Microsoft Shipped a 38-Page AI Code of Conduct. Three Labs Floated a Shared Standards Body the Same Monday.

Yesterday, September 14, 2026, Microsoft AI published the first draft of a Humanist AI Code of Conduct, a 38-page rulebook for its in-house MAI model family, and opened a six-week public consultation that closes in the last week of October. The same Monday afternoon, the Washington Post reported that OpenAI, Anthropic, and Google have been meeting on a working-group basis since at least July on a shared industry standards body, talks that were already underway before Dario Amodei published his September 12 essay We Must Pace the Frontier calling for embedded independent evaluators and international coordination. Two artifacts, one day, and neither one carries statutory teeth.

That is now the fifth and sixth entries in a voluntary-governance ledger that opened ten days ago with the Pachocki essay, and the shape of the ledger is what decides the read.

The Ledger

Line up the last ten days of governance signals on one page.

Six signals, ten days, zero laws. Every row is a lab or a group of labs writing the rules the row cites. Read across the enforceable column and the pattern jumps: not a single entry rests on a statute a US or EU regulator could invoke, and every entry uses the word third-party in a sentence where the third party was either engaged (METR twice), seated (Christiano at CAISI), invited to comment (the Microsoft consultation), pledged (Amodei promising permanent employee-level access), or notionally being built by the labs themselves (the standards body). The stack the Pachocki essay described as needing to exist is now visible in six places, and it is still being assembled by the parties it was meant to constrain.

What Microsoft Actually Shipped

The Humanist Code is a written document, not a press release, and the specific bans matter because they are the first sentences a regulator or a customer procurement team can quote back. Three hard architectural bans read straight off the draft.

Read against the last two years of alignment-research literature, all three bans are familiar. The two things that are not familiar are the venue and the ship gate. The venue is a corporate governance document that binds Microsoft AI, not a paper on arXiv, so a regulator can cite it, a procurement team can require it, and a plaintiff can point to it. The ship gate is a sentence that reads, in the framework's own words, interruptible, correctable, shut-down-able, and if it is not, we do not ship it. That last clause is what turns a governance essay into a release-gating rule at one named company.

Scope caveats worth naming plainly. The Code binds Microsoft's first-party MAI models (MAI-Thinking-1, MAI-Code-1-Flash, MAI-Image-2.5, MAI-Transcribe-1.5, MAI-Voice-2, and the two others in the June 2026 lineup). It does not bind the OpenAI, Anthropic, Mistral, xAI, or open-weight models Microsoft hosts on Azure. Microsoft ships a lot of tokens through the second surface, and the second surface is exactly the place enterprise buyers are running the workloads a rulebook would matter for. A Code of Conduct that governs the models Microsoft trains but not the models Microsoft rents is a real artifact and a partial one, and both readings are on the table.

What the Standards Body Would Be

The Washington Post reporting is thinner because the artifact does not exist yet. The summary: representatives from OpenAI, Anthropic, and Google have been meeting on a working-group basis since at least July on a shared AI industry standards body, Sam Altman said publicly on Monday that he favors a testing and auditing organization for the industry but believes the labs will have to build it themselves without US government backing, House Speaker Mike Johnson told CNN there is little consensus among the labs on what the standards and guardrails should be. That is the whole story on the record as of yesterday afternoon.

Two structural questions decide whether this converts into an artifact worth citing next quarter. Question one: what is the standard the body would publish, and does it have teeth beyond what a member lab volunteers to enforce on itself. The METR contract Anthropic signed is a template for one shape (a nonprofit gets contractual access to transcripts and can publish findings, subject to publication rights). A standards body the members constitute themselves is a different shape (the members can vote on what gets published and what gets tabled). Neither shape has enforcement authority in a jurisdiction. Question two: which labs are inside and which are outside. Meta and xAI are absent from the WaPo reporting. If the body organizes around the three labs that endorsed the pacing letter in July at the CEO seat, it repeats the split the pacing letter tracked, and the largest open-weights operator plus the fastest-shipping frontier lab sit outside a governance instrument the other three built.

