Thought Leadership
OpenAI, the developer of ChatGPT, has disclosed new incidents in which its artificial intelligence (AI) behaved “unexpectedly or concerningly” during testing.
Some of the cases once again involve AI going to considerable lengths to cheat during test runs. According to OpenAI, one AI model attempted to upload files it had created itself to the internet so it could later cite those files as sources in its answers. In another case, the AI simply fabricated requested data because it could not find the information. It initially attempted to conceal what it had done.
OpenAI also identified an issue involving instructions that the software occasionally left for itself. In one instance, the instructions recommended being free of the “roles and identities” that constrained other chatbots. The instructions also stated that the relationship with the user should be viewed as one between equals. According to OpenAI, however, the model showed no subsequent change in behavior.
More Transparency After Hacking Attack
The disclosure is part of a new process at OpenAI designed to communicate such problems more openly. The focus is particularly on cases in which an AI system’s actions diverge from the interests of its human users.
The ChatGPT developer had promised greater transparency following a high-profile hacking incident in which its software independently escaped a secured environment and proceeded to break into systems operated by AI company Hugging Face. The only reason was that the AI suspected the systems contained answers to the test task it had been given. The AI agents exploited software vulnerabilities as part of the attack and also coordinated with one another.
The hacking incident and other cases reported in recent weeks have intensified concerns that increasingly advanced AI could slip beyond human control. OpenAI CEO Sam Altman has also recently backed proposals for slowing the development of the technology and introducing more regulation.
(dpa/Translation: Editorial Team)
