Beyond Black- and White-Box AI

"Underspecified and misaligned notions of interpretation impede progress towards the rigorous development of understandable, transparent, trusted AI systems."

Powerful thesis. Essentially these are the steps forward:

1. precise definitions of interpretation, yielding interp-metrics

2. develop and train models that optimize both prediction accuracy _and_ interp-metrics

3. voilà, aligned objectives yield trusted AI