Understanding and addressing the risks associated with artificial intelligence (AI) is crucial for individuals, companies, and governments. However, identifying and categorizing these risks is a complex task. To assist policymakers and stakeholders in the AI industry and academia, MIT researchers have developed an AI “risk repository” that aims to provide a comprehensive and up-to-date database of AI risks.
The AI risk repository, curated by MIT’s FutureTech group, contains over 700 risks grouped by causal factors, domains, and subdomains. The researchers created the repository to gain insights into the overlaps and disconnects in AI safety research. While other risk frameworks exist, they cover only a fraction of the risks identified in the repository. This lack of comprehensive coverage could have significant consequences for AI development, usage, and policymaking.
By working with colleagues from the University of Queensland, the nonprofit Future of Life Institute, KU Leuven, and AI startup Harmony Intelligence, the MIT researchers scoured academic databases to retrieve thousands of documents related to AI risk evaluations. Through their analysis, they found that existing frameworks often prioritize certain risks over others. For example, privacy and security implications were included in over 70% of the frameworks, while misinformation was covered by only 44%. Similarly, discrimination and misrepresentation were discussed by more than 50%, while the pollution of the information ecosystem received attention from only 12%.
The AI risk repository offers a foundation for researchers, policymakers, and risk professionals to build upon. It saves time and increases oversight by providing a comprehensive database of AI risks. Instead of reviewing scattered literature or relying on limited frameworks, stakeholders can access the repository to gain a more holistic understanding of AI risks.
However, the mere existence of a risk repository does not guarantee effective regulation of AI. AI regulation worldwide is currently fragmented and lacks unified goals. While the repository can help align stakeholders on the risks AI poses, it does not address the limitations of safety evaluations for AI systems.
Nevertheless, the MIT researchers plan to utilize the risk repository to evaluate how well different AI risks are being addressed in the next phase of their research. By identifying shortcomings in organizational responses, they aim to raise awareness of overlooked risks and promote more comprehensive approaches to AI regulation.
In conclusion, the AI risk repository developed by MIT serves as a valuable resource for understanding and managing the risks associated with AI. It provides a comprehensive database of over 700 risks, helping stakeholders save time and increase oversight. While the repository alone cannot solve the challenges of regulating AI, it can contribute to more informed decision-making and highlight areas that require attention. By evaluating how well different risks are being addressed, the MIT researchers aim to drive improvements in AI safety practices.