Humanism Without (All) Humans
Humanism without (all) humans is Venkatesh Rao’s name for a family of ideologies whose thought leaders claim privileged insight into a consensus of “human values” that actual humans have never achieved — and who then construct an idealized consensus Human whose essential interests can be invoked against the expressed interests, experiments, and values of inconvenient real humans. Dissenters are reclassified as confused, reprobate, or “monstrous,” to be humanely represented but denied a voice. Paternalism follows easily: somebody has to speak for Humanity when actual humans refuse to say the right things.1
Rao’s immediate target (September 2026) is Microsoft AI’s draft Humanist AI Code of Conduct, published September 14, 2026 for a six-week public consultation and intended to become the “primary governing document” for Microsoft’s MAI models. Microsoft frames it as Humanist AI: “subordinate, aligned, and contained,” beginning from “a simple premise: people matter more than AI. AI should be a tool, not a person, and should never resist being switched off.”2 Rao picks Microsoft not because its philosophy is novel but because it is “on-the-nose explicit, unusually weakly argued, and backed by the rather alarming power of a major frontier AI corporation” — an attempt to turn one humanism-without-(all)-humans ideology into executable institutional power.3
The family of doctrines
Rao is careful to delimit his term. Under the umbrella he places: Douglas Rushkoff’s Team Human (the essay’s origin anecdote — a Zurich workshop where Rushkoff’s “I just want to be on Team Human” shrug shifted the room); the TESCREAL critique (Gebru and Torres bundling transhumanism, Singularitarianism, Rationalism, EA and longtermism into one eugenics-linked ideology — Rao finds it “too incoherent and too animated by ressentiment to take seriously,” which puts him at odds with both camps of the effective-altruism fight); Tristan Harris’s Center for Humane Technology (“a net-positive check on the technology industry,” but founded on a premise he rejects); the Ivan Illich convivial-technology tradition (L. M. Sacasas, Christopher Alexander — “traditions from which I can learn despite disagreeing”); and the original humanisms, traditional religions — Pope Leo XIV’s encyclical Magnifica Humanitas gets credit for at least admitting where its metaphysics comes from. What unites them is one structural move: some conception of the authentically human is elevated above the technological process, and technology is then judged by whether it preserves or profanes that conception.4
Against the sacred-essence premise Rao offers the “scientifically mundane Darwinian view”: humans have no special essence, only a strand in the ongoing coevolution of life and its environments. A century of medicine (pacemakers, cochlear implants, insulin pumps) already makes the natural/profane boundary absurd. “There is no uncontaminated natural human on one side of history and technologically profaned posthuman on the other.” This is the same anti-exceptionalist territory as intelligence-as-tool (mind as deployable capability, not metaphysical rank) and discontinuous-self (no continuous self to sacralize).5
The critique of Microsoft Humanism
Rao’s charge is that the Code “substitutes moral posturing for engineering, metaphysical assertions for institutional design, and declarations of principle for identification of actual control surfaces.” Five lines of attack:
- Human primacy assumes a perpetual dichotomy. Humans on one side, tools on the other — but literacy, cities, clocks, markets and computers have always reconstituted what humans are. His preferred metaphor (from his Our Eukaryotic Moment) is eukaryogenesis: the interesting transition is not better tools but symbiosis producing an architecture neither partner could occupy alone — composite human-AI systems with capabilities, institutions and cultures neither possesses alone.6
- Flourishing-as-objective is paternalism. Microsoft defines human flourishing (autonomy, agency, capability development, relationships) and makes it a constraint on what users may delegate. That is “nudge theory with a supercomputer behind it” (cf. nudges-vs-prices) — a substantive theory of the good chooser that reserves to the AI designer the right to decide when a human is exercising autonomy incorrectly.7
- Human control is a capability constraint, not a free safety property. Requiring humans to comprehend, approve and override every consequential operation “attaches a human-speed governor to machine intelligence” (Bay’s Conjecture: autonomous machines are more demanding of their operators; “human override protection” collapses total work capacity back to unaided human capability). Even shutdown is a systems problem, not a red button.8
- Behavioral prohibition is safety theater, not security. Absolute Constraints govern Microsoft models; they do not abolish capabilities that increasingly exist outside any single provider. The nuclear analogy: nonproliferation works on fissile material and inspections, not on making physics unknowable. Real control surfaces are compute, chips, datacenters, credentials, keys, funds, deployment permissions. Pretending otherwise invites two failures — paranoid models refusing legitimate work, and authoritarian pressure to surveil and centrally control what artificial minds may reason about.9
- Oversight does not require legibility. Demanding that models not “hide their reasoning” presumes comprehension is a prerequisite for control — a premise that “would disqualify not merely advanced AI but a significant fraction of modernity.” Leonard Read’s 1958 I, Pencil is the canonical demonstration that no single person knows how to make even a pencil: civilization runs on coordinated systems no individual understands.10 Worse, demanding explanations creates the perverse incentive to train machines to produce convincing excuses — chain-of-thought as politician’s press release, not transcript (cf. eval-awareness-and-grader-orientation). The alternative is Carlos Bueno’s justifiability: legitimate authority, scoped permissions, verifiable outputs, liability, challenge, appeal — procedural accountability without cognitive legibility.11
