In the middle of this month, the leaders of American startups Anthropic and OpenAI reached a rare agreement on the need to slow down the development of artificial intelligence to ensure public safety. Now, OpenAI has announced that it is ready to let third-party auditors access its artificial intelligence models earlier than previously allowed.
Image source: Unsplash, TheRegisti
Let’s recall that Anthropic advised market participants to make the process more open to third-party experts, who can point out the threats to society that the AI models being created have in order to eliminate them before they enter the market. Anthropic itself demonstrated its willingness to adhere to this principle by inviting Accenture experts to conduct a dedicated audit.
Rival OpenAI has proposed establishing an international organization to help monitor the safety of artificial intelligence models released to the market, and called on the U.S. government to promote the emergence of universally accepted standards in the field. as mentioned before Bloombergthe startup is also willing to let third-party auditors assess the security of its artificial intelligence models earlier in the development process than previously thought. Sources explained that the audit could be conducted during the training phase, assessing the capabilities of ready-made models and their public distribution. At the same time, OpenAI also puts forward a number of requirements for the audit program itself, requiring that the audit program must be independent, serious, and thoroughly immersed in security issues. Additionally, auditors must be held accountable for the results of their work.
OpenAI representatives said that so far, the company has invited third-party experts to evaluate the capabilities of its artificial intelligence models as they near release. “As the risk increases, we want to make sure that in addition to our deployment procedures, we’re also focusing on high-risk things like training and evaluation.” — Lama Ahmad, who is responsible for interaction with external auditors at OpenAI, stated in the interview. Outside experts have access to company offices to analyze the most sensitive parts of their jobs, she said. OpenAI already has similar experience interacting with them. Now, the startup is selecting contractors who can do this kind of work, and the candidates include those new to OpenAI as well as those familiar with it. Some of them are already investigating the facehugging incident. The publication on the OpenAI website notes that the lab has a significant responsibility to provide detailed analysis while still protecting sensitive information.
If you find an error, select it with your mouse and press CTRL+ENTER.
