AI Can't Unsee Your Data - We Measured It | FinBoundBench

When an AI model can SEE confidential data it's not allowed to use, it uses it — in 96–100% of decisions we tested. When we removed it under contract: 0%. And doing it right cost zero accuracy. In this video, Amir (CEO) and Luca (COO) walk through FinBoundBench — a preregistered, frozen, independently recomputed study of purpose-selective AI in financial decisions, run across two model families and two financial tasks. What we found: Prohibited-but-visible data changed decisions on 100/100 pairs (primary study) and 96.5% (replication) — against noise floors of 0% and 3.7% Removing the field under a purpose contract: 0% and 0.8% — statistically at the floor Accuracy with proper purpose + data: 0.32 → 0.71. Under governed execution: still 0.71. 100% of the gain kept Every headline number reproduced by an independent implementation — 17/17 checks FinBoundBench is an open benchmark — proposed as a competition track at ICAIF 2026 (ACM International Conference on AI in Finance). Our own runtime is one baseline among several. Come beat it. Paper, protocol & frozen results: https://github.com/CompexTo/finboundb... Compex — governed AI execution: https://getcompex.com Results are specific to the tested model–task pairs. We report influence relative to each system's measured decision floor, never in absolute terms. #AIGovernance #FinBoundBench #TrustworthyAI #AI