Microsoft AI Chief Sounds A Warning: 'You Can’t Steer Something You Can’t Control'

Mustafa Suleyman, chief executive of Microsoft AI, has issued one of the clearest warnings yet against the breakneck global push toward artificial general intelligence, arguing that the industry risks losing control of AI systems before it learns how to govern them. In a recent post on X, Suleyman cautioned that discussions around “AI alignment” have become misplaced.

The alignment is being pushed, he says, without addressing a more fundamental challenge: containment. "You can’t steer something you can’t control," he wrote, warning that alignment without enforceable limits is little more than 'asking nicely.'

Suleyman drew a sharp distinction between the two ideas, pushing back against a tendency within the industry to treat them as interchangeable. Containment, he said, refers to the ability to set boundaries on an AI system, restrict its agency and enforce constraints.

Alignment, by contrast, is about ensuring AI systems reflect human values and act in humanity’s best interests. Without containment, Suleyman argued, alignment efforts rest on shaky ground.

The remarks come as technology companies increasingly frame superintelligence as the next frontier of AI development. While Suleyman stopped short of naming rivals, his comments amount to a critique of what he has previously described as a potentially reckless pursuit of general-purpose systems with open-ended autonomy.

Instead, Suleyman has championed what he calls 'Humanist Superintelligence,' a model focused on domain-specific AI designed to solve concrete problems rather than mimic or exceed human cognition across all tasks.

In a recent essay on the Microsoft AI blog, he described containment and alignment as 'red lines' that should not be crossed — acknowledging in a December interview with Bloomberg that the stance remains 'novel' in the industry.

In an episode of the 'Silicon Valley Girl Podcast,' Suleyman described AI superintelligence as a system capable of self-improvement, setting its own objectives, and operating independently of human control. He emphasized the difficulty of containing such technology or aligning it with human values, calling this scenario 'the anti-goal.'

Suleyman also clarified the distinction between artificial general intelligence (AGI) and superintelligence, noting that AGI represents an intermediate stage before reaching superintelligence.

Under Suleyman’s leadership, Microsoft AI has leaned into applied use cases such as healthcare and clean energy. The company recently developed a medical AI system that achieved 85% accuracy on the New England Journal of Medicine’s complex diagnostic case challenges, compared with roughly 20% accuracy among human doctors.

Suleyman, a former co-founder of DeepMind, argues that such systems deliver superintelligence-level performance within clearly defined domains — without introducing the risks associated with unbounded, autonomous agents. Since joining Microsoft 18 months ago following the acquisition of Inflection AI’s intellectual property and staff, he has been tasked with building AI products capable of rivaling the industry’s best.

Microsoft’s evolving stance has also been shaped by changes to its partnership with OpenAI. Until a revised agreement last October, Microsoft was restricted from pursuing artificial general intelligence independently. Now, while preserving key commercial ties, it allows Microsoft to independently pursue artificial general intelligence (AGI), either on its own or with third parties.

OpenAI will transition into a public benefit corporation, with Microsoft holding an investment valued at about $135 billion, or roughly 27% on an as-converted diluted basis. OpenAI remains Microsoft’s frontier model partner, and Azure continues to have API exclusivity until AGI is formally declared.

Suleyman’s views echo broader signals from Microsoft’s leadership. Earlier this month, chairman and CEO Satya Nadella urged the industry to move away from dismissive labels such as 'AI slop,' instead framing AI as a 'cognitive amplifier' or a modern 'bicycle for the mind,' borrowing from Steve Jobs.

For now, the debate over superintelligence remains largely academic. Tools like ChatGPT have transformed how humans interact with machines, but the technology is still evolving — and so are the rules governing it.