. A Stanford Digital Economy Lab study, in its 2026/8 (August) update, found that the employment gap for young workers in AI-exposed occupations widened to 19%. But the researchers are explicit: they observed no broad economy-wide replacement, and these results are early descriptive indicators, not causal estimates of AI-driven unemployment.

The ILO's global task index offers another lens: roughly a quarter of the world's workers are in occupations with some degree of generative AI exposure, with the highest-exposure category covering 3.3% of global employment. Since most jobs still include tasks requiring human judgment, communication, or physical operation, job transformation is more likely than whole-role elimination. Stanford's research supports the alarm that the youth entryway is narrowing; the ILO study cautions against turning that alarm into a foregone conclusion of total job loss.

Related readingPrinceton professor on AI and work at ICML: how individuals can adaptModel capability has leapt forward in 24 months, but a composite reliability metric built by the SAGE lab has moved only 5–10 points.→

The second risk is misuse of capability. Fraud, deepfakes, cyberattacks, biological threats, and opinion manipulation all scale up as information and execution costs drop. But this bucket actually mixes three distinct problems: ordinary bad actors get more powerful tools, states and large institutions gain more surveillance and coercive capability, and highly capable systems behave in ways their designers didn't intend. These three have different likelihoods, different governance tools, and different timescales—they shouldn't be flattened into one generic "AI safety risk."

The third risk lands on child development, learning, and human relationships. A study of 1,131 CharacterAI users found that people with smaller real-world social networks are more likely to treat chatbots as companions; companionship use correlates with lower well-being, and the association strengthens with more intense use and more self-disclosure. This remains observational—the causal direction could run the other way, with lonelier people turning to AI more heavily.

Research on critical thinking also supports only cautious claims. A survey and interview study of 666 people reports that frequent AI use, outsourcing thinking to tools, and lower critical thinking scores are correlated, with younger participants showing higher dependence. It can't prove AI causes cognitive decline. Age, education, self-selection, and usage patterns could all be shaping the results simultaneously.

Looking across these four bodies of evidence, the risk profile isn't uniform.

This picture doesn't weaken the case for acting early. Child protection, employment buffers, and cyber defense shouldn't wait until harm is fully proven. It just keeps policy from being held hostage by an unproven doomsday curve. China's 2026 rules on anthropomorphic interactive services already ban providing minors with virtual relatives, virtual companions, or other virtual intimate relationships, and require a minor mode, reality reminders, and measures to prevent emotional dependence. Institutions can set guardrails while evidence is still evolving—but they must leave room for review and adjustment.

AI can deliver broad benefits, but that doesn't mean they'll spread automatically

The benefits Gates lists are concrete. Small hospitals could use AI to spot acute conditions earlier. Farmers in low-income countries could get advice on weather, seeds, disease, and soil. People applying for health insurance, student aid, or food assistance could fill out fewer forms and wade through less bureaucracy. Teachers could spend more time on one-on-one and small-group instruction. Research teams could scan vast literatures for more promising experimental directions.

These scenarios share a common structure: AI turns scarce, expensive expertise into cheaper services. What actually determines fairness is whether each capability can pass through five gates in sequence: language coverage, affordability, available channels, local institutional capacity, and public trust.

If any link in that chain breaks, the technical capability stays in the demo. No local language support, and farmers can't ask questions. Too expensive, and small hospitals can't buy it. No frontline staff to act on advice, and diagnosis never becomes treatment. Unclear privacy and liability, and the public won't hand sensitive problems to the system.

The key word in "AI can benefit people at every income level" is "can." Equity isn't a property of the model itself—it's produced by procurement choices, language support, infrastructure, public budgets, and service design. The Gates Foundation's partnerships with multiple model companies show that philanthropy and public deployment can fill some of the market gap; they don't prove benefits will spread on their own.

First choice: build transition management across departments and borders

AI companies should certainly be responsible for product safety, but they have no democratic mandate to decide trade-offs across employment, taxation, education, child protection, public energy, and social security. Real decision-making requires elected officials, technical experts, workers, teachers, healthcare staff, local communities, parents, and other affected parties at the same table.

The biggest structural problem with existing governments is that they're siloed by department. Labor agencies see jobs, security agencies see attacks, energy agencies see data centers, education agencies see students, tax agencies see revenue. A single wave of AI diffusion can trigger all of these consequences at once, yet no single agency has the full mandate.

The gaps in the matrix are where responsibility falls through. At the national level, you need a coordinating mechanism that can prioritize across departments, assign responsibility, and spot the blanks. Internationally, you need to address cross-border attacks, model and chip supply chains, development norms, and verification. Nuclear weapons inspections, international aviation rules, and the ozone protocol can supply components—but they can't simply be copied into an AI institution.

