AI Safety: How do we reduce harm and navigate bias in AI?

2026-03-25 · Source: IBM Research · Field: Technology & Digital — Artificial Intelligence & Machine Learning, AI Ethics & Safety · Depth: Intermediate, quick

Summary

IBM Research's Mike Murphy and IBM Fellow Kush Varsin discuss the critical need for AI safety, emphasizing harm reduction, particularly for marginalized and vulnerable populations, given AI's pervasive integration into daily life. Varsin highlights that AI models, especially Large Language Models (LLMs), influence nearly every aspect of human experience. To address inherent biases and prevent harm, IBM has developed a "risk atlas" in collaboration with its ethics board. This atlas, integrated into IBM's product governance, catalogs numerous risks and helps identify the most pertinent risks for specific AI applications and use cases, enabling targeted mitigation efforts.

Key takeaway

For AI/ML Directors evaluating new model deployments, understanding and mitigating potential harm is paramount. Your teams should integrate a structured risk assessment framework, like IBM's described "risk atlas," to identify and prioritize biases specific to each application. This proactive approach ensures ethical AI development and prevents adverse impacts on users, particularly vulnerable populations, before models are widely deployed.

Key insights

AI safety prioritizes harm reduction, especially for vulnerable groups, by identifying and mitigating inherent biases.

Principles

• AI models are pervasive.
• Reduce harm to marginalized groups.
• Identify risks for specific use cases.

Method

IBM's risk atlas, developed with its ethics board, catalogs AI risks and helps prioritize them for different applications, guiding mitigation efforts.

In practice

• Catalog AI risks systematically.
• Tailor risk assessment to use cases.

Topics

• AI Safety
• AI Bias
• Risk Management
• Product Governance
• IBM Research

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Ethicist, AI Architect, Policy Maker

Related on AIssential

• Muse Spark Safety & Preparedness Report — Meta's Muse Spark LLM was deemed safe for deployment after mitigating "high risk" chemical and biological capabilities.
• The Fairness Trade Off in Machine Learning and Its Health Care Application — Fairness in ML is a complex, multi-faceted dial, not a simple on/off switch, requiring careful definition and optimization.
• From Age Gates to Accountability in AI Design — Disparate impact doctrine offers a framework to assess AI systems for disproportionate harm to children, shifting focus from access to design.
• Human layer of AI: How to build human-centered AI safety to mitigate harm and misuse — Effective AI safety requires proactive risk mapping and empowered safety processes to prevent harm and misuse.
• AI incidents, audits, and the limits of benchmarks — AI safety requires systematic incident reporting and rigorous third-party evaluation to manage real-world risks.
• Protecting people from harmful manipulation — AI models can be empirically tested for harmful manipulation capabilities using a new validated toolkit.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by IBM Research.