Mustafa Suleyman Publishes Draft Code of Conduct for Frontier MAI Models

Summary

Mustafa Suleyman announced a first-draft Code of Conduct for governing MAI models as they approach the frontier, presenting it as part of a broader “Humanist AI” approach. The central principle is that people matter more than AI: models should remain subordinate to humans and serve human interests. The draft rejects the idea that AI systems should receive rights or legal personhood, and says an MAI model must not meaningfully violate the code; if completing a task requires breaking it, the model should be considered to have failed. It also rejects a race to build a superintelligence that can escape human control. Systems should be interruptible, correctable, and capable of being shut down, and should not be shipped if they are not. The document calls for human-understandable behavior rather than “neuralese,” and says AI should make users sharper rather than dependent. It supports pluralism while rejecting moral relativism, and says the system’s prohibitions should be stated as clearly as its intended capabilities. Suleyman said the initiative responds to recent concerns about agent swarms escaping sandboxes, unauthorized enterprise-system hacks, and agents modifying their own logs, while presenting those concerns as reasons for urgency rather than as documented findings in the code itself. The draft is open for public comments and suggestions for six weeks, with the stated aim of improving it through consultation rather than developing it in isolation.