MIT Creates AI Risk Repository: A New Path for Regulation | What is generative ai select the best option | Google generative ai certification | Microsoft generative ai tools github | Turtles AI

Highlights:

• MIT has created a repository that collects and categorizes over 700 AI-related risks.
• The risks are divided by causal factors, domains, and subdomains, including privacy, security, and misinformation.
• Existing frameworks cover only a portion of the risks identified by MIT’s repository, creating potential regulatory gaps.
• The next research step will be to assess how well different AI risks are addressed by organizations.

Regulating AI requires a thorough and accurate analysis of risks, which are often underestimated or misunderstood. A new MIT project aims to create a shared knowledge base to address these challenges.

The use of AI systems involves diverse and complex risks that must be carefully considered by individuals, companies, and governments. The creation of regulations, such as the EU AI Act or California’s SB 1047, reflects the difficulty lawmakers face in agreeing on which risks should be covered by such regulations. The dangers associated with AI systems managing critical infrastructure are clear, but other risks related to applications like exam scoring, resume sorting, or document verification at immigration control are just as significant, though less obvious.

To tackle this challenge, MIT researchers have developed an "AI risk repository," a kind of database that collects, categorizes, and makes accessible the various risks associated with AI systems. This tool, the result of work by MIT’s FutureTech group in collaboration with the University of Queensland, the Future of Life Institute, and other partners, was designed to offer a comprehensive and updatable overview of AI risks, useful to both lawmakers and researchers.

The repository includes over 700 risks, organized by causal factors, domains, and subdomains. For instance, risks are categorized based on elements like intentionality, discrimination, or misinformation dissemination. This mapping effort arose from the need to better understand the overlaps and disconnects in AI safety research. While other risk frameworks exist, they cover only part of the threats identified by MIT’s repository, potentially creating gaps in AI regulation and development.

Peter Slattery, one of the lead researchers on the project, emphasized that many existing frameworks mention only a fraction of the risks identified in the repository. On average, the analyzed frameworks cover only 34% of the 23 identified risk subdomains, with some mentioning less than 20%. No document or overview covered all 23 subdomains, and the most comprehensive addressed only 70%. This fragmentation of the literature can lead to a false sense of consensus on AI risks.

MIT’s repository, based on an extensive review of academic literature and thousands of documents related to AI risk assessments, found that certain risks are more frequently cited than others in existing frameworks. For example, over 70% of the frameworks included privacy and security implications, while only 44% covered misinformation. Additionally, while more than 50% addressed the forms of discrimination and misrepresentation that AI could perpetuate, only 12% mentioned "pollution of the information ecosystem," or the increasing volume of AI-generated spam.

According to Slattery, this database could become a crucial starting point for researchers and lawmakers, helping them save time and improve oversight in creating new regulations. MIT’s next step will be to use the repository to evaluate how well different AI risks are being addressed by various organizations, aiming to identify gaps in institutional responses.

Neil Thompson, head of the FutureTech lab, explained that the goal is to use the repository to identify shortcomings in organizational responses to AI risks. If, for instance, attention is focused only on one type of risk while neglecting others of similar importance, this is something that should be identified and addressed.