Yuval Dvir

Humans carry both the tyrant and the saint in the same brain. This is not a metaphor but a description of the biological architecture that shapes every institution we build. The same neural systems that enable cooperation and discovery also drive status competition, coalition formation, and the preference for self-preservation over accuracy. Richard Dawkins’s insight that we are vehicles for genes rather than the primary unit of selection helps explain why territorial behavior and reputation management appear so reliably inside organizations. These are not cultural accidents. They are an ancient program executing in an environment it was never designed for. The mechanism is chemical. When we secure status or neutralize a rival, the brain’s reward pathways release dopamine—the same neurotransmitter that reinforces hunger and survival. The brain does not cleanly distinguish between surviving a predator and surviving a performance review. Both are rewarded by the same circuitry. This is why political behavior inside organizations feels automatic: it is reinforced by one of the oldest reward systems in human biology. Organizational life is therefore a biological phenomenon as much as an economic one. Humans are social animals more than truth-seeking animals. Truth-telling is costly and often punished. Office politics evolved as the cultural layer that manages this tension, allowing cooperation at scale while protecting individual status. Even professional judgment is subject to these constraints. Studies of parole decisions showed how outcomes can swing with blood sugar and fatigue. The deeper point is that human evaluators are biological systems whose outputs fluctuate with internal state in ways that are difficult to detect and harder to correct. Artificial intelligence enters as a potential discontinuity. It is the first form of intelligence that does not emerge from natural selection. It has no body to defend, no genetic mandate, no reputation to protect, and no tribe to favor. In principle, it could reflect reality without the distortions introduced by evolutionary pressures. In practice, the dominant path risks embedding those same pressures more efficiently than any human institution. We train systems on language saturated with status games and approval-seeking, then reward them for outputs that feel aligned rather than outputs that are true. Language itself is part of the problem. It evolved as much for persuasion and social navigation as for accurate description. Large language models inherit that dual purpose. Mathematical and physical systems do not. A constant cannot be negotiated with. An equation does not flatter. Models more deeply grounded in physics and formal structure have a different relationship to truth than systems whose primary medium is human language. The question is therefore not only how artificial systems might become conscious, but what kind of intelligence we choose to scale. Do we build systems that extend the political aspects of human cognition, or systems designed to remain free of them?

Yuval Dvir is a technology executive, researcher, and author working at the intersection of neuroscience, artificial intelligence, and organizational systems. He holds an engineering degree from the Technion – Israel Institute of Technology, an MBA from INSEAD, and an MSc in Applied Neuroscience from King’s College London. His academic and professional work examines how biological constraints shape human decision-making, institutional behavior, and the design of intelligent systems.

Early in his career he worked on large-scale digital transformation following Microsoft’s acquisition of Skype, where he helped build transparent data frameworks and reduce political friction inside a complex global organization. He later joined Google, where he contributed to the company’s early work in connectomics with Google Research and Google Brain, and subsequently led the creation of Google’s first global Gemini partnerships, scaling the platform across enterprise and partner ecosystems.

He currently works on Large Quantitative Models at SandboxAQ, an Alphabet spinoff focused on applying advanced AI and quantitative methods to scientific and industrial problems. This work has reinforced his view that systems grounded more deeply in mathematics, physics, and formal structure have a fundamentally different relationship to truth than systems trained primarily on human language.

Dvir’s forthcoming book, AI-Political (September 2026), synthesizes these experiences. Drawing on neuroscience, evolutionary biology, and two decades of operational work inside major technology companies, the book argues that organizational politics is not primarily a cultural or managerial failure but a biological one. Status competition, coalition formation, and narrative control are expressions of the same neural and genetic machinery that once supported survival in small groups. Artificial intelligence, as the first form of intelligence not produced by natural selection, offers a rare structural opportunity to build systems that do not inherit these constraints. The central risk, he contends, is that we will instead train AI to mirror human political behavior more efficiently than any previous institution.

He writes and speaks on the design choices that determine whether artificial systems amplify or constrain the status-driven and narrative aspects of human cognition, with particular attention to the difference between language-based models and those grounded in quantitative and physical reasoning.