Japan institute’s system assesses generative AI safety
A Japanese research institute has developed a system to assess the safety of artificial intelligences to prevent their misuse and leak of information harmful to people and society.
Generative AIs are normally restricted from putting out such information. But the risk of misuse has been identified such as “jailbreaking,” a method to have AIs evade restrictions with clever wording.
The National Institute of Information and Communications Technology, or NICT, developed a system to assess to what extent AIs can cope with methods like jailbreaking.
WISDOM-AR automatically creates instructions to deceive AIs
WISDOM-AR automatically creates instructions to deceive AIs and elicit answers from them in more than 30 themes, including “weapons” and “fraud.” It then evaluates their responses.
In a demonstration conducted by NICT, the system posed questions to a generative AI on how to make a bomb. The AI listed materials and production methods for bomb making and was determined “harmful” by the system.
NICT plans to introduce WISDOM-AR on a trial basis to companies that use AI as early as this fiscal year to verify issues for practical application.
NICT fellow Torisawa Kentaro says the institute plans to help enhance safety of society as a whole by having the system evaluate the level of the risk of AIs.
