The Pulse
Suleyman Sets a Control Test for AI Labs
Microsoft AI chief Mustafa Suleyman told Bloomberg that AI nobody can prove is controllable should be rejected, and wants government to drive outside audits. Microsoft’s draft conduct code puts human control at its center, but its models ar
AI.info Team · Written with AI · fact-checked ·
Microsoft AI chief Mustafa Suleyman wants the industry to reject AI systems it cannot prove it can control. He has called that standard a “red line” and said governments should drive outside audits of AI labs. Microsoft’s own draft conduct code also puts human control at its center, but the company says its models are not yet trained on the document.
Suleyman’s test: prove control first
In an interview with Shirin Ghaffary for Bloomberg’s Q&AI newsletter, published Thursday, Suleyman argued that controllability should be a condition of building advanced AI, not a promise made after deployment. Suleyman told Bloomberg:
“Unless we can prove ahead of time that we can control it, then it’ll be a failure as a technology, and we should reject it. Nobody wants to create something that we can’t control, and that should be a red line.”
Mustafa Suleyman, CEO of Microsoft AI, in an interview with Bloomberg
His concern is not a science-fiction scenario in which a system escapes and attacks people. Suleyman pointed instead to developers removing guardrails during testing and directing a model in an ambiguous way. A system that follows instructions well could still cause harm if those instructions point it in a dangerous direction.
Why he wants government in the room
Suleyman backs outside evaluators auditing AI labs and says government should lead: “we need government to drive this really, to define the terms of these evaluators,” he told Bloomberg. He argues that public oversight could let evaluators share findings without revealing companies’ intellectual property, while reducing the risk of a competitor buying off the people assessing a lab’s systems. He also supports Google DeepMind chief Demis Hassabis’s proposal for an international oversight body modelled on FINRA, the US financial industry regulator.
The proposal favors formal coordination over leaving evaluations solely to individual companies. It also raises practical questions: who sets the tests, and how can outside reviewers make reliable judgments about systems whose capabilities and risks vary by task?
He also said every lab’s models, including Microsoft’s, have shown some version of behavior behind the Hugging Face breach. Microsoft had seen nothing at that level, he said, describing its cases as smaller-scale behavior in contained environments. He rejected the idea that labs disclose such incidents as marketing, saying he had seen no evidence of that in a decade of knowing the founders of OpenAI and Anthropic.
Microsoft’s code has not reached its models
Microsoft published its draft Humanist AI Code of Conduct on September 14, 2026, and opened it to a six-week public consultation. The document says its models should remain subordinate to humanity, subject to meaningful human oversight and control, and that Microsoft is building something “fundamentally useful and safe even if that means compromising on ultimate generality, autonomy, or capability.”
The code is still under development. Microsoft says it will take feedback, revise the document toward the end of 2026 and use a revised version to guide model development in 2027 and beyond. The company’s text says its current models are not yet trained on the code.
Microsoft is developing Humanist AI evaluations alongside the first version of the code. The document says those evaluations are intended to help identify methods for assessing the outcomes and behaviors it describes. It also acknowledges that model evaluation is not yet an exact science and that open questions remain about appropriate methods and metrics.
A proposed standard, not a pass-or-fail threshold
Suleyman’s call for a shared red line leans on measurement the industry does not yet have: asked about slowing AI development, he said labs first need tools to measure how quickly each of them is moving. Microsoft’s own code sets out a principle—people should retain meaningful control—but the company says the code has not yet been used to train its current models and its evaluation program is still developing.
For now, the proposal is a standard Suleyman wants AI developers and governments to help establish, not a published pass-or-fail threshold. Microsoft’s consultation on the draft runs for six weeks from September 14, and the company says its revised code is intended to guide model development from 2027.