Detecting Feature Eligibility Illusions in Enterprise AI Autopilots
summary
Describes a concrete method to detect three failure modes—inverse causality, observability illusion, and end-state bias—that make ML models look accurate on static tables but fail in production. Uses audit logs when available to infer field update order and start/end distributions, or reconstruct a “movie” from time-stamped snapshots using heuristics to locate record age and sampling rate. Experiments across multiple enterprise tables compare log vs snapshot results and show snapshots are good enough to reveal most illusions and avoid blind deployments.
AI-written summary · about
chapters (8)
- 0:00Start
- 8:19Scoping the Problem
- 14:12Inverse Causality
- 16:41Observability Illusion
- 18:29End state bias
- 20:00Uncovering Illusions (0) Logs
- 21:09Uncovering Illusions (W) Taking Snapshots
- 23:38Take-home message
related talks
49mAn AI with an Agenda
50mRealities & Misconceptions about Enterprise Applications of AI/ML and Data
01h 05mNavigating the Security and Privacy Landscape of Modern AI
28mHindsight bias: A machine learning story
31mThe Dark Side of AI
18mAI Observability: How To Fix Issues With Your ML Model
25mWhy Good Models Fail After Deployment
16mOperationalizing AI in Finance: Real-Time Analytics