TLDR: This paper introduces a framework using LLMs to assess and explain AI risks from various stakeholder perspectives. It demonstrates how different stakeholders perceive risks differently in AI systems (e.g., medical AI, autonomous vehicles, fraud detection) and uses rule-based explanations to highlight conflicts. An interactive visualization tool is proposed to enhance transparency in understanding these disagreements, emphasizing the importance of stakeholder-aware explanations for responsible AI governance.
In the rapidly evolving landscape of artificial intelligence, particularly with the widespread adoption of large language models (LLMs), ensuring responsible deployment is paramount. A new research paper, “Who Sees the Risk? Stakeholder Conflicts and Explanatory Policies in LLM-based Risk Assessment,” delves into a crucial aspect often overlooked: how different individuals and groups perceive risks in AI systems. This work, authored by Srishti Yadav, Jasmina Gajcin, Erik Miehling, and Elizabeth Daly, introduces a novel framework that uses LLMs to act as “judges” to predict and explain AI risks from various stakeholder perspectives.
The core idea behind this research is that risk perception isn’t universal. A doctor, a patient, a regulator, or a developer might view the same AI system and its potential risks very differently. Traditional AI risk assessment methods often fall short by being “stakeholder-agnostic,” meaning they don’t account for these diverse viewpoints. This paper addresses this gap by proposing a stakeholder-grounded approach, making AI governance more context-sensitive and inclusive.
The framework leverages LLMs to generate stakeholder-specific, interpretable policies. This means the LLMs are prompted to assess risks not just generally, but from the viewpoint of a particular persona – for example, a surgeon using an AI medical diagnosis assistant, or a passenger in an autonomous vehicle. By doing this, the system can identify not only what risks exist but also how and why different stakeholders might agree or disagree on those risks.
To achieve this, the researchers utilized the IBM AI Risk Atlas Nexus for identifying potential risks and the GloVE explanation method to generate rule-based explanations. These explanations use “IF” and “DESPITE” clauses to articulate why a risk might be relevant for a specific stakeholder. For instance, one stakeholder might say, “IF the decision is always reviewed by a human, then it’s not a risk,” while another might argue, “DESPITE human oversight, it’s still a risk.” Identifying these contrasting rationales is key to understanding conflicts.
The methodology was demonstrated across three real-world AI use cases: medical AI, autonomous vehicles, and fraud detection. The results clearly showed that stakeholder perspectives significantly influence how risks are perceived and the patterns of conflict that emerge. For example, fraud analysts, family members, patients, and transportation regulators often showed higher proportions of risk-labeled predictions, indicating their direct exposure or impact from AI decisions.
A significant contribution of this paper is the introduction of an interactive visualization tool. This tool helps to reveal how and why conflicts emerge across stakeholder perspectives, enhancing transparency in conflict reasoning. Imagine a visual representation where different stakeholders are color-coded bubbles, and lines connect them to show areas of agreement or disagreement on specific risks. Users can then click on a stakeholder to see the detailed “IF” and “DESPITE” explanations that led to their assessment, highlighting the clauses causing the conflict.
Also Read:
- Rethinking AI Oversight: Why Healthcare Needs Capability-Based Monitoring for Large Language Models
- Enhancing AI Agent Safety and Trustworthiness with a New Response Framework
This research emphasizes the critical need for stakeholder-aware explanations to make LLM-based evaluations more transparent, interpretable, and aligned with human-centered AI governance goals. While the current approach uses synthetically generated stakeholders, the framework lays a strong foundation for integrating real-world feedback and more nuanced risk categories in the future. For more in-depth information, you can read the full research paper here.
