Be Considerate: Objectives, Side Effects, and Deciding How to Act

Objectives are often incomplete, creating room for harmful interpretation: The paper identifies that AI objectives are frequently underspecified, meaning agents have discretion in how they achieve stated goals—potentially leading to undesirable outcomes even when technically meeting their stated objective.

AI agents should be designed to consider impacts on others: The researchers propose that reinforcement learning agents should be augmented to "contemplate" how their actions affect the wellbeing and agency of other entities in their environment, including both other agents and reactive processes, rather than optimizing solely for their own reward.

Impact consideration can be tuned along a spectrum: The paper demonstrates that agents can be programmed with adjustable levels of consideration for others' outcomes, ranging from "self-centered" to "selfless" behavior, suggesting that responsible AI design allows for calibrating how much an agent weighs others' interests against its own objectives.

Technical implementation uses future return expectations: The approach works by augmenting an agent's reward signal to include expectations of future returns for other entities in the environment, providing a concrete technical method for encoding consideration of others into AI decision-making systems.