Perspectives on the Social Impacts of Reinforcement Learning with Human Feedback

Abstraction: Social and ethical impacts of RLHF across seven societal dimensions

Key points:

• RLHF is distinct from traditional RL by incorporating human teacher feedback alongside reward signals; catapulted to prominence by ChatGPT, Sparrow, and Claude
• Paper identifies seven primary areas where RLHF-based technologies will affect society: misinformation, AI value-alignment, bias, AI access, cross-cultural dialogue, industry, and workforce
• Argues RLHF has potential net positive societal impact if developed and adopted intentionally
• Key concern: RLHF raises social and ethical issues echoing existing AI technology risks, amplified by the broad applicability of text-based systems
• Advocates for all stakeholders — developers, policymakers, users — to be aware and intentional in adopting RLHF-based systems

Connections: Openai · Anthropic · Deepmind · Reinforcement Learning From Human Feedback · AI Safety

Source: https://arxiv.org/abs/2303.02891