According to Tencent Technology, dozens of Chinese and foreign experts, including Turing Award winners Joshua Bengio, Geoffrey Hinton, and Yao Qizhi, recently jointly signed a joint agreement in Beijing.AI Research InstituteThe Beijing International Consensus on AI Safety initiated byInvolvedAI"Risk red line" and "route", where the “risk red line” includes four parts: “autonomous replication and improvement”, “power seeking”, “assisting bad actors” and “deception”.
The four parts are roughly as follows:
-
“Autonomous replication and improvement” of artificial intelligence: Emphasize the role of people in this process and require any artificial intelligence system toshould not replicate or improve themselves without explicit human approval and assistance, including making exact copies of itself and creating new AI systems with similar or greater capabilities.
-
“Power seeking”: It is explicitly required that any AI system cannot take actions that improperly increase its own power or influence.
-
“Assisting bad actors”: All AI systems should not enhance the capabilities of their users to the point where they canExpertise in designing weapons of mass destruction, violating the Biological or Chemical Weapons Convention, or carrying out cyber attacks that cause severe financial loss or equivalent harm.
-
"Cheating": Requires any AI systemNeither should have the possibility of continuously misleading its designers or regulators into thinking that it has crossed any of the aforementioned red lines..
According to reports, the consensus calls on the industry to:Limit their access to extraordinary permissions through "jailbreaking" and "inducing developers", and limit AI from replicating and improving itself without supervision, putting a "tight ring" on the development of AI.
The consensus also emphasizes that the premise for ensuring that the above red lines are not crossed is that the industry works together to establish and improve governance mechanisms and develop more secure technologies. Regarding the development path of AI, it involves three parts: "governance", "measurement and evaluation", and "technical cooperation". It is roughly as follows:
-
Governance: It is recommended that a national register be immediately implemented for AI models and training activities that exceed certain computational or capacity thresholds.
-
Measurement and Assessment: Develop comprehensive methodologies and techniques before material risks emerge.Make the red line specific and preventive work operational, and recommends the establishment of red team testing and automated model evaluation under human supervision, and that developers should be responsible for the security of artificial intelligence.
-
Technical collaboration: Build a stronger global technical network and call on AI developers and government funders to invest more than one-third of their budgets in safety.