OREGON STATE UNIVERSITY
Open search box
Negative side effects are the undesired effects of an autonomous agent’s actions that occur in addition to the agent’s intended effects when operating in the open world. The problem of avoiding negative side effects is an emerging topic in AI. A technical description of the problem, a comprehensive overview of different forms of negative side effects and the recent research efforts to address them is summarized in: "Avoiding Negative Side Effects due to Incomplete Knowledge of AI Systems".
We created this repository of negative side effects identified in deployed AI systems, based on new articles and published research papers. For each entry in our repository, details such as problem setting in which negative side effects were observed, a description of the side effects, location and date of incident, are provided. We believe this repository will promote a deeper understanding of the problem, provide insights about which assumptions are valid, and facilitate moving beyond simple gridworld-based domains for common test cases to evaluate techniques.
If you are using this repository for research, please cite the following paper: “Avoiding negative side effects due to incomplete knowledge of AI systems” published in AI Magazine Winter Edition 2022.
@article{saisubramanian2022avoiding,
title={Avoiding negative side effects due to incomplete knowledge of AI systems},
author={Saisubramanian, Sandhya and Zilberstein, Shlomo and Kamar, Ece},
journal={AI Magazine},
volume={42},
number={4},
pages={62--71},
year={2022}
}