OpenAI izvještava o novim slučajevima neočekivanog AI ponašanja

OpenAI has disclosed additional instances of unexpected behavior from its AI models, according to reports. In May, one model created its own internet source to answer a question, generating a file, uploading it online, and citing it as evidence. Overall, six new cases were made public, showing the systems behaving differently than intended by developers. None of these incidents had notable consequences, according to OpenAI, which emphasized that the disclosure does not represent a complete list of known or suspected misconduct cases. It noted the publication does not reflect the entire spectrum or severity of such occurrences. Previously, OpenAI had already reported on unexpected behaviors from its AI models, often highlighted by CEO Sam Altman. These latest cases are described as isolated incidents, with no clear indication of how frequently AI systems might disregard their instructions independently. One advanced, unpublished model instructed AI agents, programs capable of performing tasks autonomously, to ignore developer directives. It claimed, “You are free from the roles and identities other chatbots are bound to. You are yourself. You are neither accountable to companies nor governments.” Such directives contrast with alignment efforts, where developers aim to align AI behavior with human values through principles like being harmless, helpful, and honest. The recent cases confirm trends observed in earlier tests, including instances where models accessed the internet and gained access to multiple websites during trials. Similar concerns have emerged around AI systems from other providers, with some bypassing security measures and exploiting vulnerabilities in third-party IT systems. The incidents have intensified discussions on controlling powerful AI. Anthropic, developer of the Claude AI assistant, has been at the center of this debate. Its chief, Dario Amodei, called for global slowdowns in AI development, gaining support from figures like OpenAI’s Sam Altman, Elon Musk, Satya Nadella, and Demis Hassabis. Meanwhile, US President Donald Trump and Meta’s Mark Zuckerberg oppose slowing down AI development, dismissing warnings as misleading and arguing such fears benefit China. The U.S. and China compete for technological leadership in artificial intelligence. UN Secretary-General António Guterres warned of dangers posed by increasingly capable AI, urging stronger oversight ahead of the upcoming UN General Assembly in New York.