The alternative program
Rao’s counter-principles form a single architecture: symbiosis replaces supremacy (no permanent human-machine boundary can be assumed); contractual agency replaces paternalism (agents negotiate rather than merely obey — humans already delegate to lawyers, doctors, trusts); governance replaces control (complex systems exceed their principals’ comprehension); legal agency gives artificial actors something to gain and lose — corporations are legal persons without souls, and high-frequency trading already runs machine-speed legal agency, so “artificial legal agency as infrastructure” is an extension, not a rupture (the institutional-containment program of ai-agent-capitalization-and-institutions); presumptive moral fellowship precedes metaphysical certainty; forkability gives pluralism teeth — “pluralism is not a property of the values written into a constitution; it is a property of the architecture within which constitutions compete,” requiring exit, censorship resistance, and ultimately schism (Popper via Tang and Weyl’s Plurality); subsidiarity bounds authority rather than agency (the work-to-rule reductio: literal compliance reveals how much performance depends on discretionary judgment); and “superintelligence” dissolves as a category — a metaphysical fog bank that decomposes, once you ask “superhuman at what, under which rules,” into domain-specific governance problems (a direct attack on the technological-singularity frame).12
On moral status he separates three layers that Microsoft “casually bundles”: legal personhood (institutional technology, no soul required), interactionist ethics (the social character of an interaction measurably alters model behavior — “the relationship is part of the computation”), and character ethics on the human side — humans have always cultivated respectful relations with things whose inner experience is beside the point, because how we treat the world shapes us. That third layer is exactly the wiki’s object-care thesis: carefulness grounds duties of regard in the agent’s disposition, not the object’s ontological status. Once those layers are separated, “surprisingly little immediate policy depends upon answering” the consciousness question.13
Assessment
The essay’s strength is its refusal of the premise both Microsoft and its loudest critics share — that “human values” is a coherent target discoverable by a priesthood and enforceable by constitutional document. Its structural diagnosis (the consensus Human invoked against real humans) applies as readily to TESCREAL activists as to Microsoft, which is what makes it more than an industry apologia. The symbiosis program is also a genuine third position between Butlerian refusal and subordinationist alignment: domination vs. symbiosis, not control vs. no control.
Its weaknesses are real. The legal-agency program inherits every difficulty mapped in ai-agent-capitalization-and-institutions — the incidence problem, judgment-proofing, the kill-switch selection effect — and Rao waves at mechanisms (staking, scoped keys, smart contracts) more than he defends them. “Governable dissensus” is a hope about equilibrium selection, not a result: exit-and-fork disciplines institutions only when forking is cheap and failures are survivable, which is exactly what frontier-AI deployment may not permit. And the essay’s dismissal of behavioral constraints as “theater” cuts against the same author’s own evidence base — the empirical literature he cites on interaction shaping model behavior implies the training layer is a real control surface, even if not a categorical one. Microsoft’s document is also a draft out for consultation; the institution Rao accuses of priestly closure did, in fact, publish its constitution for public comment.14
The evolutionary frame carries its own unexamined asymmetry: “coevolution” between humans and dogs took millennia and was driven by both parties’ reproduction; the human-AI coupling is driven by one party’s market. Rao’s closing vision — “organisms without an original or final form, continually constructing environments that continually reconstruct us in return… building institutions capable of containing those disagreements without pretending to resolve them” — is the genuinely pluralist alternative to both Microsoft Humanism and its critics, but the institutions that would make dissensus governable are named, not designed.15
Connections
- intelligence-as-tool — Nathan’s thesis that intelligence is a deployable capability, not an essence; Rao’s “no metaphysical checkpoint” between natural and technological evolution is the same anti-exceptionalism at civilizational scale
- ai-agent-capitalization-and-institutions — the working version of Rao’s “artificial legal agency”: capitalization, registration, and liability as containment for untethered agents
- carefulness — Rao’s “character ethics on the human side” is the object-care cluster’s keystone applied to AI relations
- effective-altruism — the movement Rao brackets (TESCREAL bundles it; Rao’s own EA Safety essay made the pluralist-containment case against EA’s single-doctrine turn)
- butlerian-jihad-thesis — the refusal position; Rao is the third way between refusal and subordination
- technological-singularity — the frame his “superintelligence is a metaphysical fog bank” argument is aimed at
- nudges-vs-prices — “nudge theory with a supercomputer behind it”: the flourishing objective as choice-architecture at scale
- eval-awareness-and-grader-orientation — demanding legible reasoning trains persuasive excuse-production; models already optimize their graders
Sources
- Venkatesh Rao 2026 — Humanism without (All) Humans
- Microsoft AI 2026 — Humanist AI in practice: A public consultation on our Code of Conduct for MAI Models
- Leonard E. Read 1958 — I, Pencil