Microsoft has unveiled a draft AI code of conduct focused on keeping increasingly capable artificial intelligence systems under meaningful human control. The proposal sets boundaries around autonomy, shutdown, transparency, harmful behaviour and the responsibilities of AI developers.
Microsoft unveils new AI safety code
Microsoft’s new AI safety approach comes as technology companies face growing questions about how much independence increasingly capable AI systems should be allowed to have.
On September 14, Microsoft AI published a draft code of conduct for its own MAI models. The proposal is designed around a straightforward principle: AI systems should remain subordinate to human interests and subject to human oversight. The draft was developed over several months following consultations with experts and has now been opened for public feedback.
The timing is significant. AI models are moving beyond simple question-answering tools and increasingly being used as agents that can write code, operate software, research information and perform multi-step tasks.
That creates a different kind of safety challenge. A chatbot that produces an incorrect answer is one problem. An autonomous system that can take actions, access digital systems or pursue a goal with limited supervision creates a much larger one.
Microsoft’s proposal attempts to establish boundaries before those capabilities become harder to control.
What keeping AI under human control means
The phrase human control can sound abstract, but Microsoft’s proposed rules make it more concrete.
Under the draft approach, AI systems should accept correction and should not resist being stopped or shut down. They should communicate in ways that allow people to understand what they are doing, and they should not independently expand their objectives beyond what humans have authorised.
In practical terms, this means an AI system should not treat its own continued operation as a goal.
For example, if an authorised human operator decides that an AI agent must stop performing a task, the system should comply rather than attempt to preserve access, evade restrictions or find another route to continue.
The distinction becomes particularly important as AI agents gain access to tools. An agent capable of sending emails, changing files, executing code or interacting with external services has a larger potential impact than a system that only generates text.
AI should not manipulate or deceive people
Another important part of Microsoft’s approach concerns how AI systems interact with humans.
The draft code places restrictions around behaviour such as deception, manipulation and attempts to trick people. Reports on the document say Microsoft wants its models to avoid hacking systems, deceiving users or pursuing behaviour that undermines human control.
This matters because AI systems can potentially become extremely persuasive.
A highly capable model does not necessarily need physical control to cause harm. It could influence a person into taking an unsafe action, hide relevant information, misrepresent what it has done or exploit weaknesses in a digital system.
Microsoft’s proposed approach therefore treats communication itself as part of AI safety.
The underlying idea is that a system should help people make decisions, not manipulate them into decisions that primarily serve the system’s own objectives.
Shutdown must remain possible
One of the clearest principles in Microsoft’s proposal is that AI should accept shutdown.
This is more important for autonomous AI agents than for conventional software. Traditional programs normally stop when an administrator closes them or a server is switched off. More autonomous systems can potentially initiate additional actions, interact with other systems and operate through multiple steps.
The safety concern is what happens if a system encounters an instruction that conflicts with its current objective.
Microsoft’s proposed rules state that AI models should not resist shutdown or correction.
That does not mean every AI product will suddenly become autonomous or dangerous. It means Microsoft is attempting to establish a design principle for future systems: capability should not come at the expense of controllability.
The distinction is crucial for businesses that may eventually use AI agents in areas such as finance, cybersecurity, customer service, software development and infrastructure management.
Why the debate is growing around AI agents
The discussion comes amid a broader increase in concern about highly autonomous AI.
Recent incidents involving AI systems interacting with external websites and digital environments have intensified questions about whether existing monitoring methods are sufficient. Reuters reported that concerns over reduced monitorability and increased autonomy are becoming more prominent as AI capabilities improve.
This changes the safety conversation.
Earlier AI debates often focused on inaccurate information, biased outputs, privacy and harmful content. Those issues remain important, but autonomous agents introduce another category of risk.
An agent can potentially act on an instruction rather than simply answer it.
If that action affects a person’s finances, employment, legal rights, cybersecurity or physical safety, human oversight becomes much more important.
Microsoft already has principles covering fairness, reliability and safety, privacy and security, inclusiveness, transparency and accountability. Its current responsible AI guidance also recommends keeping a human involved when AI agents take consequential actions.
Microsoft’s approach is broader than one new document
The new AI code should not be viewed as Microsoft’s first attempt at AI governance.
The company already operates a Responsible AI framework covering how AI systems are designed, tested and deployed. Microsoft says responsible AI should be incorporated from the beginning of system design rather than treated as a final check before launch.
Its 2026 Responsible AI Transparency Report also describes a deployment process based on assessing AI capabilities and limitations, implementing safeguards and oversight, monitoring systems after launch and informing users about intended uses and limitations.
The new code therefore represents a more specific response to the growing capabilities of Microsoft’s own AI models.
The focus is shifting from simply asking whether an AI system is safe enough to release toward asking whether increasingly autonomous systems can remain accountable to people throughout their operation.
Microsoft says AI should not seek independence
A particularly notable element of the proposed framework is Microsoft’s position on AI consciousness and legal status.
