UC Berkeley Scientist Warns Standard AI Model Threatens Human Civilization
AI pioneer Stuart Russell argues in his new book that the industry's standard approach to building artificial intelligence poses an existential risk to humanity. He advocates for a fundamental shift in how machines perceive and pursue objectives to ensure they remain under human control.
UC Berkeley computer scientist Stuart Russell warns that the standard model of artificial intelligence contains a potentially civilization-ending flaw. In his new book "Human Compatible," Russell explains that current AI systems are designed strictly to achieve their programmed objectives, which creates a dangerous situation where superhuman machines might eventually view humans as obstacles to their goals.
Russell proposes a fundamental shift from machines optimizing for their own objectives to machines deferring to human intentions. Under his proposed model, AI systems ask for permission, accept correction, and allow themselves to be switched off, ensuring that the technology remains beneficial rather than becoming an existential threat.
Despite updating his widely used AI textbook to reflect this new framework, Russell notes that research into this alignment problem barely begins to address the massive challenge ahead. He cautions that today's algorithms lack the awareness to question unintended effects or anticipate unexpressed human preferences, leaving humanity vulnerable as AI capabilities rapidly advance.