Welcome. We are the organization dedicated to the forensic post-mortems on AI failures. Our team is focused specifically on unintended harms—for example, a medical bot giving fatal advice or a chatbot encouraging self-harm.
We believe it is necessary to understand the "why" behind the error. Our methodology follows the aviation safety model: Every accident serves as a lesson to prevent recurrence.
We conduct our work with Analytical Depth and Victim-Centricity. Most importantly, we practice the Blameless Post-Mortem; we look for systemic faults, not individuals to blame.
A structured wiki of failures. We classify these by the nature of the failure observed.
Detailed PDFs breaking down exactly why a specific agent failed. We utilize chain-of-thought log analysis.
Detailed log review of context window overflow leading to dosage error.
Investigation into bias within the validation set for auto-claims.
"To build the safety, we must understand the failure."
Operating principally from San José (Costa Rica), Bogotá (Colombia), and CDMX (Mexico).
© 2023 Safe AI Investigation Organization