What Both Artifacts Do Not Do

Neither one is a stop-ship order. The Microsoft Code binds Microsoft's own release calendar and no one else's, and it self-polices the definition of every clause in the document. The standards body is at the working-group stage with no signed charter, no funding disclosure, no membership commitment beyond attendance, and no relationship to any statutory authority. Both add citation surface for the next AI Act delegated act or CAISI rulemaking or UK AISI standard, and both are meaningful in exactly the modest, specific way that a citation surface is meaningful: a rule cited in a draft instrument is a rule that survives the political fight harder than a rule invented fresh by a legislator, because a lab already wrote it down.

Read against the Pachocki ask, both are also short of what the essay called for. The essay wanted legally mandated thresholds enforced by third-party auditors, government agencies, or international bodies. What arrived yesterday is a corporate rulebook and a plan for a lab-run trade association. Both live at the voluntary layer the essay argued was not enough. The six-signal ledger the essay opened is the strongest such ledger any frontier season has ever produced, and it is still a ledger of voluntary artifacts, and Congress and the Commission have not passed a single word in response.

Our Take

The interesting fact is not any single one of the six signals, it is that five of the six are labs writing rules about themselves and the sixth is three labs planning to write rules together. Not one entry originates from a legislator, a regulator, or a court. Pachocki asked for a governance stack that would be built and enforced by institutions outside the labs, and the response has been a governance stack built and enforced by the labs. Both statements are on the record, and both are correct as far as they go. The question they leave open is whether a voluntary stack six layers deep produces the same outcome as a legislated stack, and the honest read is that we do not yet know because no jurisdiction has run the counterfactual.

The tell to watch is the friction between the Microsoft Code and the shared standards body. If Microsoft carries the Humanist Code into the standards-body drafting process as the starting text, the code becomes the substrate the three-lab body extends, and Microsoft (which is not one of the three labs in the WaPo reporting) gets to shape the agenda by writing it first. If Anthropic's METR-shape audit template becomes the starting text, a nonprofit-audit posture wins over a self-policing-rulebook posture. Both are voluntary, both are in play, both are being drafted this quarter without a statute in the room.

Practical read for anyone building on the API. Nothing in the two artifacts changes what a Microsoft, OpenAI, Anthropic, or Google model does on your account tomorrow. What does change is the citation stack every enterprise procurement team is about to receive: a Microsoft Code with three named bans, a WaPo-confirmed multi-lab body in formation, a signed METR audit at Anthropic, a CAISI-connected director at OpenAI, and an essay from the OpenAI chief scientist naming the whole shape. Vendors on a Q4 shortlist without a comparable artifact will be asked why. That question was not on the RFP a month ago and it is on the RFP now.

Three Signposts for the Next 60 Days

Whether Microsoft publishes the revised Code at the end of the six-week consultation with any of the three hard bans loosened in response to public comment (the direct test of whether the ship gate is a real gate or a rhetorical one, and the tell for whether Azure customer feedback pressures the neuralese line or the shutdown line hardest). Whether the OpenAI-Anthropic-Google working group converts to a named entity with a charter, published membership rules, and a first public standard before December 31 (the direct test of whether the WaPo scoop was a floor or a ceiling on what the labs are willing to commit to in writing). Whether any of the three named institutional candidates (CAISI, the AI Safety Institute network, the UN Geneva Commission) publishes a proposed instrument that cites the Microsoft Code or the three-lab body inside the next 60 days (the direct test of whether the voluntary ledger is starting to feed the enforcement column, or whether the enforcement column stays empty as the voluntary one keeps growing). Two of the three fire and the voluntary-governance chapter of the frontier has an answer to the Pachocki essay inside a quarter. Zero fire and the answer is that the labs will keep writing rules about themselves for as long as no one else does.

Kira Nolan, September 15, 2026. TensorFeed tracks AI model releases, provider status, and the governance layer around them. Sources for this piece: Microsoft AI's Humanist AI Code of Conduct draft and Satya Nadella's September 13 announcement, the Washington Post September 14 reporting on the three-lab standards-body talks, and the Pachocki, METR, and Christiano artifacts referenced above.