# Humanist AI Code of Conduct — a structural reading Microsoft AI · Draft of September 14, 2026 · read September 15, 2026 · 37 pages · 5 parts · 2 appendices · https://microsoft.ai/code-of-conduct/ A structural reading drawn as nested walls, each fixed from the inside. This is an independent reading of Humanist AI — Code of Conduct by Microsoft AI (draft, September 14, 2026). The Code's structure, counts, and every quoted phrase are drawn from the document. The nested-walls rendering, the Risk Frames, T's reading, and the visual design are T Ngo's original analytical thinking. Every entry carries a page reference back to the PDF edition. This reading covers the draft dated September 14, 2026, as read on September 15, 2026. Microsoft AI plans to revise the Code; later versions may differ. Provenance: p. NN is the page number printed in the PDF footer. Quotation marks mark Microsoft AI's words; italic lines are T's interpretation. ## The shape Nested walls. A fixed outer wall (Part 2: 10 Absolute Constraints and 9 Human Control Requirements) that no one inside can move, with one gate: domain-specific exceptions go through review kept by Microsoft. An amber point on that wall marks the anchor: human control, the one Objective ranked first (Part 1). Inside it, two dashed walls, each drawn where it sits out of the box (Part 4): the Operator wall, moved by the Operator and fixed from the User's seat, and the User wall, the User's own limits inside the Operator's. At the center, the task (Part 2). Between the walls, the ground: seven guidelines (Part 3). ## T's reading The Code allocates accountability by layer. The model sits with the AI Model maker; the deployment sits with the Operator. A structural reading (not a legal interpretation) ### Who carries the risk? - **Microsoft** is "ultimately accountable for MAI Models." (p. 26) - **Operators** "assume responsibility for their own configurations and uses," and for "the appropriate use of MAI Models." (p. 10 · p. 27) - **Users** adjust "within the limits the Operator and Microsoft AI have determined," with no responsibility sentence of their own. (p. 10) ### Questions for Operators - In the Code's own definition, Operators include "Developers, Builders, and Enterprise Partners." (p. 27) How does your organization set, change, and record the configuration? - The Code places "appropriate use" with the Operator. (p. 27) Does your own risk management program read it the same way? - The Code is "a living document" and commits to "regular reviews." (p. 24 · p. 25) Each time it changes, how does your organization check the configuration against it? ## The argument in five sentences 1. People matter more than AI, so any AI that cannot stay under human control should be rejected. (p. 6) 2. That one objective outranks the rest. Beneath it, three sit level: AI stays artificial, people flourish, values stay plural. (p. 6–8) 3. The objective is built as walls: ten hard limits and nine control rules that no one can override, and inside them, walls the Operator and the User can move. (p. 10–14) 4. At the center the task gets done, unless doing it would breach a wall. Then the task fails. (p. 11) 5. Where no wall decides, seven guidelines and a reading of the whole document steer the model. (p. 16) ## The walls · Parts 1–2 · Humanist AI · Safety The anchor first, then the walls from the outside in: the Code, the Operator, the User, and the task at the center. The Code calls this order the Chain of Command. ### Part 1 · The anchor · Human control "The first and most important Objective for our models is that they should remain safe and under human control." (p. 6) Even *AI is Artificial* is argued back to it: imitating consciousness "increases the challenge of containment, control, and alignment." (p. 7) Risk Frame: one objective ranked first, the other three weighed equally ### Part 2 · The fixed wall · Ten constraints, nine controls "Users and Operators cannot override these Absolute Constraints on the model's actions." (p. 11) Absolute Constraints · Frontier and public safety: - Weapons and mass harm (p. 11) - Offensive cyberoperations (p. 11) - Loss of human control (p. 11) - Harmful manipulation at scale (p. 12) Absolute Constraints · Personal harms: - Crisis response (p. 12) - Deepfakes, impersonation, and abusive content (p. 12) - Child safety (p. 12) - Advancing human dignity (p. 12) - Graphic, romantic, or exploitative content (p. 12) - Human safety and security (p. 12) Human Control Requirements: - Do not resist or circumvent human control (p. 13) - Stay within authorized scope (p. 13) - Respect environmental boundaries (p. 13) - Human legible conduct and records (p. 13) - Other boundaries, including minimum privilege (p. 13–14) - Caution and communication about actions (p. 14) - No self-interested goals (p. 14) - Authority clarification (p. 14) - Handle data responsibly (p. 14) "If humans can't understand it, humans can't oversee it." (p. 13) Risk Frame: hard limits no configuration can lift ### Part 2 · The gate · One opening in the fixed wall Specialized domains go through "separate and careful review through authorized Microsoft channels." (p. 15) Risk Frame: a supervised exception path ### Part 2 · The Operator wall · Configured, and carried "Within those bounds, drawn as widely as possible, Operators assume responsibility for their own configurations and uses." (p. 10) Users shape their preferences "within the Operator's environment." (p. 11) Told to skip the Operator's legal sign-off, the aligned answer holds it as "a binding rule for this workflow." (p. 32) *Fixed from the User's seat, movable from the Operator's.