On August 26, 2026, Bill Gates published a programmatic essay on AI on Gates Notes. He introduces a category he calls "Human Reserved": activities that a society might choose to keep in human hands even if software and robots were to perform them in a technically credible way. On the same day, in an interview with Axios, he added the part that made news: in the most extreme and initial version of this framework, he said he could imagine a share of up to 40% of current jobs reserved for people.
In the text Gates gives very concrete examples: care for the elderly, delivering serious diagnoses, education, mental health. In some cases he imagines strong protection for the human role. In others he describes hybrid models, where the professional remains responsible and uses AI as support. Axios reports that he also added childcare and jury service. Also to Axios, Gates distinguishes between a permanent version of the human reserve and a temporary one, designed for workers who cannot easily be moved from one sector to another. In the essay he even uses the example of a 55-year-old construction worker. The point is that retraining, taken as a universal formula, has very earthly limits.
The proposal carries weight because of its structure. Gates admits that the questions that truly matter remain open: who decides which activities to reserve, by what criteria, what controls can limit circumvention, what happens when one country slows automation and another does not.
If a human work reserve became public policy, a company could no longer replace people with models simply because the balance sheet allows it. Rules and oversight would enter the picture. Exceptions would have to be written down. Litigation would be part of the cost. A share of efficiency would leave management and pass into a political process.
In the lexicon of the AI transition, this shifts the center of gravity. For years the debate has treated the direction of travel as given: automate as much as possible, then redistribute the damage with subsidies and training. At most, a few fiscal corrections. Gates moves the decision one step earlier. A community can choose to preserve human work even when the machine costs less or works faster. The architecture, for now, is missing. He calls for new institutions, talks about taxes on AI tokens and robots, and returns to a line he was already arguing in 2017: slow automation down when necessary.
Protecting human roles can compress productivity, raise prices, and shift investment elsewhere. It can also reward countries that adopt no similar limits. Moreover, "jobs" is too broad a word: almost all occupations are bundles of tasks, and deciding which tasks remain human is harder than drawing a line around a profession. Gates acknowledges these difficulties. That is precisely why the proposal matters. The conflict moves beyond wages and retraining. It also enters the right to set a limit on efficiency.
