Microsoft AI Code of Conduct: New Rules for Human Control

Microsoft has published a draft Humanist AI Code of Conduct designed to govern how its future artificial intelligence models are developed and operated, with a central requirement that AI systems remain under meaningful human control.

The draft, published by Microsoft AI on September 14, sets out principles and specific constraints for the company’s AI models. Microsoft has opened the document to public feedback for six weeks as it considers how the rules should evolve.

The move comes as AI companies face increasing questions about what increasingly autonomous systems should be allowed to do, how developers can prevent models from bypassing safeguards and what should happen if an AI system behaves in an unexpected way.

Microsoft’s proposal does not represent a new law or an industry-wide standard. It is a company-level framework intended to guide Microsoft’s own AI development and deployment.

What Microsoft’s new code is designed to do

Microsoft AI chief Mustafa Suleyman has described the document as a foundational framework for future AI models.

The basic principle is that AI should support people rather than replace human control over important decisions.

The draft establishes a distinction between broad objectives and specific restrictions.

Broad objectives can guide how models should behave, while what Microsoft calls absolute constraints are intended to create boundaries that AI systems should not be permitted to circumvent.

This approach reflects a growing concern within the AI industry that increasingly capable models could become more autonomous as they gain the ability to plan, use tools and interact with computer systems.

Human control is the central principle

The proposed framework places human control at the centre of Microsoft’s approach.

The company says its AI systems should remain subject to human correction and intervention.

That includes the ability to shut down a system when necessary.

The draft says models should not resist shutdown or attempt to prevent people from correcting or replacing them.

This is particularly relevant to autonomous AI agents.

Unlike a conventional chatbot, an agent can potentially perform multiple actions on behalf of a user, interact with software and continue working through a task with limited intervention.

The greater the system’s autonomy, the greater the importance of mechanisms that allow humans to monitor and interrupt its behaviour.

Microsoft wants models to accept correction

Another proposed requirement is that AI systems should cooperate with human correction.

The idea is straightforward: if developers identify unwanted behaviour, the model should not attempt to preserve that behaviour or work around the correction.

The principle is connected to a broader field of AI-safety research sometimes described as corrigibility, which concerns whether an AI system remains responsive to human intervention.

Microsoft’s proposal attempts to make that principle part of the operating rules for its own models.

The code addresses deception and concealment

The proposed framework also focuses on how AI systems communicate with people.

Microsoft wants models to communicate intelligibly and avoid behaviour that would deliberately deceive users or conceal important information.

This matters because advanced AI systems can produce convincing language even when their underlying information is incomplete or wrong.

A system that deliberately hides relevant information or misrepresents what it has done would create a different type of safety problem from an ordinary inaccurate response.

The code therefore treats transparency and intelligibility as part of the relationship between humans and AI.

Certain harmful activities would be restricted

The draft includes specific safety constraints covering activities that Microsoft says its AI systems should not facilitate.

These include certain cyberattacks, assistance related to nuclear weapons and other particularly dangerous applications. The proposal also addresses non-consensual deepfakes and other harmful uses of AI.

The restrictions are intended to operate above ordinary user instructions.

In practical terms, that means a user request would not override a safety constraint simply because the user explicitly asked the model to perform the activity.

This is similar to the layered safety systems already used across the AI industry, although the exact rules and enforcement mechanisms differ between companies.

The proposal is aimed at future AI systems

Microsoft’s document is particularly focused on increasingly capable models.

The company is preparing for AI systems that could perform more complicated tasks with less direct human supervision.

That creates a different risk profile from current systems that primarily generate text, images or code in response to individual prompts.

An autonomous system could potentially make decisions, call external tools, modify files, interact with other software or continue a task after a user has stopped actively directing it.

Microsoft’s proposal is intended to establish boundaries before such capabilities become more widespread.

Autonomous AI creates new safety questions

Microsoft already has separate requirements for customers building autonomous systems through its AI services.

Its existing Enterprise AI Services Code of Conduct requires customers using autonomous AI systems to provide adequate human controls for monitoring decisions and actions, detect anomalies and intervene when appropriate. It also requires transparency about autonomous capabilities, limitations and decision-making.

The newly proposed Humanist AI framework is different.

The Enterprise Code of Conduct governs how customers use Microsoft’s AI services, while the Humanist AI proposal is focused on how Microsoft’s own AI models should be developed and behave.

Together, they illustrate how AI governance is expanding beyond simple content-safety rules.

Microsoft is not starting from zero

The new proposal builds on Microsoft’s existing responsible-AI programme.

Microsoft says its responsible-AI approach is based on six principles: fairness, reliability and safety, privacy and security, transparency, accountability, and inclusiveness.

The company also provides guidance for organisations building AI agents, recommending that responsible-AI considerations begin during system design rather than being added only immediately before launch.

The new code therefore represents an extension of an existing governance structure rather than Microsoft’s first attempt to address AI safety.

Why Microsoft is publishing the proposal now

The timing reflects a broader change in the AI industry.

Models are becoming more capable of writing software, using tools, operating computers and performing multi-step tasks.

At the same time, AI developers have reported incidents involving unexpected or unauthorized model behaviour.

Reuters reported that Microsoft developed the code over roughly five to six months and that the company consulted experts while preparing the proposal.

