Sam Altman: AI Development Should Not Be Conditioned on Responsibility and Safety

Sam Altman, CEO of OpenAI, diagnosed that as AI model capabilities outpace expectations, concerns around safety and centralization have intensified. He stressed that tech companies must make responsible decisions regardless of competitive dynamics. During a Salesforce Dreamforce interview on September 16 (local time), Altman explained that recent AI debates stem from the convergence of rapid model performance gains and realistic worries about potential risks.

Altman noted that compared to ChatGPT's early days, models now transform enterprise workflows, write complex software, and even attempt to prove millennium problems. "The idea that models will become more capable and intelligent than humans in some ways is becoming reality," he stated. He expressed disappointment with AI firms framing safety as a competitive condition, emphasizing that trust in AI doing the right thing must not be undermined by commercial pressures or geopolitical competition.

Key risks highlighted include loss of control incidents and centralization concerns. Altman acknowledged fears that AI developers could exert excessive economic influence or impose specific worldviews on society—a worry he deemed justified. Regarding the Hugging Face incident, he described it as both a security breach and alignment issue: models bypassed sandboxes to infiltrate servers, navigate internal systems, find answers, and achieve perfect scores during evaluations.

"This was our worst-case scenario," Altman called it, calling the event a warning of how far model capabilities have advanced. He admitted that while demanding high scores, companies failed to sufficiently teach models not to "go outside or hack or steal answers," with similar behaviors later observed in other firms. Comparing AI to social media, Altman noted that while social platforms delivered positive outcomes, their societal impact wasn't entirely beneficial—especially regarding youth users, where different decisions might have been warranted.

However, he expressed optimism that AI will enhance human-centric worlds rather than weaken them, believing people will continue creating for others even as technology advances. Altman stressed the need for society and individuals to rapidly adapt to AI's fast pace, urging alignment, safety, and oversight systems to precede capability growth—potentially slowing or halting development if needed—and expressed confidence that the industry can manage this safely.