Writing in the Financial Times, Cristina Criddle reports that Bill Gates wants governments to consider designating certain jobs as “Human Reserved”: work deliberately left to people even when artificial intelligence or robotics could technically perform it.
At first, the proposal sounds defensive—as though humans need a protected enclosure because machines are becoming more capable. But it raises a deeper question.
The relevant test is no longer simply: “Can AI perform this task?” We must also ask: “What is lost when a human no longer performs it?”
A service is not merely its output
A teacher does more than transmit information. A counsellor does more than identify a problem and recommend an intervention. A lawyer does more than retrieve authorities. A caregiver does more than monitor medication, meals, and movement.
These roles contain functions that are difficult to measure: trust, presence, moral accountability, encouragement, judgment, and the experience of being recognised by another person.
Gates illustrates this through the caregivers who supported his father during Alzheimer’s disease. They learned to understand needs he could no longer clearly express. A sufficiently advanced machine might eventually recognise the same behavioural patterns. That does not necessarily mean it should replace the relationship.
Efficiency measures whether an output was produced using fewer resources. It does not tell us whether the process itself carried human value.
Children particularly need human friction
This distinction becomes especially important in education and child development.
AI companions can be endlessly available, patient, and agreeable. Those qualities make them attractive—but they may also make them developmentally incomplete.
Human relationships require children to tolerate misunderstanding, negotiate differences, read emotional cues, repair conflict, and recognise that another person has needs independent of their own. An artificial companion optimised to maintain engagement may remove precisely the interpersonal friction through which social maturity develops.
AI can still be enormously useful in education. It can provide personalised explanations, reduce teachers’ administrative workload, and offer support where specialist services are scarce. But the appropriate model is augmentation: technology extending the reach of a responsible human, not removing human oversight from the relationship.
The strongest objection is access
There is a serious counterargument. Reserving services for humans could make them slower, more expensive, and less available. A student without a tutor may benefit from an AI tutor. A rural patient may prefer an AI-supported consultation to no consultation. Someone unable to afford legal advice may gain meaningful assistance from an automated legal tool.
“Human Reserved” must therefore not become a professional protection scheme that preserves expensive monopolies while excluding people from help.
The better principle is not that AI should never deliver important services. It is that institutions should not remove meaningful human access merely because automation is cheaper.
Where AI expands access, reduces administrative friction, or improves professional judgment, it should be welcomed. Where it removes accountability, exploits dependency, or replaces relationships that are themselves part of the service, greater restraint is justified.
Policy must decide what markets cannot
Commercial markets are unlikely to resolve this distinction themselves. If an automated system is cheaper and scalable, organisations will face pressure to adopt it—even where the social costs appear somewhere else: in weaker relationships, reduced professional development, displaced workers, or children whose primary interactions are increasingly engineered by corporate systems.
Gates also highlights a structural imbalance: employers pay taxes when hiring people while investment in automation may receive favourable treatment. Society therefore subsidises the replacement of labour and then treats the resulting disruption as an individual failure to retrain.
Not every technically replaceable role should be legally protected. But neither should capability be confused with permission.