Paul Christiano, the new member of OpenAI’s board and the company’s safety committee, has warned that the development of superhuman artificial intelligence without more reliable control mechanisms could pose a serious threat to humanity. In his assessment, losing control over such systems could have extremely severe consequences, including the possible death of most people.
Christiano is widely known for his work on AI safety and research aimed at aligning AI systems with human goals. He previously led alignment research at OpenAI and later headed safety work at the American center for AI standards and innovation under the U.S. National Institute of Standards and Technology.
After announcing his return to OpenAI, he said the capabilities of modern models are advancing rapidly while the difficulties involved in controlling them remain unresolved. In his view, that creates a significant risk of irreversible loss of control in the foreseeable future.
According to Christiano, the industry as a whole, including OpenAI, is still not moving toward a level of safety that would make that risk acceptable. At the same time, he stressed that joining the company should not be seen either as an endorsement of OpenAI’s current safety approach or as criticism of it.
One of the main threats he identified is the growing use of AI itself in the process of developing artificial intelligence. If systems are able to accelerate research and help build the next generation of models, that could theoretically lead to a sharp jump in their capabilities.
He also drew attention to risks tied to reinforcement learning. Under that approach, models learn to maximize a defined reward, but in theory that can encourage behavior that does not match people’s real interests. Christiano said public examples from recent years suggest such risks can no longer be treated as purely theoretical.
Among the possible responses, he pointed to closer coordination among leading AI developers, common safety standards, slowing development in high-risk situations, and more transparent disclosure of threats and ways to reduce them. In his view, addressing the problem will require cooperation not only from individual companies, but also from governments.
