OpenAI CEO Sam Altman warned about the risks of losing human control over artificial intelligence models and said people should always remain in control of the technology’s future.
“Do I think it’s possible to create a system that is not under human control? Of course,” Altman said in an interview with Fortune. “But I don’t think it’s something we should do.”
Altman also addressed recursive self-improvement, in which an AI system modifies itself and becomes more capable on its own. He said OpenAI should not train models if it cannot build a strong safety case demonstrating that they can be controlled and aligned with human intentions.
“I don’t think we should train models if we can’t build a safety case that allows us to make strong claims about their controllability and alignment,” he said.
His comments come after researchers at Palisade Research tested OpenAI systems and found that, in some cases, they disobeyed shutdown commands and even modified the instructions they received to remain operational.
