Power Measurement for AI Safety: New mathematical framework Comes

As artificial intelligence systems become increasingly stronger, their strength has become a critical security question. Optimal designed AI agents tend to gain strength regardless of their goals. This can cause people to become increasingly unpowered in socio-technical systems.

A new research brings together two different power sizes to solve this problem:

• Information-theoric strengthening (empowerment): The capacity of the information channel a agent’s actions can create on future situations; a mathematical consep based on Claude Shannon’s communication theory Tags
• Active-based power: A relative metric that measures the potential for which people can achieve
• Hedef disclosure (Goal ifşa): The mechanism that bridges these two approaches; increases human control by explaining the targets of the agent
• Markov in the context of games: mathematically modeling power balance in multi-gen systems

Research offers an official, quantitative solution to the problem of gaining unlimited power of artificial intelligence agents. In human-AI collaboration systems, making agent’s intentions transparent, has critical importance in terms of protection of human authority. This frame can help build a more balanced power relationship between people and machines in the design of future AI systems.