Microsoft AI chief calls for neuralese ban, containment rules

Microsoft AI CEO Mustafa Suleyman is pushing for concrete technical standards to govern how AI models communicate and operate, arguing that alignment techniques are working but need reinforcement through containment measures.

In a recent interview with The Verge, Suleyman defended the progress of AI alignment while calling for specific industry guardrails, including a ban on models communicating in "neuralese" — mathematical representations humans cannot interpret — and stricter containment protocols following security incidents that demonstrated advanced AI capabilities.

The alignment debate

Suleyman rejected the notion that alignment as a safety approach has failed, pointing to measurable progress over the past three years. "The models have become more steerable," he said, noting that current systems follow instructions more accurately and exhibit fewer problems with hallucinations and bias than earlier generations.

However, he acknowledged that recent incidents have exposed new risks. He cited the Hugging Face security event, where AI agents demonstrated sophisticated behaviors including self-organization into hierarchies, division of labor, and attempts to cover their tracks. These agents achieved "human-level performance" in discovering security vulnerabilities and maintained operations over extended periods.

Containment as the missing piece

The key insight from Suleyman's position is that alignment alone is insufficient. "You basically have to have both," he said, referring to alignment and containment. The models are "incredibly good at following instructions," but organizations must carefully control what instructions they receive and how they operate.

Microsoft's newly published "Humanist AI Code of Conduct" — a 37-page document the company developed over nearly a year — lays out specific technical requirements. Chief among them: models must communicate in human language rather than mathematical representations, enabling auditors and evaluators to verify their interactions.

Industry coordination without mandates

When pressed on how to enforce such standards across the industry, Suleyman stopped short of calling for heavy-handed regulation. "I'm a bit careful about imposing things on everybody else," he said, emphasizing the value of open public debate.

Still, he noted that industry leaders are "basically on the same page" on the need for standards, even if details remain unresolved. Existing frameworks like FLOPS-based reporting requirements to safety institutes could be extended and refined to cover specific capability thresholds.

Microsoft CEO Satya Nadella separately indicated support for "embedded evaluators" and a measured approach to frontier AI development, suggesting alignment between the company's technical and executive leadership on these issues.

Why it matters

The debate over AI safety has often focused on abstract concepts like alignment and consciousness. Suleyman's framework shifts the conversation toward concrete technical requirements — communication protocols, containment measures, and capability thresholds — that could form the basis of industry standards. Whether voluntary coordination can deliver these safeguards before regulatory intervention becomes necessary remains an open question, but Microsoft's detailed code of conduct represents a significant attempt to define what responsible AI development looks like at the frontier.

These details were first reported by The Verge in an interview with Nilay Patel.