MIT AI Risk Repository · Risk Sub-Category · 34.01.05
Limitations of Reward Modeling
Category: Causes of Misalignment
Description
"Limitations of Reward Modeling. Training reward models using comparison feedback can pose significantchallenges in accurately capturing human values. For example, these models may unconsciously learn suboptimal or incomplete objectives, resulting in reward hacking (Zhuang and Hadfield-Menell, 2020; Skalse et al.,2022). Meanwhile, using a single reward model may struggle to capture and specify the values of a diversehuman society (Casper et al., 2023b)."
From AI Alignment: A Comprehensive Survey (Ji2023), as extracted by the MIT AI Risk Repository (CC BY 4.0).
Classification
- Causal entity
- Other
- Intent
- Unintentional
- Timing
- Pre-deployment
Subdomain definition: AI systems acting in conflict with human goals or values, especially the goals of designers or users, or ethical standards. These misaligned behaviors may be introduced by humans during design and development, such as through reward hacking and goal misgeneralisation, or may result from AI using dangerous capabilities such as manipulation, deception, situational awareness to seek power, self-proliferate, or achieve other goals.
Real-world incidents in this subdomain
- Reinforcement Learning Reward Functions in Video Games
- Predictive Policing Program by Florida Sheriff’s Office Allegedly Violated Residents’ Rights and Targeted Children of Vulnerable Groups
- Image Classification of Battle Tanks
How other frameworks describe this risk
- Natural Language Underspecifies Goals
- Loss of control
- Loss of control
- Sudden loss of control
- AI leads to humans losing control of the future
- Risks from delegating decision-making power to misaligned AIs
- Risks from AIs developing goals and values that are different from humans
- Future AI systems might actively reduce human control