The Reputational Risks of AI

Abstraction: Framework for diagnosing and responding to AI reputational failures

Key points:

• Analysis of 106 AI controversy cases identified three failure modes: privacy intrusion (~50%), algorithmic bias (~30%), and explainability failures (~14%)
• Privacy failures stem from "data creep" (using data without consent) and "scope creep" (using data beyond consented purpose); DeepMind/NHS and Target are cited examples
• Root cause of algorithmic bias is data integrity — biased training data or model drift ("model creep") — not the algorithm itself
• Stakeholder perceptions split into capability (technical competence) vs. character (governance/ethics) dimensions; privacy failures are most often perceived as character failures (43% of cases)
• Capability failures require technical fixes with transparent timelines; character failures require governance reform and explicit opt-in consent mechanisms
• Microsoft's Tay chatbot (2016) illustrates a capability failure response: public apology, continued technical fixes, and eventual relaunch as "Zo"

Connections: Facebook · Amazon · Microsoft · Algorithmic Bias · AI Ethics · Explainability

Source: https://cmr.berkeley.edu/2022/01/the-reputational-risks-of-ai/