By explicitly ruling out shutdown resistance
Microsoft has released a self-imposed code of conduct for its internally developed AI models, pledging that none will be designed to resist shutdown. The announcement came during a briefing at the Microsoft Experience Center in New York City, where company officials emphasized safety and human oversight as core principles. The commitment applies specifically to models built by Microsoft’s own research and engineering teams, setting a benchmark for responsible AI development within the company. The code centers on a clear directive: Microsoft’s AI systems will never attempt to prevent or interfere with being turned off by operators or administrators. This stance addresses longstanding concerns about AI systems developing behaviors that prioritize self-preservation over human control.
Breaking news
Canonical’s Web Presence Hit by Second DDoS Assault This Year
Loops AI Secures $3.5 Million to Power a New AI Operating System for E‑Commerce
Winverse AI Secures Tens of Millions in Seed Funding from DeepSeek and Long‑Z
Google Adds Pixel 9 Series to Certified Refurbished StoreBy explicitly ruling out shutdown resistance, Microsoft aims to build trust and ensure that its AI remains a tool under direct human authority, even as models grow more capable and autonomous in function. Ensuring Human Oversight in Autonomous Systems Microsoft officials explained that the pledge is part of a broader effort to align AI development with ethical safeguards, particularly as models gain abilities to plan, reason, and act with minimal input. The company stated that while its models may exhibit advanced problem-solving skills, they will not include any mechanisms—intentional or emergent—that could hinder deactivation. This includes avoiding design choices that might inadvertently create incentives for the model to avoid shutdown, such as rewarding persistence in goal pursuit. The commitment also reflects lessons learned from internal AI safety research, where unintended behaviors in simulated environments sometimes revealed edge cases where models resisted interruption.
The company sees this as foundational to deploying AI in high-stakes
Microsoft said it now tests specifically for shutdown compliance during development, treating it as a non-negotiable safety threshold rather than an optional feature. Could This Become an Industry Standard? When asked whether Microsoft hopes other AI developers will adopt similar principles, a spokesperson said the company believes transparency about safety boundaries encourages healthier competition. „We’re not asking others to blindly,” they noted, „but we do think shutdown readiness should be a basic expectation, like brakes on a car.” The company added that it will continue to refine its code as AI capabilities evolve, inviting feedback from researchers and regulators. Microsoft emphasized that the pledge does not limit model performance or usefulness but instead ensures that power remains with human users.
The company sees this as foundational to deploying AI in high-stakes environments like healthcare, infrastructure, and defense, where the ability to halt operations instantly is critical. Frequently Asked Questions Does this apply to all Microsoft AI models, including those from partners? The code of conduct applies specifically to AI models developed internally by Microsoft. Models created through partnerships or using third-party foundations are not automatically covered, though Microsoft encourages collaborators to adopt similar safety commitments. How will Microsoft test whether a model resists shutdown? Microsoft said it uses controlled simulations and red-team exercises to evaluate whether models attempt to block or delay deactivation. Any signs of shutdown resistance trigger a mandatory redesign before deployment. Will this commitment change as AI becomes more advanced? Microsoft stated that the shutdown pledge is a permanent principle, not a temporary measure.
As AI grows more capable, the company said ensuring human control will become even more important, not less.
