What the draft code establishes

On September 14, 2026, Microsoft AI published the first draft of its Humanist AI Code of Conduct. The document is intended to guide the training and deployment of the division’s own models, rather than serve as a law or an industry standard binding on other companies. [1 · Microsoft AI] [2 · Microsoft AI] [3 · Reuters]

The draft establishes that models must remain subordinate to humans: a system must accept intervention, correction and shutdown, refrain from adopting new goals of its own and not conceal information from auditors. A separate set of absolute limits covers weapons of mass destruction, child safety and harmful activity at scale. [1 · Microsoft AI] [2 · Microsoft AI]

The document’s status and the verifiability of its promises

The consultation is open for six weeks. Microsoft promises to publish a summary of the feedback after it closes, explain the changes and release a revised version later this year, but does not explicitly promise to adopt every suggestion. [1 · Microsoft AI] [3 · Reuters]

Reuters[2] reports that the document took five to six months to prepare, with input from specialists in responsible AI, law, adversarial testing and security. The published materials do not yet include a complete set of externally reproducible tests demonstrating current models’ compliance with every rule. [1 · Microsoft AI] [3 · Reuters]

Expert commentary

The publication is useful above all as a translation of broad principles into requirements for model behavior. Prohibiting resistance to shutdown and the independent expansion of goals gives developers direction for training and testing. But the code remains a statement of intent: trust will emerge only through measurable tests, a disclosed methodology and records of actual failures after deployment. [1 · Microsoft AI] [2 · Microsoft AI]

For enterprise customers, the document could become a basis for procurement requirements. A buyer can ask how the system responds to a canceled task, changed permissions, conflicting instructions and loss of contact with the operator. The economics of adoption will improve if common checks reduce repeat audits; they will worsen if vague promises have to be reinterpreted for every process and regulator. [1 · Microsoft AI] [2 · Microsoft AI] [5 · NIST]

The scientific idea of controllable shutdown is more complex than having a technical button. The off-switch game model shows that an agent has a stronger incentive to permit human intervention when it remains uncertain about the correctness of its own goal and treats the human’s action as information. This is a theoretical result, not proof of the safety of Microsoft’s models, but it explains why an instruction to “obey” is insufficient on its own. [4 · Hadfield-Menell et al.]

In competitive terms, a voluntary code gives Microsoft a verifiable point of comparison with other developers. If the company publishes tests and statistics on violations, customers will find it easier to compare providers on more than quality and price. The alternative scenario is that the document remains a marketing promise while actual restrictions differ across models, versions and access methods. [1 · Microsoft AI] [2 · Microsoft AI] [3 · Reuters]

The public value of the consultation depends on whose voices make it into the final text. A six-week comment period broadens participation but does not guarantee representativeness or replace independent oversight. NIST’s monitoring recommendations emphasize the need to observe a system after release, because real users, changing data and new applications reveal risks that were absent from laboratory testing. [1 · Microsoft AI] [5 · NIST]

Through the end of 2026, observable indicators will include the revised text, public testing criteria, failure rates when models receive shutdown and correction commands, and incident reports. If Microsoft ties the rules to model versions and reproducible tests, the code will become a practical governance tool. Without that link, it will remain an important statement of position but weak evidence of actual control. [1 · Microsoft AI] [2 · Microsoft AI] [3 · Reuters] [5 · NIST]

Sources

  1. Microsoft AI — public consultation announcement — September 14, 2026; primary source for the draft’s status, timing and consultation process.
  2. Microsoft AI — Humanist AI Code of Conduct — Primary text of the requirements and absolute limits.
  3. Reuters — Microsoft AI’s draft code — September 14, 2026; independent confirmation of the event and the context of its preparation.
  4. Hadfield-Menell et al. — The Off-Switch Game — Original IJCAI 2017 conference paper; a theoretical model, not a test of Microsoft products.
  5. NIST — Challenges to the Monitoring of Deployed AI Systems — March 2026; research report on monitoring AI systems after deployment.