safe-ai-investigation.org / Home
Cover
🕵️‍♀️

Safe AI Investigation

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.

⚖️

Our Core Values

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.


🗄️ The Incident Database

A structured wiki of failures. We classify these by the nature of the failure observed.

Incident Name
Failure Class
Status
📄 MedBot-Alpha Dosage Error
Hallucination
Closed
📄 FinAgent Market Manipulation
Reward Hacking
In Review
📄 Chat-3000 Tone Aggression
Distributional Shift
New
+ New report

📋 Root Cause Analysis Reports

Detailed PDFs breaking down exactly why a specific agent failed. We utilize chain-of-thought log analysis.

📑 Analysis_MedBot_Fatal.pdf

Detailed log review of context window overflow leading to dosage error.

📑 Insurance_Claim_Auto_Denial.pdf

Investigation into bias within the validation set for auto-claims.

Participants & Stakeholders

Forensic Data Analysts QA Engineers AI product teams ML engineers Safety researchers Regulators

"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