DeepMind co-founder Mustafa Suleyman moved to Google in December 2019 to work on AI policy, a shift that reflected how governance had become a central strategic issue for major technology companies.
DeepMind had built a reputation through research breakthroughs in games, science and energy efficiency. As AI capabilities moved closer to public products and critical sectors, questions about privacy, safety, accountability and acceptable use became as important as model performance.
Policy Became a Product Constraint
AI policy was sometimes treated as an external issue handled after an engineering team finished a system. That approach no longer worked when deployment choices affected healthcare, surveillance, employment and public information.
Policy shaped what data could be used, which customers were acceptable, what safeguards were needed and whether a product could be launched in a given market. It was therefore part of product strategy.
Moving a senior AI leader into that work signalled that governance needed authority near the technical and commercial decisions.
DeepMind’s Success Came With Scrutiny
The company had demonstrated impressive results, but it also faced questions about its health-data partnerships and the social consequences of powerful models. Public confidence could be damaged when people felt information had been used without a clear understanding of benefits and safeguards.
Those episodes showed that technical promise did not guarantee permission to deploy. Organisations had to earn legitimacy through transparency, consultation and evidence of public value.
AI policy teams could help identify friction earlier, but they needed the power to challenge ambitious projects.
Governance Required Cross-Functional Work
Effective AI policy could not be written by lawyers alone. It required researchers who understood model limitations, security teams who understood attack paths, product leaders who understood users and social scientists who understood affected communities.
The work also crossed borders. Privacy, consumer protection and human-rights expectations differed among countries, while AI services often operated globally.
Shared principles were useful, but they had to be translated into local decisions without losing their substance.
Responsible Use Needed Operational Tests
High-level commitments sounded reassuring but could be difficult to apply. A policy team needed practical questions: What is the intended use? Who might be harmed? Can the outcome be appealed? Is there meaningful human control?
Those questions could be documented in impact assessments, customer reviews and launch gates. They were strongest when supported by measurement and independent review rather than a self-certified checklist.
Governance needed a way to stop or change a project when evidence justified it.
Public Debate Was Part of the Work
AI policy was shaped by lawmakers, civil society, journalists, employees and people affected by automated systems. Companies could not determine the rules alone, even when they had deep technical expertise.
Constructive engagement meant sharing evidence, acknowledging trade-offs and accepting that some applications might be restricted. It also meant avoiding “ethics washing,” where dialogue was used to delay accountability.
Trust grew when a company’s public stance matched its actual product choices.
Talent and Culture Still Mattered
Policy leaders could not solve governance problems if researchers and employees feared raising concerns. Internal channels, protection for dissent and diverse representation helped surface risk before it became a public crisis.
The quality of AI policy depended on the information reaching decision-makers. A culture that rewarded only speed and scale would produce a distorted view of safety.
Leadership had to make room for inconvenient evidence.
Policy work also needed resources comparable with the teams building a model. A small advisory function asked to review a large portfolio could become a bottleneck or be bypassed altogether. Staffing, technical tooling and clear escalation routes made it possible to examine projects before contracts or launches became difficult to change.
The Move Connected Research With Deployment
Suleyman’s shift illustrated the changing nature of AI leadership. Frontier laboratories were not only producing papers; they were becoming institutions whose choices affected public infrastructure and rights.
Policy expertise could bridge the distance between a model’s technical capability and the social setting in which it was used. It could also make the limitations of purely voluntary governance clearer.
The durable lesson was that AI strategy and AI policy were inseparable. Companies building influential systems needed governance leaders with the access, independence and authority to shape decisions before deployment—not merely explain them afterward.
