For decades, the evaluation of artificial intelligence has been centered around the comparison of human and machine performance. However, this approach has been criticized for its limitations in capturing the true capabilities and risks of AI systems in real-world environments.

Current benchmarking methods, which focus on isolated tasks with clear right or wrong answers, do not account for the complex interactions between AI systems, human teams, and organizational workflows. This misalignment can lead to misunderstandings of AI’s capabilities, overlooked systemic risks, and misjudged economic and social consequences.

To address this issue, researchers propose shifting from narrow methods to benchmarks that assess AI systems’ performance over longer time horizons within human teams, workflows, and organizations. This approach, known as HAIC benchmarks (Human–AI, Context-Specific Evaluation), aims to provide a more comprehensive understanding of AI’s real-world performance.

A study of real-world AI deployment in various organizations, including small businesses, healthcare, and non-profit institutions, highlights the need for a more nuanced evaluation approach. The study reveals that even AI models with impressive technical scores can struggle in real-world environments, where they must interact with human users, existing systems, and regulatory requirements.

For instance, FDA-approved AI models for reading medical scans may achieve high accuracy in controlled tests but require additional time and effort to interpret their outputs in the context of hospital-specific reporting standards and regulatory requirements.

By adopting a more human-centric approach to AI evaluation, we can better understand the strengths and limitations of AI systems and ensure their safe and effective deployment in real-world environments.

Photo by I am Alex on Pexels
Photos provided by Pexels