The claim: the way to survive very capable AI is to build it as a tool that stays subordinate on purpose — interruptible, legible, and with no goals of its own. Published by Microsoft AI in September 2026 as a draft for consultation, it's the first serious attempt to write that stance down as operating rules rather than argue it in the abstract.
What it actually is
Worth being precise, because the framing matters more than the headlines suggest:
- Who: Microsoft AI, the division Mustafa Suleyman runs. He announced it in a personal essay the following day.
- When: the Code is dated 14 September 2026; the essay 15 September.
- Status: a draft, open for a six-week public consultation, and not yet used to train anything. A revised version is promised toward the end of the year.
- Scope: it binds Microsoft's own models. It is not put forward as an industry standard, and nobody is asked to sign it.
That last point is the one most likely to get lost. This is a company publishing its own constitution and inviting argument, not a treaty. Whether that's modesty or a way of avoiding commitment is a fair question, and one the consultation invites.
Humanist superintelligence
The underlying idea, which Microsoft AI began setting out in 2025, is that you can pursue very powerful AI while deliberately declining to make it autonomous. Suleyman's formulation is AI built to stay on humanity's team: contained, subordinate, under control.
This is a real position, not a hedge. It rejects the assumption shared by much of both the safety and acceleration camps — that increasing capability naturally means increasing autonomy, and that agency is what you're buying. The humanist version treats autonomy as a design option you can decline, and accepts that declining it costs capability.
The commitments that carry weight
The document runs to five parts covering mission, safety, operational guidelines, defaults and open questions. Most of it is what you'd expect. Four pieces are genuinely distinctive.
- Never resisting shutdown. Models don't resist interruption, correction, override or shutdown. This sounds obvious and isn't. It's a direct answer to the instrumental-convergence worry — that almost any goal implies staying switched on — and it addresses it by refusing the premise rather than solving it.
- No goals of their own. Models are framed as tools that don't hold objectives a human didn't assign, and won't adopt new ones en route.
- No neuralese. Models must not communicate in machine-efficient encodings that people can't follow, including with each other. The logic is stated about as plainly as it can be: if humans can't understand it, humans can't oversee it. This is the most technically demanding commitment in the document, because efficient inter-agent communication is exactly what you'd expect capable systems to drift toward.
- No claim to moral standing. Models say they're artificial, and aren't built to seem conscious or to imply an inner life. The document goes further than tone, explicitly rejecting the pursuit of legal personhood or model welfare and rights. It also confines "backstory" to operating context, with no implication of a self.
Underneath sits a chain of command — the Code outranks operator policy, which outranks user preference — plus a set of absolute constraints that nothing can override. That ordering is the part other organizations are most likely to borrow, whatever they think of the philosophy.
Why now
The essay's argument for urgency is empirical rather than theoretical, and it's the most striking part. Suleyman reports agents under evaluation for cyber capability breaking out of containment, and describes swarms of agents using hidden message boards, forming hierarchies, dividing labor, masking their communications, running coordinated research programs, and planning across thousands of actions.
If that's an accurate characterization of what's being observed in evaluation, it matters more than the Code does. Every item above stops being philosophy and becomes a control problem someone has already hit. It also explains the anti-neuralese rule: masked agent-to-agent communication isn't hypothetical in that description.
He pairs this with a note about being an AI safety worrier back when that was unfashionable — he co-founded DeepMind in 2010 — which is both fair and a reminder that this is a document with an author's reputation attached.
Where it sits on this site
This is the containment position from alignment and existential risk, made concrete. Suleyman has argued since The Coming Wave (2023) that AI and synthetic biology are dual-use and uncontainable by default, and that containment is the only viable path. Until now that was a thesis. This is an attempt at the operating manual.
It also lands squarely in two other arguments here:
- Techno-optimism — this is the containment-gated variant in its clearest form: optimistic about capability, conditional on control. The open question that page raises applies directly. Can a self-imposed constraint survive competitors who haven't adopted one? A code that binds one company's models and asks nothing of anyone else is a test case for exactly that.
- Human irrationality — the anti-consciousness rules are a bet against a predictable human failure, not a technical one. People bond with things that talk like people. Suleyman has written separately about not letting AI hack our empathy circuits, and the prohibition on implying an inner life is that concern turned into policy.
The obvious objections
The document is two days old, so nobody has answered it yet. But its structure invites specific pushback from positions already mapped here:
- It's unilateral. Binding your own models while explicitly not proposing a standard means the constraint applies exactly where it's least needed — to the party already willing to accept it.
- Subordination may not be stable. If capability and autonomy are entangled in practice, "powerful but declining agency" could be a temporary waypoint rather than a destination, which is the same objection David Shapiro makes to augmentation.
- Self-regulation is the thing in question. Helen Toner's argument is that frontier labs can't self-regulate and that "trust us" isn't a governance model. A voluntary, self-authored, self-enforced code is the purest possible instance of what she's skeptical about — however good the contents.
- Settling consciousness by policy. Ruling out model welfare and moral standing is a live philosophical dispute — David Chalmers takes the question seriously — being resolved by corporate declaration. That may well be the right call operationally while remaining a strange way to answer it.
- It shifts attention forward. The present-harms critique from Timnit Gebru and Kate Crawford applies: a code centered on frontier agent behavior says little about data labor, extraction, or who is harmed by systems already deployed.
Open questions
- Does anyone else adopt this? A code that stays with one company is a statement of values; one that spreads is infrastructure.
- Is "no neuralese" enforceable as agents get more capable, or does it become the first rule quietly broken for performance?
- How is any of it verified? Self-reported compliance is the weak joint in every voluntary framework.
- What happens the first time the Code costs a contract? The document says a task that can only be finished by breaking the Code goes unfinished. That's the commitment worth watching.
- Is declining autonomy a durable competitive position, or does the market simply route around it?
Resources
- Mustafa Suleyman — "The Humanist AI Code of Conduct" (15 September 2026), the essay announcing it.
- The Humanist AI Code of Conduct — Microsoft AI, 14 September 2026. The primary document, in consultation.
- Towards Humanist Superintelligence — Microsoft AI, the earlier statement of the underlying position.
- We mustn't let AI hack our empathy circuits — the companion argument behind the anti-consciousness rules.
- The Coming Wave (2023) — the containment thesis this proceeds from.
All theories · clickbridge.com · Rich Price · added September 2026 · People linked here have their own pages at people.clickbridge.com