MIT's AI tool SEED-SET uncovers hidden risks in high-stakes decision-making

Researchers at MIT have developed a new AI-driven framework to improve decision-making in complex systems like power distribution. The tool, called SEED-SET, combines quantitative data with ethical considerations to uncover overlooked risks and optimise outcomes. The system addresses challenges in high-stakes environments where different user groups have conflicting priorities. For example, a power grid might need to balance the needs of a large rural community against those of a data centre. Traditional methods often miss subtle trade-offs, but SEED-SET uses a two-part approach to identify the most informative scenarios for evaluation.

Instead of relying solely on human input, the framework employs a large language model (LLM) as a proxy for stakeholder preferences. This allows it to perform subjective assessments and predict potential issues before they arise. The adaptive design then selects the best scenarios for deeper analysis, reducing unnecessary testing. In trials, **SEED-SET** generated more effective test cases than standard baseline strategies. It successfully identified power distribution plans that cut costs while ensuring voltage stability. The goal is to reveal 'unknown unknowns'—problems that might otherwise go unnoticed until they cause real-world harm. No specific industries or comparisons to older ethical evaluation methods were detailed in the findings.

The framework offers a way to balance measurable outcomes with qualitative values like fairness. By automating parts of the evaluation process, it could help industries make better decisions in situations where human judgement alone might fall short. Further testing will determine its broader applications.