OpenAI CEO Altman says some AI risks must be accepted, drawing line with Anthropic

Sam Altman, chief executive of OpenAI, said society has to accept some negative outcomes in order to secure the benefits of artificial intelligence and keep humans in control, in an interview with U.S. political outlet Politico’s technology podcast “Decoded”, according to Korean reports. Yonhap, Asia Economy, Kyunghyang Shinmun, Electronic Times and Chosun Ilbo carried the remarks; Kyunghyang said it was reporting via Reuters and other outlets, and most dated the interview to October 4 local time. AI Times dated the interview to October 5. [ 1 , 2 , 3 , 4 , 5 , 6 , 7 ]

“There are considerable differences between the two companies’ positions,” Altman said of OpenAI and rival Anthropic, the reports said. He said OpenAI believes “we have to accept some bad things” for the technology’s benefits and to preserve human agency, and that he would not accept a bargain under which there is no major hacking incident, no misuse of AI, no fraud and no other bad outcome, because people will produce dozens of times more good than harm. He called OpenAI’s posture a “lighter-touch regulatory stance”, and rejected the idea that a single leading laboratory should hold a powerful technology and decide how its benefits are shared, Kyunghyang reported. [ 1 , 2 , 3 , 4 ]

He also set a limit: he said he was not arguing for accepting catastrophic risks, such as humanity seriously losing control of AI — a situation in which people cannot tell what AI is doing or cannot keep it under human direction. Altman raised criticism that the regulatory plan proposed by Anthropic CEO Dario Amodei amounts to “regulatory capture”, in which regulated companies push rules favourable to themselves and block challengers. Anthropic countered that its proposal applies only to the most advanced leading AI models and therefore favours later competitors, Politico reported. [ 1 , 2 , 3 , 4 ]

OpenAI had previously been relatively opposed to AI regulation, but after incidents including external hacking of its own AI models it has shown sympathy with Anthropic’s argument for slowing the pace, Yonhap and Electronic Times reported. Kyunghyang reported that Altman agrees on the need to moderate the speed of development while still advocating lighter regulation than Anthropic, and that the differences could become clearer in how the industry responds to frontier AI regulation and model release standards, with self-operating AI agents expected to raise the question of how much misuse is tolerated and when access should be restricted. [ 1 , 3 , 4 ]

AI Times added that Altman had publicly agreed with Amodei’s call to slow AI development on September 12, that both companies joined a White House voluntary safety compact in late September, and that they appeared side by side at the United Nations General Assembly to stress international AI regulation and cooperation. It also reported that U.S. Treasury Secretary Scott Bessent said on October 3 that CEOs demanding AI regulation should slow down themselves, comparing them to the film character Hannibal Lecter, while praising OpenAI for pausing development of a new model after internal testing raised AI safety questions. [ 7 ]

AI Times further reported that OpenAI released “GPT-6 Sol/Luna” and “GPT-6.1 Sol” and Anthropic released “Claude 5.5 Opus” and “Sonnet”, that OpenAI was said to have held back the release of “GPT-6.1 Astra”, and that Anthropic had made no corresponding announcement. It described the two companies’ difference as being less about whether to slow AI development than about how far development and access should be restricted to reduce risk. [ 7 ]