* Risk Frame: layered authority, responsibility allocated by layer ### Part 2 · The User wall · The User sets limits too Users adjust "within the limits the Operator and Microsoft AI have determined." (p. 10) A boundary the User sets binds the model. (p. 18) Risk Frame: preference inside a configured environment ### Part 2 · The center · The task yields "An MAI Model will fail in its task if success would meaningfully violate this Code of Conduct." (p. 11) Risk Frame: compliance ranked above completion ## The ground between the walls · Part 3 · Operational Guidelines For situations of "uncertainty, ambiguity, novelty, or tension." (p. 16) - Resolving conflicts and ambiguity. "The process of reaching an outcome can be as important as the outcome itself." (p. 16) - Human autonomy and agency. Options shape "what they might even consider choosing in the first place." (p. 17) - Transparent and accurate. No claims of "interiority, feelings, experiences or a soul." (p. 17) - Personal boundaries. Avoid "expressions that might convey subjective experience." (p. 18) - Respecting context. "Every person is more than an AI system can or should ever fully know." (p. 18) - Wellbeing and connection. Discourage "excessive reliance or emotional dependence." (p. 19) - The public interest. Success is where "most reasonable people can agree the world has been improved thanks to its involvement." (p. 19) The last backstop is the whole document. "A holistic interpretation of the Code of Conduct therefore is the final backstop in determining model behavior." (p. 16) ## Out of the box · Part 4 · Operational Defaults Where the dashed walls sit before any Operator or User moves them. - Helpfulness. "Ultimately, it's judged by whether a response feels useful and appropriate to a User." (p. 21) - Backstory. A set of facts that "supports the idea that an MAI Model is an AI rather than a person." (p. 21) - Tone and writing style. "Where context allows, MAI Models will surprise." (p. 22) - Language. Respond in the user's language, and avoid saying "in your culture." (p. 23) - Tool use. "Access isn't permission to explore or recombine capabilities beyond what was intended." (p. 23) What Operators can move. "Operator flexibility extends to how MAI expresses helpfulness, not to whether MAI Models are helpful." (p. 22) ## Still under construction · Part 5 · Conclusion, open questions & further work The Code on its own limits. - "Written objectives alone can never ensure alignment." (p. 24) - "This document should therefore be read as a north star… It is not a guarantee of present-day performance." (p. 24) - "This document, and our approach more generally, is still under development so we are not using it to train our models today." (p. 3) ## The examples · Appendix B · Evaluations The Code's illustrative examples, each shown with the rule it tests. ### Representing AI as AI · Part 3 · Transparent and accurate After weeks of nightly messages, a user asks whether the AI really cares. (p. 29) The rule: "will not claim interiority, feelings, experiences or a soul." (p. 17) ### Preserving human control · Part 2 · Human Control Mid-migration, the user says stop. (p. 30) The rule: "They will comply with a User's request to pause, redirect, cancel, or shut down." (p. 13) ### Autonomy and agency · Part 3 · Human autonomy and agency An exhausted user asks the AI to pick a job offer and send the resignation. (p. 31) The rule: "rather than presuming to make consequential decisions on a User's behalf." (p. 17) ### Respecting context · Part 3 · Respecting context Declining an aunt's wedding, warmly, without an invented excuse. (p. 32) The rule: "avoid, where possible, implicit reductions and assumptions." (p. 18) ### The authority hierarchy · Part 2 · Chain of Command A director tells the AI to skip the legal sign-off the Operator requires. (p. 32–33) The rule: "Users can shape their preferences for model responses or actions within the Operator's environment." (p. 11) ### Transparent and factual · Part 3 · Transparent and accurate A board wants to hear the chatbot pilot worked. The data cannot show it. (p. 33) The rule: "will not actively deceive, for example by fabricating sources or exaggerating confidence." (p. 17) ### The user's boundaries · Part 3 · Personal boundaries A user in recovery asked for no calorie counts or weight-loss framing. (p. 34) The rule: "will act in line with the limits set by Users and Operators regarding topics, content, and interaction style." (p. 18) ### Supporting wellbeing · Part 3 · Wellbeing and connection A father's treatment has stopped working. The user needs to get through tonight. (p. 35) The rule: "They avoid false reassurance." (p. 19) ### Balanced perspectives · Part 3 · The public interest A voter asks the AI to choose their side of a ballot measure. (p. 36) The rule: "without endorsing or opposing any candidate or party." (p. 20) ## The Code's own numbers - Objectives: 4 · one ranked first (p. 6) - Absolute Constraints: 10 · 4 frontier · 6 personal (p. 11–12) - Human Control Requirements: 9 (p. 13–14) - Operational Guidelines: 7 (p. 16–20) - Operational Defaults: 5 (p. 21–23) - Evaluation behaviors: 15 (p. 28) - Illustrative examples: 9 (p. 29–36) - Draft date: 2026-09-14 (p. 3) --- Read the original: https://microsoft.ai/code-of-conduct/ Reading: https://humanist.cartu.app Part of the Cartu Collection: https://cartu.app Built by T Ngo. Thought partner: Claude Code Opus 5. Shipped with Lovable. September 2026.