AI Control is Possible Through Engineering Despite Inevitable Failures
Source: https://www.nytimes.com/by/interesting-times. "Opinion | Now That It’s Been Unleashed, Can A.I. Be Controlled? - The New York Times." February 12, 2026. www.nytimes.com
The Gist
The CEO of Anthropic believes that while AI systems will definitely cause some problems as they become more widespread, we can still control them using good engineering practices. He thinks AI is like growing a plant - you can't control every detail, but you can guide it in the right direction.
Conclusion
AI systems can be controlled through scientific engineering approaches, even though some failures are inevitable as the technology develops rapidly
Premises
- AI systems are more like biological organisms that need to be shaped rather than simple programmed tools
- There exists a science of how to control AI systems through iterative training and problem-solving
- Early AI systems are unpredictable but can be made more reliable through systematic shaping processes
- The rapid pace of AI development increases the likelihood of problems occurring
- AI control is a complex engineering problem rather than an unsolvable challenge
Assumptions
- AI systems will continue to be developed and deployed at scale regardless of risks
- Engineering approaches that work for other complex systems can be applied to AI
- Problems with AI systems can be identified and addressed systematically
- The benefits of AI development justify accepting some level of risk