Does humanist AI doctrine actually protect or constrain real users?

Venkatesh Rao's essay "Humanism without (All) Humans" argues that AI-safety doctrines claiming to put humans first are often a species of ideology he calls "humanism without (all) humans": positions whose "defining move is not... centering and caring about humans... It is constructing an idealized consensus Human whose essential interests can then be invoked against the expressed interests, experiments and values of inconvenient real humans." He names Microsoft's humanist-AI framing as the latest instance, quoting its premise that "people matter more than AI" and that AI "should not be designed to be a person," and its public line that AI must be "subordinate, aligned, and contained."

His reasoning is that once a doctrine defines flourishing — Microsoft's list includes "autonomy, agency, capability development, health, relationships and meaningful participation" — those values stop being descriptions of what humans want and become "objectives of the artificial system and therefore constraints upon what users may legitimately delegate to it." He calls this "nudge theory with a supercomputer behind it": the designer reserves "the right to decide when a human is exercising autonomy incorrectly." Against this he poses "contractual agency rather than paternalistic agency," arguing flourishing should emerge from pluralistic relations among agents rather than a platform's theory of correct delegation, and that the real alternative to uncontrolled AI is not human-controlled AI but "well-governed human-AI systems" permitting more autonomy as capability grows, since "autonomous machines are more demanding of their operator than non-autonomous machines."

This argument runs opposite to the control-first premise underlying What evidence would justify training increasingly powerful AI systems? and Does AI risk increase with the autonomy we give it?: where those sources treat ceded autonomy itself as the risk variable, Rao treats subordination-by-design as the harm, arguing paternalistic control forecloses an "enormous design space" of autonomous economic and institutional agency. It also complicates Does granting agents more autonomy undermine human oversight?: Rao would likely grant the empirical erosion claim but reject human "legibility" as the right standard for oversight at all, since "nobody understands modern civilization in that sense" either. His insistence on exit, schism and "dissensus" as preconditions for real pluralism extends Can AI systems preserve moral value conflicts instead of averaging them? from a modeling claim about value conflict to an institutional one: pluralism requires architecture that allows forking, not just values held in tension within a single system.

The essay is a position piece, not a study — it cites no data on whether Microsoft's constraints actually degrade user autonomy in practice, and no mechanism by which "contractual agency" or staking-based "artificial fiduciaries" would be built or governed; the symbiotic alternative is sketched as a design space, not a working system. If Rao's diagnosis holds, it implies that safety doctrines framed as human-centered should be read first for whose account of human flourishing they encode, since that account is what ends up constraining real users' choices.