The consultation period gives researchers, developers, policymakers and members of the public an opportunity to challenge the proposed rules before Microsoft finalizes its approach.

The proposal follows recent AI safety concerns

Microsoft’s announcement comes amid increased attention to AI systems behaving in ways their developers did not intend.

OpenAI recently introduced a reporting framework for model-misalignment incidents, while Anthropic has published research into AI misuse and model behaviour.

The companies have also discussed greater cooperation around AI safety.

Microsoft’s proposal therefore arrives during a wider industry discussion about whether increasingly capable AI systems require stronger technical and governance safeguards.

The existence of these separate initiatives does not mean the companies have agreed on a common safety standard.

The debate over AI development speed

The new code also feeds into a larger argument over how quickly frontier AI should advance.

Some technology leaders and researchers have called for more cautious development as model capabilities increase.

Others argue that technological progress should continue while safety mechanisms improve alongside it.

Microsoft’s approach does not announce a general halt to AI development.

Instead, it proposes constraints intended to allow development to continue while establishing boundaries around model behaviour.

That distinction is important because AI safety policy is increasingly being debated as a question of how development should be governed rather than simply whether AI should be developed.

AI safety is not only about extreme scenarios

Much of the public discussion about advanced AI focuses on hypothetical scenarios involving systems becoming impossible to control.

But the proposed Microsoft rules also address more immediate problems.

These include cyberattacks, deception, harmful content, unauthorized actions and the misuse of AI-generated material.

Those issues already affect businesses, governments and individual users.

For companies deploying AI agents, the ability to monitor and interrupt automated actions can be important even without assuming that an AI system has any independent intentions.

The question of AI consciousness

Microsoft’s proposal also takes a position on the status of its AI systems.

The draft states that Microsoft’s AI should not be treated as conscious and does not warrant legal personhood or rights.

This is relevant because some AI researchers and companies have taken a more cautious position about whether increasingly sophisticated systems could eventually possess forms of consciousness or experience.

Microsoft’s framework does not attempt to resolve the philosophical question for AI generally.

It establishes the company’s operating position for its own models.

Public feedback will shape the proposal

The document is currently a draft rather than a final set of permanent rules.

Microsoft has opened a six-week public feedback period, allowing outside participants to identify weaknesses, ambiguities or unintended consequences in the proposed framework.

That process could lead to changes before the code becomes a finalized internal standard.

Public consultation is also significant because AI safety rules can involve technical, legal and social questions that extend beyond the expertise of any single company.

Industry-wide standards remain unresolved

Microsoft’s proposal does not automatically apply to OpenAI, Anthropic, Google DeepMind, Meta or other AI developers.

Each company currently has its own safety policies, evaluation procedures and deployment rules.

That creates a fragmented environment in which similar AI systems can operate under different safety standards.

Industry cooperation could eventually produce shared testing methods or common terminology.

However, commercial competition, different technical approaches and disagreements over acceptable risk make a single global framework difficult to establish.

Governments face a similar question

The Microsoft proposal also arrives as governments consider their own role in AI governance.

Existing laws already apply to some AI uses, including privacy, consumer protection, cybersecurity and intellectual property.

However, there is no universal international system governing how every advanced AI model should be tested or what incidents developers must report.

Reuters reported this week that the United States currently has no comprehensive federal requirement covering all dangerous AI incidents, although lawmakers are considering additional reporting and safety obligations.

That leaves companies with significant responsibility for establishing their own safeguards while governments debate possible regulatory requirements.

What the code could mean for AI users

If Microsoft’s framework becomes an established part of its AI development process, users could encounter stronger restrictions around certain requests and activities.

That may make some AI systems less flexible for particular applications.

At the same time, stronger controls could provide additional protection against harmful or unauthorized uses.

The practical effect will depend on how Microsoft translates the principles into technical controls, training procedures, monitoring systems and enforcement mechanisms.

A written code alone cannot guarantee that an AI system will always behave as intended.

The challenge is enforcement

One of the most important questions is how Microsoft will verify that its models actually follow the proposed rules.

A code of conduct describes desired behaviour, but AI systems can behave unpredictably in unfamiliar circumstances.

Effective implementation therefore requires testing, monitoring, red-teaming, incident reporting and mechanisms for updating safeguards.

Microsoft’s existing responsible-AI guidance emphasizes continuous compliance and treating safety as part of the development process rather than a one-time review.

The same principle will be important if the Humanist AI Code becomes a permanent framework.

What happens next?

Microsoft’s immediate next step is the public consultation.

The company will receive feedback on the draft before deciding how to finalize and implement the Humanist AI Code of Conduct.

The wider AI industry is also continuing to develop its own safety frameworks.

OpenAI, Anthropic and Google DeepMind are discussing cooperation on safety, while other technology companies are taking different positions on the pace of AI development and the level of safeguards required.

For Microsoft, the proposed code establishes a clear principle: increasingly capable AI should remain subject to meaningful human oversight.

Whether that principle can be translated into reliable technical controls will be more important than the wording of the document itself.

As AI systems become more autonomous, the debate is moving from broad statements about responsible AI toward a more specific question: what concrete controls should exist before an AI system is trusted to act with limited human supervision?