MIT AI Risk Repository · Risk Sub-Category · 15.01.07

Implementation

Category: First-Order Risks

Description

"This is the risk of system failure due to code implementation choices or errors."

From The Risks of Machine Learning Systems (Tan2022), as extracted by the MIT AI Risk Repository (CC BY 4.0).

Risks from Malfunctions
International Scientific Report on the Safety of Advanced AI (Bengio2024)
Fine-tuning related (Degrading safety training due to benign fine-tuning)
Risk Sources and Risk Management Measures in Support of Standards for General-Purpose AI Systems (Gipiškis2024)
Operational harms (financial markets)
Risk Sources and Risk Management Measures in Support of Standards for General-Purpose AI Systems (Gipiškis2024)
Extreme Risks
Cataloguing LLM Evaluations (InfoComm2023)
Safety & Trustworthiness
Cataloguing LLM Evaluations (InfoComm2023)
Limitations of Human Feedback
AI Alignment: A Comprehensive Survey (Ji2023)
Uncertainty concerns
AI Hazard Management: A Framework for the Systematic Management of Root Causes for AI Risks (Schnitzer2024)
Choice of untrustworthy data source
AI Hazard Management: A Framework for the Systematic Management of Root Causes for AI Risks (Schnitzer2024)