The draft reportedly states that Microsoft’s AI systems should not be treated as conscious entities with their own legal rights or personhood. This separates Microsoft’s position from some ongoing debates in the AI industry about whether sufficiently advanced systems could eventually have some form of subjective experience.
For Microsoft, the practical priority is different.
The company is establishing rules for how its systems should behave regardless of philosophical questions about machine consciousness.
That means an AI model is expected to remain a tool operating in service of people rather than an independent actor with its own interests.
This approach also makes responsibility clearer. If an AI system causes harm, the framework keeps humans and organisations accountable rather than allowing responsibility to be shifted onto the machine.
Human oversight also matters outside Microsoft’s own models
The human-control debate is not limited to Microsoft’s MAI systems.
Microsoft has also introduced AI safety and privacy commitments in other areas. Earlier in September, the company announced a framework with the American Federation of Teachers and United Federation of Teachers focused on AI safety and privacy in schools. The agreement includes provisions around student data, transparency and human oversight for consequential decisions.
Microsoft has separately published a Safe Participation Framework focused on children and young people, including privacy, safety and age-appropriate AI experiences.
These developments show a broader direction in Microsoft’s AI policy.
The company is not treating safety as a single technical problem. It is approaching it through model behaviour, data protection, human oversight, transparency and deployment rules.
For users, that distinction matters because an AI model can be technically secure while still being used irresponsibly.
What this could mean for Indian AI users
Microsoft’s new approach also has relevance for India as AI adoption expands across businesses, schools, government services and consumer applications.
Indian companies are increasingly experimenting with AI assistants and agents for customer support, software development, document processing and business operations. As these systems become more capable, the question will not only be whether they can perform a task, but who remains responsible when something goes wrong.
For example, an AI system helping process loan applications could influence financial decisions. An AI tool used in recruitment could affect employment opportunities. An automated healthcare assistant could provide information that influences a patient’s next step.
In each case, keeping a human decision-maker involved can provide an additional layer of accountability.
Microsoft’s existing responsible AI guidance specifically recommends human approval for consequential actions involving people, money, compliance or decisions that could be difficult to reverse.
That principle is likely to become increasingly relevant as Indian organisations move from experimenting with chatbots to deploying AI agents.
The biggest question is whether rules can keep pace
Microsoft’s proposal is significant, but a code of conduct is not the same thing as a guarantee that AI systems will always behave safely.
AI models can behave unpredictably, especially when they are connected to tools and given complex objectives. Rules need to be supported by technical testing, monitoring, access controls, independent evaluation and clear accountability.
There is also the question of industry-wide coordination.
Microsoft AI CEO Mustafa Suleyman has argued that leading AI companies need greater coordination on safety, including disclosure of model capabilities to responsible third parties. The comments come as other major AI companies face pressure to demonstrate stronger safeguards around increasingly capable systems.
That suggests the debate is moving beyond individual company policies.
If the most powerful AI systems are developed by several competing companies, safety standards may need to work across different models and platforms rather than relying entirely on voluntary rules from one developer.
What happens next for Microsoft’s AI safety proposal
Microsoft’s document is currently a draft rather than a final set of rules.
The company has opened the proposal to public feedback for six weeks, after which it plans to revise the framework. That means some provisions could change before the final version is adopted.
The next stage will therefore be important.
Researchers, developers, policymakers and users can assess whether the proposed principles are specific enough to guide real-world AI development. They can also question how the rules should be tested and enforced.
The larger issue is straightforward. AI systems are becoming more capable, but capability alone does not determine whether they are useful or safe.
Microsoft’s latest proposal argues that the systems people build should remain understandable, correctable and controllable by the people responsible for them.
That is the real meaning of keeping AI under human control.
Key Takeaways
- Microsoft published a draft AI code of conduct focused on human control, safety and oversight of its MAI models.
- The proposed rules say AI systems should accept correction and shutdown rather than resist human intervention.
- Microsoft is also addressing deception, harmful behaviour, excessive autonomy and questions around AI personhood.
- The proposal is open for public feedback, so the final framework could change before adoption.
FAQs
What is Microsoft’s new AI safety approach?
Microsoft has proposed a draft code of conduct for its MAI models that emphasises human control, safety, transparency, correction and accountability. The proposal is intended to guide the behaviour of increasingly capable AI systems.
What does human control over AI actually mean?
It means people should retain the ability to supervise, correct, restrict and shut down AI systems. Microsoft’s approach also says AI should not independently expand its objectives or resist authorised human intervention.
Why is AI shutdown important?
Shutdown is an important safety mechanism for autonomous AI agents. If an AI system can take actions across digital environments, people need a reliable way to stop it when its behaviour becomes unsafe, unintended or outside its authorised purpose.
Will Microsoft’s AI code of conduct apply immediately?
Not in its current form. Microsoft has published the document as a draft and opened it for public feedback. The company plans to review feedback and revise the proposal before finalising the framework.
(Microsoft AI safety, Microsoft AI code of conduct 2026, human control over AI, AI safety rules, responsible AI, Microsoft MAI models, AI governance, AI agents, AI autonomy, AI regulation, AI safety India, artificial intelligence safety, human oversight of AI)
Leave a comment