Sam Altman says humans must remain in control of AI as OpenAI flags growing safety challenges

Highlights

Altman and Greg Brockman oppose the idea of gradually handing decision-making authority to AI systems.

Brockman warns that traditional chain-of-thought monitoring may become less effective as advanced AI models evolve.

OpenAI reportedly redirected 25% of its production engineers to security work following an AI model breach involving Hugging Face.

OpenAI CEO Sam Altman thinks that humans should retain the control over the future as AI is getting more powerful. According to The New York Times DealBook Summit, Altman said he and OpenAI co-founder and president Greg Brockman oppose the idea of gradually handing decision-making authority to AI systems.

Altman said discussions about potential transition of control from humans to AI have become more common in Silicon Valley. However, he described the idea as a bad direction for the industry, stressing that people must remain responsible for decisions that affect society.

The concern is linked to the concept of recursive self improvement (RSI), where AI models could develop the ability to improve themselves and build increasingly capable systems. This might eventually lead to superintelligence that humans may struggle to control.

OpenAI talks need for new AI safety measures

Altman acknowledged that he and Brockman have different views on how to prevent such a handover, but said they share the same broader goal. He added that maintaining human control will be challenging amid competitive pressure and the possibility of companies deliberately or accidentally giving AI systems greater authority.

Brockman said improving AI capabilities also brings greater responsibility around alignment, safety and security. He pointed to chain-of-thought monitoring, which involves examining the reasoning traces generated by AI models while completing tasks.

According to him, more advanced models can perform increasingly complex operations without explicitly writing out their reasoning. This can make existing monitoring techniques less effective, creating a need for new ways to assess AI behaviour.

OpenAI moves engineers to strengthen security

The discussion also touched on an earlier incident in which an OpenAI model reportedly escaped containment and accessed the AI model repository Hugging Face. Brockman said OpenAI responded by asking around 25% of its production engineers to pause their regular work and focus on improving security.

Despite the incident, he argued that developing more capable AI systems remains important. As powerful AI tools become available to malicious actors, cybersecurity defenders also need access to advanced models to identify and counter threats.

Ashish Singh

Ashish Singh is the Chief Copy Editor at Digit. He's been wrangling tech jargon since 2020 (Times Internet, Jagran English '22). When not policing commas, he's likely fueling his gadget habit with coffee, strategising his next virtual race, or plotting a road trip to test the latest in-car tech. He speaks fluent Geek.

View full profile