Artificial intelligence steps in to optimize decision-making in various contexts. For example, an autonomous system might accurately identify a power distribution strategy that curtails costs and enhances voltage stability. But does such a technically optimal output denote fairness? What’s next if the lowest cost power dissemination strategy results in more vulnerability for disadvantaged neighborhoods to suffer power outages than their high-income counterparts?
This is where MIT researchers step in to develop an automated method of evaluation that harmonizes cost efficiency with social justice. This innovative method aids in the swift identification of potential ethical dilemmas before implementing AI-powered solutions. It factors in both measurable outcomes such as cost-effectiveness and reliability, and qualitative values such as fairness that are more subjective in nature.
The system distinguishes objective evaluations from user-defined human values. To achieve this, it employs a large language model as a representative of human values to capture and include stakeholder preferences. The adaptive framework shortlists the most suitable scenarios for a thorough evaluation, which saves time, effort, and money involved in the manual selection of scenarios. These test-cases can depict situations wherein autonomous systems align well with human values, as well as instances that fall short of the ethical expectations set.
Distinguished associate professor at MIT, Chuchu Fan, and her team have contributed extensively to this development project. Their research will be on display at the upcoming International Conference on Learning Representations.
A power grid is a prime example of a system where an autonomous AI model’s ethical alignment is beyond challenging. Traditionally, testing frameworks depend heavily on pre-gathered data, which often lacks ethical context. As ethical standards and AI systems are constantly evolving, traditional evaluation methods need frequent updates and revisions. To address this, MIT researchers developed an experimental design framework— Scalable Experimental Design for System-level Ethical Testing (SEED-SET), to identify the most informative scenarios.
For different user groups, such as a rural community or a data center, the priority of ethical considerations fluctuates even if their core requirements are the same. SEED-SET address this by splitting the problem hierarchically into objective and subjective evaluations. Their method begins by forming an objective model that considers how the system performs on tangible metrics such as cost. A subjective model that takes into account stakeholder judgements like perceived fairness is then built on this base.
SEED-SET, with its exceptional ability to adapt to multiple objectives without any pre-existing evaluation data, could, for example, highlight instances where power distribution prioritizes higher-income areas during peak demand periods, leaving underprivileged neighborhoods more prone to outages.
In practical tests of their system, researchers used SEED-SET to evaluate realistic autonomous systems, such as an AI-driven power grid and an urban traffic routing system. The result? The system generated over twice as many optimal scenarios within the same timeframe compared to baseline strategies and revealed many scenarios that were overlooked by other methods.
“As the user preferences switched, the list of situations SEED-SET unveiled changed drastically. This indicates that the evaluation strategy adapts to user preferences effectively,” says Anjali Parashar, the lead researcher.
Looking to the future, the researchers hope to conduct user studies to assess the practical benefits of SEED-SET in real decision-making situations. They are also keen to explore the use of more efficient models that can handle larger problems with increased criteria such as the testing of Large Language Model decision-making.
This groundbreaking research was funded partly by the U.S. Defense Advanced Research Projects Agency. Looking to incorporate AI automation into your business operations? Visit implementi.ai to discover cutting-edge tailored solutions designed for your business requirements.
To get more insights on this research, read the original news article here.
This website uses cookies.