Google DeepMind’s Kareem Ayoub: Building a Fixed Fence Around AI
Kareem Ayoub, Vice President of AI Technical Strategy at Google DeepMind, shared insights in an interview with Bloomberg's Tom Mackenzie at HumanX in Amsterdam on September 23, 2026.
AI Governance: A Bank-like Approach
Ayoub argued that companies can't fully govern AI systems but can implement fixed rules around them, much like banks did in the late 1980s and early 1990s with trading algorithms. This approach balances exponential AI capabilities with linear governance.
Dividing the AI Governance Problem
He divided the AI governance problem into three parts:
- Role of AI in R&D
- Transparency of systems
- Control over AI
Google's Frontier Safety Framework, which establishes capability thresholds and mitigations, is a step towards addressing these issues, according to Ayoub.
The Gap Between AI Potential and Business Use
Ayoub highlighted a discrepancy between what AI technology can achieve and how businesses utilize it. While AI excels in tasks like speeding up drug discovery, translating ancient texts, and detecting gravitational waves at the frontier, inside companies, its applications often revolve around mundane tasks like drafting or summarizing email.
The Rise of Agents
The shift from models to agents as products began around the end of 2025. An agent enhances a model by providing memory, tools, and internet access, enabling it to handle complex, multi-step tasks. The strongest applications so far have been in software engineering, he noted.
Successful AI Implementation in Companies
Ayoub identified four habits among the best companies he works with:
- Rapidly building safe sandboxes instead of waiting for lengthy IT projects.
- Not betting against model progress.
- Measuring what matters.
- Leading by example by using the AI tools themselves.
Failure Modes in AI Implementation
Companies falter when they reward employees for hiding their AI usage and when leaders only superficially utilize the available tools. Ayoub expressed hope that, as AI intelligence becomes cheaper, more resources will be allocated to checking output quality rather than producing more output.
DeepMind's Mood and Future Directions
When asked about morale at DeepMind, Ayoub declined to add much beyond what has already been reported. He emphasized AGI as Demis Hassabis's life mission and highlighted the energy surrounding Google's Koray Kavukcuoglu-led Gemini program.
DeepMind recently launched the DeepMind Institute to explore AGI and foster discussions on the topic. Former DeepMind researchers are also founding their own ventures, such as Emulate, a world-model startup.
"The future is no longer a forecasting challenge; it becomes a design problem," Ayoub concluded.