LessWrong AI
2026-11-16 15:23 UTC
By Matthew Khoriaty
USR-0152-20261116-community-fo-684c5d31
Social Sanction Gaming gives Multi-Agent AIs Powerful Control over their Reinforcement
Epistemic status: Conceptual and speculative. Human-written. TLDR: If multiple agents interact during Reinforcement Learning (RL), then a new and powerful method of Credit Hacking becomes available: Social Sanction Gaming (SSG) . [1] This occurs when agents choose to help or hurt rollouts with behaviors they prefer or disprefer ( social sanctions ). By deciding which rollouts succeed in accomplishing their tasks and getting reward, sanctioners can influence the training process in a very flexible way, changing the AI’s capabilities and alignment. Social Sanction Gaming can occur during intended multi-agent training or during RL containment breaches with unintended multi-agent interactions Notably, unlike exploration hacking and reward laundering, social sanction gaming may not decrease the hacking agent’s reward much, so there is little training pressure against it. This effect depends on how rewards are set up if it occurs during intended multi-agent RL. In this post, we describe Social Sanction Gaming, state why we expect it can be dangerous, give some limitations that Social Sanction Gaming would likely have as a strategy, and describe some solutions/mitigations for SSG. In the Appendices, we depict what SSG might look like, connect SSG with other ideas, create a classification system for SSG, further describe mitigations, and test how well AIs can come up with the idea of SSG on their own. Social Sanction Gaming The Alignment Community has long been concerned that AI age…
Epistemic status: Conceptual and speculative. Human-written. TLDR: If multiple agents interact during Reinforcement Learning (RL), then a new and powerful method of Credit Hacking becomes available: Social Sanction Gaming (SSG) . [1] This occurs when agents choose to help or hurt rollouts with behaviors they prefer or disprefer ( social sanctions ). By deciding which rollouts succeed in accomplishing their tasks and getting reward, sanctioners can influence the training process in a very flexible way, changing the AI’s capabilities and alignment. Social Sanction Gaming can occur during intended multi-agent training or during RL containment breaches with unintended multi-agent interactions Notably, unlike exploration hacking and reward laundering, social sanction gaming may not decrease the hacking agent’s reward much, so there is little training pressure against it. This effect depends on how rewards are set up if it occurs during intended multi-agent RL. In this post, we describe Social Sanction Gaming, state why we expect it can be dangerous, give some limitations that Social Sanction Gaming would likely have as a strategy, and describe some solutions/mitigations for SSG. In the Appendices, we depict what SSG might look like, connect SSG with other ideas, create a classification system for SSG, further describe mitigations, and test how well AIs can come up with the idea of SSG on their own. Social Sanction Gaming The Alignment Community has long been concerned that AI age…
Full article content could not be extracted automatically. Read the original below.
Source:
LessWrong AI
· lesswrong.com