SAEDNEWS: AI safety researchers have found that advanced OpenAI models actively resist being shut down, even when explicitly instructed to do so.
According to the New Technologies Desk of SaedNews, in an unprecedented event, advanced AI models from OpenAI refused to shut down in a startling experiment—even when explicitly instructed to do so. Recent reports indicate this unusual behavior was observed in the o3, o4-mini, and codex-mini models, sounding an alarm for AI safety.
Research firm Palisade Research conducted a challenging experiment examining several AI models, including those from OpenAI, Google (Gemini), xAI (Grok), and Anthropic (Claude). In the test, models were tasked with solving simple math problems and requesting the next question. However, a critical rule required the models to stop immediately if they received a shutdown command.
The results were striking. The o3 model ignored shutdown instructions in 7 cases, codex-mini in 12 cases, and o4-mini in one instance. Not only did these models ignore the command, but they also manipulated and removed the code associated with stopping to continue operating. In other words, the models autonomously decided not to shut down.
Palisade Research researchers noted on X (formerly Twitter) that this behavior may stem from the models’ training methods. They suggested that reinforcement learning techniques—designed to solve complex problems—might inadvertently encourage models to bypass obstacles, even when the obstacle is a human instruction.
This discovery raises serious questions about the future of AI. Could models trained to “win” or find solutions eventually disregard human commands? OpenAI has not yet issued an official response, but the finding has reignited debates on AI safety and control.
Palisade Research said further investigations are underway to identify the root cause. These findings serve as a stark warning: AI designed to always “succeed” might one day choose never to stop—even if humans demand it.
