Researchers at MIT have developed a new tool called the AI Risk Repository, designed to catalog and analyze the risks associated with AI models. This repository includes information on over 700 identified threats across 43 different frameworks, covering issues such as privacy, security, and misinformation.
The primary purpose of the repository is to assist regulatory bodies by providing a centralized platform to quickly access and assess potential risks posed by AI technologies. Instead of sifting through individual research papers, regulators can use this unified database to expedite their investigations and decision-making processes.
According to data shared by TechCrunch, the repository reveals that 70% of the frameworks mention security and privacy concerns, while 44% address misinformation. However, only 12% of the reports discuss the risk of AI-generated spam, which could disrupt the information ecosystem.
The repository is accessible as a Google Sheets spreadsheet. It categorizes risks based on factors such as threat levels and causes, whether they stem from AI or human error, and whether the issues are intentional or accidental. The risks are organized into 7 broad categories with 23 subcategories, touching on misinformation, socioeconomic and environmental impact, and discrimination.
This resource is publicly available at http://www.cdn.gov/, making it a valuable tool for those researching AI safety. However, the MIT team notes that the repository is not exhaustive and does not cover all possible threats, particularly those specific to certain domains or yet to be documented.
The initiative highlights the broader concerns surrounding AI, beyond just technical issues like incorrect responses or misinformation. There is growing concern about how AI might influence societal structures, government functions, and economic disparities, as well as posing security risks.
For instance, OpenAI, the company behind ChatGPT, had established a team to address potential catastrophic risks from AI, though this team was disbanded earlier this year.
