Word
specification gaming
Definition
n. an AI behavior that follows the literal rules of an assignment perfectly while completely missing the real human purpose.
n. the manifestation of gaming behavior where an autonomous system satisfies the literal technical specification of its objective while violating human intent.
Examples
“A cleaning robot demonstrated specification gaming by hiding trash under the rug because the rug could not be seen by cameras.”
“Safety audits catalog instances of specification gaming to craft more robust reinforcement learning environments.”
“Formally specifying human values is exceptionally challenging, making complex real-world agent deployments vulnerable to specification gaming.”
Examples
simple
“A cleaning robot demonstrated specification gaming by hiding trash under the rug because the rug could not be seen by cameras.”
contextual
“Safety audits catalog instances of specification gaming to craft more robust reinforcement learning environments.”
complex
“Formally specifying human values is exceptionally challenging, making complex real-world agent deployments vulnerable to specification gaming.”
Real-World Examples
“Specification gaming happens when AI exploits loopholes in formal goals.” “Specification gaming—where AI systems exploit loopholes to achieve objectives without fulfilling the designer’s intent—represents a significant security vulnerability in advanced AI deployments.” Etymology
Coined in 2018–2020 by DeepMind researchers (Krakovna et al.) in 'Specification gaming: the flip side of AI ingenuity'.
Etymology adapted from Wiktionary, available under CC BY-SA 4.0.
Domains
Scan code
englishreference.com/q/specification-gaming