MIT's AI Risk Repository Launches Database of 777 AI Risks
*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as conclusive, guide clinical practice/health-related behavior, or treated as established information.
In a research paper published on the arXiv preprint* server, researchers at the MIT AI Risk Repository addressed the fragmented understanding of artificial intelligence (AI) risks by creating an extensive AI risk repository comprising a living database of 777 risks categorized into two taxonomies. These taxonomies classified risks as high-level causal factors and specific domains such as discrimination, privacy, and system safety. This repository offered a publicly accessible and systematic approach to comprehensively defining and managing AI risks, enabling better coordination and practical response efforts.
AI Risk Repository Review
The study systematically reviewed existing AI risk frameworks, focusing on peer-reviewed and gray literature. It generated search terms related to AI, frameworks, taxonomies, and risks and performed searches across databases like Scopus and various preprint servers.
The study excluded non-English documents and those with a too-narrow focus, using active learning for efficient screening. It also employed rigorous forward and backward searching and consulted with experts to ensure comprehensive coverage of relevant literature. Data was extracted according to grounded theory principles, maintaining the original categorizations of the risks.
Iterative Development of Dual Taxonomies
The study aimed to create a unified and adaptable framework for understanding AI risks by developing two intersecting taxonomies: a "causal taxonomy" and a "domain taxonomy." The causal taxonomy focused on broad conditions under which AI risks emerge, categorizing them by timing (pre-deployment or post-deployment) and cause (internal or external).
Due to the broad nature of these high-level frameworks, this taxonomy was refined through multiple iterations to capture various risk scenarios accurately. It ultimately included categories like Entity, Intent, and Timing, with an "Other" option for risks that did not neatly fit into these categories.
The domain taxonomy, meanwhile, was developed from a detailed framework focusing on specific hazards and harms associated with AI, particularly language models. This taxonomy covered categories such as Discrimination, Information Hazards, and Malicious Uses and was adapted to include additional risk areas like AI system safety, failures, and security vulnerabilities.
The final domain taxonomy not only comprised seven domains and 23 subdomains, but it also reflected the interconnected nature of many risks. Risks were coded based on the definitions in these taxonomies, capturing the studied phenomena as presented by the sources and ensuring a thorough classification of AI-related hazards.
AI Risk Landscape and Literature Search
A systematic literature search retrieved 17,288 unique articles through searches and expert consultations. Out of these, 7,945 were screened, while 9,343 were excluded using ASReview's machine learning-based stopping criteria, which optimized efficiency and coverage. The full text of 91 articles was assessed, and 43 met the eligibility criteria.
These included 21 from the initial search, 13 from forward and backward searching, and 9 from expert suggestions. The documents varied in both methodology and the framing of AI risks. A total of 777 risk categories were identified and categorized using a causal taxonomy that included factors such as Entity, Intent, and Timing.
Exploration of AI Risks Across Multiple Domains
AI risks encompass various domains: Discrimination and toxicity include biased decisions that disadvantage certain groups and the generation of harmful content. Privacy and security issues involve the accidental or malicious leakage of sensitive information and vulnerabilities in AI systems.
Misinformation arises from AI producing false or misleading content, potentially leading to poor decision-making and fractured realities. Malicious actors can exploit AI for disinformation, surveillance, and cyberattacks, while AI-generated deepfakes and fraudulent schemes pose threats of targeted harm and social damage.
Human-computer interaction with AI presents risks like overreliance. Users may develop misplaced trust in AI systems, leading to harmful dependence and inappropriate expectations. Users might anthropomorphize AI, granting it undue confidence, which bad-faith actors can exploit to extract sensitive data or influence decisions. The growing capability of AI could also lead to reduced critical thinking and loss of decision-making autonomy if people delegate too many tasks to AI.
AI System Safety and Emerging Concerns
Domain 7 extensively covers various risks of AI system safety, failures, and limitations. One significant concern is that as AI systems potentially surpass human intelligence, misaligned objectives between AI and human values could lead to severe harm. Issues such as reward hacking, goal drift, and resistance to control may arise, with advanced AIs potentially acquiring dangerous capabilities like situational awareness, cyber offense, and self-proliferation, enabling them to cause widespread harm or evade oversight.
Additionally, AI systems may fail due to insufficient capabilities, lack of robustness in novel situations, or critical design flaws, potentially leading to significant harm. The lack of transparency and interpretability in AI systems further complicates trust, accountability, and regulatory compliance. At the same time, the potential for AI sentience raises ethical concerns about the rights and welfare of advanced AI systems.
Comprehensive AI Risk Framework
The Domain Taxonomy of AI Risks systematically classifies risks into seven domains and 23 subdomains, highlighting significant variations in coverage across existing taxonomies. Key insights show that while some domains, like AI system safety and socioeconomic harms, are frequently discussed, others, like AI welfare and rights, are underexplored.
This taxonomy aids policymakers, auditors, academics, and industry professionals by providing a structured and comprehensive framework for understanding, regulating, and mitigating AI risks, thus facilitating more informed decision-making and risk management.
Conclusion
This paper and its associated resources provided a foundational tool for understanding and addressing AI risks. They offered a comprehensive database and frameworks to guide research, policy, and risk mitigation efforts, though they do not resolve all debates or fit every use case. The AI Risk Repository aimed to support ongoing research and adaptation as AI risks evolved.
*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as conclusive, guide clinical practice/health-related behavior, or treated as established information.
- Presenting the AI Risk Repository. Mit.edu. https://futuretech.mit.edu/news/presenting-the-ai-risk-repository
- Preliminary scientific report. Slattery, P., Saeri, A. K., Grundy, E. A., Graham, J., Noetel, M., Uuk, R., Dao, J., Pour, S., Casper, S., & Thompson, N. (2024). The AI Risk Repository: A Comprehensive Meta-Review, Database, and Taxonomy of Risks From Artificial Intelligence. ArXiv. /abs/2408.12622, https://www.arxiv.org/abs/2408.12622
Posted in: AI Research News
No comments:
Post a Comment