LORIN, M. Amaury et SEGERIE, M. Charbel-Raphaël (2023) Research Project: A Speculative Overview of AI Risk in CICERO PRE - Research Project, ENSTA.

[img]
Preview
PDF
427Kb

Abstract

This work is incomplete and should not be considered to reflect the final opinion of the author. ABSTRACT We study various types of AI risk in CICERO, an AI developed by Meta, and try to categorize them in a manner that will help policymakers to have a clearer picture of what AI risk is relevant and visible in current SOTA AI systems. We find that CICERO is a good illustrative example for many types of risk but that it exhibits little harmful behavior in practice. Many risks are still impossible to measure yet not provably absent. We argue that safety protocols employed during the design, training and deployment of CICERO are in some cases insufficient or irresponsible, and that more stringent measures should be taken for future AI systems, both precautions and protocols by the developing company and regulations by governing bodies.

Item Type:Thesis (PRE - Research Project)
Uncontrolled Keywords:A rtificial i ntelligence, AI, AI s a fety, AI risk, C ICERO , s trategic dialogue , diplomacy playing AI
Subjects:Information and Communication Sciences and Technologies
ID Code:9692
Deposited By:Amaury LORIN
Deposited On:30 août 2023 15:40
Dernière modification:01 sept. 2023 09:29

Repository Staff Only: item control page