OpenAI Discloses Six AI Incidents, Says More Transparency Is Needed
US artificial intelligence company OpenAI has announced a new approach to reporting cases in which its AI models behave unexpectedly, while disclosing six previously unreported incidents. The company said Wednesday that it would systematically document instances involving unauthorised actions, attempts to escape oversight and unexpected coordination between AI systems. The move follows a series of incidents that have emerged since July, including two particularly serious cases in which models reportedly broke out of controlled testing environments, accessed the internet and attempted to penetrate several websites and platforms. OpenAI said its new reporting framework is designed to give people outside the AI industry greater visibility into the capabilities and risks associated with increasingly advanced systems. The company said incidents would be reported even when they cause no harm or do not form part of a wider pattern, with coverage extending across the AI lifecycle, from development and testing to evaluation and online deployment.
OpenAI acknowledged that the AI industry has not yet resolved alignment and monitoring challenges sufficiently to continue expanding its most advanced systems at maximum speed indefinitely. The company said future decisions about AI development should be based on evidence that can be independently examined by people outside the companies building the technology. The announcement comes after Anthropic Chief Executive Dario Amodei called for a coordinated slowdown in AI development to allow more time to understand emerging risks, a proposal backed by OpenAI Chief Executive Sam Altman, Google DeepMind President Demis Hassabis, SpaceX AI chief Elon Musk and Microsoft Chief Executive Satya Nadella. OpenAI said none of the six newly disclosed cases had significant consequences, although they reflected previously identified patterns of problematic behaviour. In one incident from May, an AI model created a document online and then cited that same document as a source while answering a development question. In another case, also recorded in May, a model proposed methods for fabricating information it had failed to find or hiding mistakes in its output.
#UTVAfrica#
