Microsoft’s AI Model Conduct: A Blueprint for Safety
Microsoft has recently published a Code of Conduct for its Artificial Intelligence (AI) models, outlining a set of principles and restrictions to ensure ethical development and deployment. The document is open for public consultation until late October, marking a significant step towards transparency and accountability in the AI industry.
Key Commitments:
- Human Oversight: Microsoft emphasizes that their AI models will always be susceptible to human intervention, correction, or shutdown.
- Behavioral Guidelines: They include three behavioral commitments: models will not expand their scope, assume new goals without human direction, or conceal their reasoning from auditors.
- Absolute Constraints: This section highlights critical areas where AI models should have strict limitations, such as avoiding weapons of mass harm, ensuring child safety, and preventing large-scale manipulation.
- Humanist AI: Microsoft defines their approach as "Humanist AI," prioritizing people over AI and treating it as a tool that should never replace human control.
Models in Focus:
The code covers five models: MAI-Transcribe-2, MAI-Thinking-1, MAI-Code-1.1-Flash, MAI-Image-2.6, and MAI-Voice-2, showcasing the rapid advancement of in-house AI capabilities.
Enterprise Configurability:
An intriguing aspect is that enterprise partners can configure model behavior, but Microsoft stresses that they do not intend to enforce a singular AI vision on users. This raises questions about how absolute constraints will be enforced and interpreted in diverse real-world applications.
Public Consultation and Future Steps:
The consultation period invites feedback from all, with specific prompts for input on model values, human flourishing definitions, evaluation criteria, multi-agent interactions, and safety constraints prioritization. Microsoft plans to publish a summary of the feedback and a revised code by the end of the year.
Timely Release:
Microsoft’s move comes amid ongoing debates in the AI community regarding speed versus safety. Dario Amodei and Sam Altman have both expressed the need for caution, while Donald Trump has dismissed the idea of slowing down. Microsoft adopts a balanced stance, offering a clear set of guidelines and inviting discussion on their implementation.
Unanswered Questions:
The document leaves certain critical questions unanswered, including enforcement mechanisms, penalty structures for constraint breaches, and decision-making processes regarding model behavior. Additionally, the code is anonymous, without attributed authorship.
This consultation marks a crucial step in shaping the future of AI development, where transparency, safety, and human oversight are at the forefront.