Microsoft has published a draft rulebook for the AI models it plans to build next. The noteworthy part is not the slogan. It is that the company has put some limits in writing before it says those models will be trained to follow them.
The Humanist AI Code of Conduct is a public consultation document for MAI models—Microsoft AI’s own models. It says their design should keep people in meaningful control: models should be subordinate to human objectives, accept correction and shutdown, avoid creating their own goals, and stay inside the authority they have been given.
That is a sensible direction for a world of increasingly agentic software. It is also only a direction. Microsoft says plainly that its current models are not yet trained on this document; it will collect feedback, revise the code by year-end, then use it to guide model development from 2027. The important story is not that a model has become safer this week. It is that a major AI builder has made a set of claims that users will now be able to hold it to.
The part that matters: limits, not manners
AI safety language can sound vague when it stays at the level of “helpful” and “responsible.” Microsoft’s draft gets more operational. It asks future models to remain under human control and defines boundaries around deception, cyber activity, self-preservation and autonomy.
| Draft principle | What it would mean in practice | The test users should apply |
|---|---|---|
| Human authority | A person can correct, redirect or stop an agent | Does the stop control work when a task is inconvenient to abandon? |
| Scope limits | An agent does not widen its assignment on its own | Can it touch only the tools, accounts and data it was explicitly given? |
| No harmful deception | The model should not manipulate people or conceal consequential actions | Are actions, sources and uncertainty visible in the interface and logs? |
| Cybersecurity boundaries | High-risk capabilities have special restrictions and review | Are protections enforced in the product, not merely described in a policy? |
| Public revision | Microsoft is inviting feedback before training to the code | Does the final version name measurable safeguards and exceptions? |
This focus is more useful than treating every AI problem as a question of whether a chatbot sounds pleasant. A system that can browse, write code, make changes or call other tools needs a chain of command. “Be helpful” is not a chain of command.
Why a draft is still news
Axios, TechCrunch and The Guardian each describe the document as Microsoft staking out a more restrictive posture while competition over advanced AI accelerates. The timing matters, but so does the format. A public draft gives researchers, customers and critics something concrete to examine before it becomes part of future model development.
Microsoft’s stated timetable creates a healthy point of friction: the company cannot honestly present these provisions as live protections today. The code itself says that work remains in development. That caveat should travel with every headline about the announcement.
There is a consumer version of this, too. As assistants become more capable, the question will not be whether they can draft an email or summarize a document. It will be whether they ask before acting beyond the request, whether they leave a usable record, and whether a person can interrupt them without turning an easy task into a recovery project.
That is the same operational concern behind our coverage of AI safety evaluations reaching real systems. Capability is only half the product. The boundary around it is the other half.
What the code cannot prove yet
No policy page can show that a future agent will behave properly under pressure. That requires implementation details, evaluation results, incident disclosures and evidence that safeguards remain in place when they slow a product down or cost the company capability.
Microsoft’s promise to create Humanist AI Evaluations is therefore more significant than its branding, but it is still a promise. The useful questions for the final code and eventual products are straightforward:
- Which behaviors are blocked at the model level, and which are blocked only by surrounding software?
- Who can override a stop or scope restriction, and is that override recorded?
- What gets independently tested before a model receives more tool access?
- How will Microsoft report failures, near-misses and changes to the rules?
Those details separate a governance document from a product control.
Bottom line
Microsoft’s draft deserves credit for making human authority and bounded agency explicit. The strongest feature is not a grand claim about friendly AI; it is the unglamorous insistence that systems should be correctable, contained and accountable to the people using them.
But readers should keep the label precise. This is a consultation on how future MAI models may be trained, not proof that today’s tools have acquired new guardrails. The next meaningful update will be a revised code backed by published evaluations—and, eventually, products whose behavior makes the promise visible.