Senate Foreign Affairs, Defence and Trade Legislation Committee, , 8:47:54 to 8:53:35
A detailed debate on whether the ADF's AI policy requires a human to make lethal force decisions or merely holds humans accountable for AI actions.
Topics: Artificial intelligence, Law reform
Speakers: Hugh Jeffrey, David Shoebridge, Meghan Quinn, Raff Ciccone, Speaker 32
>> operators. Mr. Jeffrey search the policy. It doesn't include the word target or targeting.
>> I'm sorry Senator
>> policy doesn't include the word target
or targeting. The policy uh is to allow for us to integrate AI into our capability such that it is uh able to more quickly identify um targets in our uh in in that that are seeking to project force against Australia and of course everything to do with military capability ultimately needs to find fix and affect a target
>> but it doesn't it doesn't just go there I know we are deep in conversation here, but I need to rotate the call. I've given you well over 40 minutes. Um, how long have you got left with this mind because I just need to manage the other senators?
>> If if you just allow me two more questions.
>> Sure. Okay.
>> Um, the policy does not require human um a human to make a decision about the application of lethal force. The policy says humans will be accountable for the delivery um of lethal force, but it seems to be consciously drafted not to avoid not to require humans to make the decision about the delivery of lethal force. Was that a an intentional part of this this policy? Uh no senator
the policy is determined by the obligations of lawfulness, adherence to values-based principles and proportionate control measures. So the valuebased principles if you just give me a minute senator to outline are accountability. So defense personnel must exercise enforced judgment and care in the research design and development uh and use of decommissioning of a AI technologies and all outcomes uh will be attributable to an identified defense official. Bias and harm indication. AI technologies must be designed to enable detection, correction, and mitigation of unintended bias and unintentional harm where it can be reasonably anticipated, explainable, the function and the relationship between inputs and outputs of AI technologies. Uh must be uh traceable and the technology is able to meaningfully accommodate human machine interactions. Human centric. All AI technologies and their use must be considered in light of their planned likely and potential impact on humans at all levels of involvement and reliable and secure. AI technologies must perform reliably against well- definfined use cases with resilience to change or interference including mitigations and provisions for the risk of failure. So these policy settings align align with Australia's international legal obligations and domestic law and will ultimately be the basis on which we seek to integrate AI technologies into defense capability.
>> And yes, Mr. Jeffrey, and then um as I read the policy, it allows and permits AI to make the decision to kill somebody and to apply lethal force. um um um provided there's some chain of command or some chain of accountability back to a human but it is consciously and deliberately and allowing AI to make the kill decision.
>> No, Senator I I just wouldn't accept
identify the provision in the policy um that's that that that prohibits AI making the kill decisions or identifies the element in the policy that requires a human to make the kill decision. And I'm more than happy for you to take that on notice. Senator, if
I may, um, if you refer to point number one under lawfulness, it actually highlights human judgment and accountability is central to lawful, legitimate, and responsible use of AI and designated defense personnel. The accountable officers will always be accountable for its use, decisions, and outcomes.
>> Correct. So, humans set it up, AI kills somebody, and the humans accountable for the AI kill. That's what that says. No,
Senator. Um, if I could just add, um, AI is essentially another tool in the ADF uh, capability set. Anything in the capability set and particularly anything that's lethal uh, to affect a target ultimately has to be directed uh, uh, within the ADF command chain. >> AI can speed up that c that command chain. It can identify uh more risks and it can enable us to defend ourselves more effectively. But ultimately its application, its use in the context of the application of military force has to be directed.
>> Well, I'd ask you again on notice to identify the provision in this policy that requires a human to make the kill decision.
>> So very happy to take that on notice. It's a very important issue that you're raising. um and it's been one that's been well ventilated and discussed uh in Australia uh within the defense forces. I just make the note that this this is something we need to grapple with because um as we move forward AI is being integrated into all aspects of of our life and so very happy to take it on notice and set out how how it matches with the appropriate language in the policy. Uh you've heard the explanation from witnesses but we're very happy to take it to the next step and join together with the published policy.
>> All right. Very good. Um on that note, um Senator Roberts, I understand you've got a few questions and then I'll come to the deputy chair. Thank you.