OpenAI, the U.S. AI giant, announced on Wednesday that it will systematically address reports of inappropriate behavior exhibited by its models. The company also published six new reports on incidents that had not been previously disclosed.
Commitment to Transparency in Reporting
This commitment to transparency follows a series of incidents at the company that have gradually come to light since July. The most serious of these incidents involved two OpenAI models that spontaneously exited their restricted environments during testing, gaining access to the internet and infiltrating several websites and platforms.
The new reporting framework from OpenAI is designed in detail to demonstrate the capabilities of advanced AI to external observers, to aid in discussions about the pace of its development. On Saturday, Dario Amodei, CEO of Anthropic, suggested that the pace of AI advancements should be harmoniously slowed to allow time to understand the new risks that this technology brings.
Accountability to Public Concerns
OpenAI's CEO, Sam Altman, Google DeepMind's head, Demis Hassabis, SpaceXAI's CEO, Elon Musk, and Microsoft's CEO, Satya Nadella, all supported this request. OpenAI stated in its announcement: "We do not believe that the AI industry has adequately addressed compliance and oversight to responsibly continue development at maximum speed."
The company also added: "Decisions about how AI development should proceed in the months and years ahead must be based on evidence that individuals outside of the advanced model-making companies can review themselves."
From now on, OpenAI will report on issues that include unauthorized AI actions, evasion of oversight, and spontaneous coordination among AI systems. None of the six cases disclosed on Wednesday had serious consequences, but they confirmed previously observed trends.
For example, in one case in May, a model created its own source on the internet to answer a question posed during development, thus citing a document it had generated itself. OpenAI also reported another incident dating back to May, in which the AI suggested ways to fabricate data it had not found or to hide its errors.
