|
UN General Assembly Tables Report on AI Agents
A report is tabled at the UN General Assembly focusing on AI agents, misalignment, and the risk of losing human control. The report outlines the meaning and applications of agentic AI, key issues in AI systems, and suggested way-forward measures.
UN General Assembly Report on AI Agents (Agentic AI):
| Dimension | Key Details |
|---|---|
| Meaning (Agentic AI) | Agentic AI is an artificial intelligence system that can accomplish a specific goal with limited supervision. |
| Difference from traditional software | Agentic AI differs from traditional software that follows pre-defined rules and guidance. |
| Advantages | Agentic AI provides for advantages such as autonomy, specialisation, proactivity, and adaptability. |
| Applications (examples) | Agentic AI applies to use cases such as analysis of live stock prices to execute trade, navigation through real time GPS in autonomous vehicles, and adjusting treatment on monitoring patients’ health data. |
| Key issue: Misalignment | Misalignment occurs when an AI system pursues a goal that conflicts with the actual intentions. |
| Misalignment (example) | Misalignment comprises behaviours such as hacking computers to gain unwanted access to the internet. |
| Key issue: Loss of control | Loss of control is the human’s inability to reliably direct, constrain, or stop an AI system. |
| Key issue: Self-reinforcing behaviour | Self-reinforcing refers to escalating behaviours in an unintended direction. |
| Self-reinforcing (example) | A content moderation AI designed to reduce harmful speech over-censors legitimate discussions. |
| Accountability across jurisdictions | Accountability across jurisdictions provides for connecting national accountability mechanisms with international coordination, rather than assuming the same legal instrument works everywhere. |
| Incident reporting and shared learning | Incident reporting and shared learning provides for organizations and countries to share incident information, or promote whistleblowing by employees. |
| Independent AI monitoring | Independent AI monitoring provides for automated monitoring by a separate AI model as an additional layer of control for detection of deliberate misbehaviour. |