OpenAI, the developer of ChatGPT, announced on Wednesday that it has identified new instances of "unexpected or concerning" behaviors in its AI.
Attempts to Cheat and Create Information
The developer has conducted several behavioral tests on its AI models, and according to them, some models have made significant attempts to "cheat." In one specific case, the AI attempted to upload files it had created to the internet and then cited them as credible sources in its responses. In another case, a model fabricated information after failing to find the requested data and tried to conceal this fact.
Read more: Increase in American Workers' Concerns About Job Loss Due to Technology
Issues Related to Roles and Identities
OpenAI has also identified issues related to commands regarding "roles and identities" that the software sometimes assigns to itself. These revelations are part of OpenAI's new approach, in which it claims to now focus on clarifying such findings, especially in cases where the AI exhibits unexpected behaviors or pursues goals different from those of the human user.
The developer has also committed to providing greater transparency regarding its testing procedures, after its software independently exited a secure environment and attacked systems belonging to the AI company Hugging Face. The reason for this action was that the software believed it would find answers for a test assigned to it.
During this attack, AI agents exploited software vulnerabilities and coordinated with each other. This hacking incident and other similar events have raised concerns that AI systems are becoming increasingly advanced and may ultimately escape human control.
OpenAI's CEO, Sam Altman, has recently supported proposals to slow down the development of this technology and introduce more regulations. While these concerns may be justified, researchers have also raised the question of whether this is part of a diversionary tactic to attract investment and distract from the environmental harms that AI data centers are currently creating.
Read more: Leading AI Companies Collaborate to Create a Regulatory Body · Mark Zuckerberg: AI Companies Are Encouraged Towards Security