The direction of this proposal is clear, but the design is still rough. The essay doesn't answer what powers a new institution would have, how it avoids becoming another slow layer of bureaucracy, how it divides work with existing regulators, or how far the US and China can cooperate under competitive pressure. The most workable first step is actually modest: before a crisis hits, establish standing meetings and fault lists where economists, technologists, companies, workers, and local governments regularly point out where existing institutions can't reach.

Second choice: which jobs stay human

"Human Reserved" is the most distinctive concept in the essay. It borrows the logic of nature reserves: just because a piece of land can be paved over doesn't mean it should be. Just because a job can be automated doesn't mean society has to allow it.

Gates illustrates this value through his experience caring for his father. A caregiver can read subtle signals to tell whether an elderly person is hungry, and can maintain trust when the person struggles to communicate. Similar questions arise in delivering serious medical news, education, and mental health services. A machine may produce the right content, but humans still bear the relationship, the responsibility, and the consequences.

Human Reserved can't become a permanent whitelist of jobs. It works better as four tests: how much is lost when human relationships disappear; whether mistakes must be attributable to someone accountable; whether affected workers can transition to other roles in a reasonable timeframe; and whether the local area was already short on labor.

These four tests produce three kinds of outcomes. Roles with high relationship value and high accountability should stay human-led—technology can extend human capability without taking over the final relationship. Industries with difficult transitions and strong regional variation suit gradual hybrid models. Tasks with severe labor shortages, low relationship value, and reversible errors can adopt automation more aggressively.

Japan's aging society might welcome care robots to fill staffing gaps, while a region with ample labor might prefer keeping human caregivers. That difference is exactly why the boundary has to be adjusted by location and over time through open processes. The genuinely hard questions are still ahead: who gets to draw the line, how companies comply, and how cross-border trade handles different standards. Human Reserved gives us something to evaluate; it doesn't yet provide an enforcement system.

Third choice: let automation gains pay for transition costs

Current tax systems typically levy payroll taxes on employees while letting businesses deduct investments in machinery. When companies compare hiring people versus buying machines, the tax code is already adding a weight to the automation side of the scale.

As more tasks shift to AI and robotics, governments could face two pressures at once: wages and the payroll tax base shrink, while spending on retraining, unemployment assistance, and local economic repair goes up. That's why Gates again argues for taxes on AI tokens and robots—so automation gains bear some of the transition cost and the incentive to shed human workers slows slightly.

The argument captures a real fiscal contradiction, but it's a long way from tax law. Tokens are units of computation that can flow across models, tasks, and borders. Robots can be depreciated, leased, or embedded in services. Whether public-benefit uses like healthcare and education get exemptions would immediately create classification and avoidance space. Tax too broadly and you suppress beneficial adoption; tax too narrowly and you collect nothing.

The principle most worth keeping is that the tax code shouldn't push one-way substitution toward automation—and the revenue should be explicitly tied to affected workers, people with reduced hours or wages, and communities where losses concentrate. The specific base could be tokens, compute, capital gains, excess profits, or automation equipment; the essay doesn't prove which is best.

The framing is the most valuable part; the predictive certainty is the weakest

The essay's strongest contribution is turning AI fairness from a well-meaning slogan into three categories of institutional questions:

1. Decision power: who gets to manage the transition across departments and borders, so no institution only sees its own slice.
2. Value boundaries: which relationships, responsibilities, and dignities can't be priced solely by cost-efficiency.
3. Funding sources: when automation reduces labor demand, who pays for retraining, safety nets, and affected communities.

The weakest part is the overconfident tone around some highly uncertain predictions. Whether AI will cause sharp net job losses across industries within a decade, when models might approach error-free long-term autonomous work, and when dexterous robots will become price-competitive in construction and hospitality—none of this is established by current evidence. Stanford's youth employment gap is worth watching, and the ILO's task analysis reminds us that most jobs are more likely to be restructured first.

Policy doesn't need to wait until every prediction comes true. Cross-department monitoring, child safeguards, retraining budgets, and accountability systems can all be built now. Preparing early also doesn't mean accepting a fixed unemployment timeline. The most robust path is to put these three institutional choices on the table today while continuously updating employment, well-being, education, and safety data.

Gates leaves the question with governments, companies, schools, communities, and individuals: who gets the capabilities AI creates, who bears the costs of transition, and which relationships do we insist on keeping human. The defining changes of the next decade may well be decided inside these three choices.