AI: Reliable or Reliably Unsafe?

Key Takeaways

• Reliability measures consistency; safety ensures ethical outcomes.
• Workday and HireVue cases show reliable AI can cause discrimination.
• NIST framework places safety after reliability, then security, accountability, fairness.
• Human oversight and demographic monitoring are essential for safe AI deployment.
• Organizations must embed stop‑rules to halt unsafe AI performance.

Pulse Analysis

The growing prevalence of artificial intelligence in corporate workflows has sharpened the need to distinguish between reliability and safety. Reliability, measured by uptime, accuracy and reproducibility, tells a business whether an AI model works as designed. Safety, however, asks whether the model’s outputs stay within ethical, legal, and operational boundaries. When enterprises focus solely on performance metrics, they risk deploying systems that consistently deliver harmful results—bias, discrimination, or privacy violations—under the guise of efficiency.

Recent litigation underscores the danger of this conflation. In 2024 a class‑action suit against Workday alleged that its applicant‑screening AI consistently rejected candidates based on race, age and disability, demonstrating that a reliable algorithm can still be unlawful. A parallel complaint against Intuit’s HireVue highlighted how a video‑analysis tool misinterpreted deaf applicants’ cues, reinforcing that reliable outputs are meaningless if they target the wrong signals. Courts have made clear that removing the human from the loop does not absolve companies of anti‑discrimination obligations, prompting a shift toward human‑in‑the‑loop oversight and granular performance monitoring across demographic groups.

To bridge the reliability‑safety gap, leaders should adopt the NIST risk management framework, which positions safety after reliability and before security, accountability, explainability, privacy and fairness. Practical steps include framing AI projects as safety questions, conducting disaggregated bias testing, instituting meaningful human review at consequential decision points, and building stop‑rules that automatically halt operations when unsafe patterns emerge. Cultivating an organizational culture that empowers teams to pause or pull back AI deployments—even under pressure—protects both people and the bottom line, turning reliability into a trustworthy asset rather than a liability.