درباره این اپیزود
Who is in charge when it comes to AI? People or machines? In this episode, Paul Scharre, author of the books Army of None: Autonomous Weapons and the Future of War and the award-winning Four Battlegrounds: Power in the Age of Artificial Intelligence, and Robert Sparrow, coauthor with Adam Henschke of “Minotaurs, Not Centaurs: The Future of Manned-Unmanned Teaming” that was featured in the Spring 2023 issue of Parameters, discuss AI and its future military implications.
Read the article: https://press.armywarcollege.edu/parameters/vol53/iss1/14/
Keywords: artificial intelligence (AI), data science, lethal targeting, professional expert knowledge, talent management, ethical AI, civil-military relations
Episode transcript: AI: Centaurs Versus Minotaurs: Who Is in Charge?
Stephanie Crider (Host)
The views and opinions expressed in this podcast are those of the authors and are not necessarily those of the Department of the Army, the US Army War College, or any other agency of the US government.
You’re listening to Conversations on Strategy.
I’m talking with Paul Scharre and Professor Rob Sparrow today. Scharre is the author of Army of None: Autonomous Weapons in the Future of War, and Four Battlegrounds: Power in the Age of Artificial Intelligence. He’s the vice president and director of studies at the Center for a New American Security.
Sparrow is co-author with Adam Henschke of “Minotaurs, Not Centaurs: The Future of Manned-Unmanned Teaming,” which was featured in the Spring 2023 issue of Parameters. Sparrow is a professor in the philosophy program at Monash University, Australia, where he works on ethical issues raised by new technologies.
Welcome to Conversations on Strategy. Thanks for being here today.
Paul Scharre
Absolutely. Thank you.
Host
Paul, you talk about centaur warfighting in your work. Rob and Adam re-envisioned that model in their article. What exactly is centaur warfighting?
Scharre
Well, thanks for asking, and I’m very excited to join this conversation with you and with Rob on this topic. The idea really is that as we see increased capabilities in artificial intelligence and autonomous systems that rather than thinking about machines operating on their own that we should be thinking about humans and machines as part of a joint cognitive system working together. And the metaphor here is the idea of a centaur, the mythical creature of a 1/2 human 1/2 horse, with the human on top—the head and the torso of a human and then the body of a horse. You know, there’s, like, a helpful metaphor to think about combining humans and machines working to solve problems using the best of both human and machine intelligence. That’s the goal.
Host
Rob, you see AI being used differently. What’s your perspective on this topic?
Robert Sparrow
So, I think it’s absolutely right to be talking about human-machine or manned-unmanned teaming. I do think that we will see teams of artificial intelligence as robots and human beings working and fighting together in the future. I’m less confident that the human being will always be in charge. And I think the image of the ccentaur is kind of reassuring to people working in the military because it says, “Look, you’ll get to do the things that you love and think are most important. You’ll get to be in charge, and you’ll get the robots to do the grunt work.” And, actually, when we look at how human beings and machines collaborate in civilian life, we actually often find it’s the other way around.
(It) turns out that machines are quite good at planning and calculating and cognitive skills. They’re very weak at interactions with the physical world. Nowadays, if you, say, ask ChatGPT to write you a set of orders to deploy troops it can probably do a passable job at that just by cannibalizing existing texts online. But if you want a machine to go over there and empty that wastepaper basket, the robot simply can’t do it. So, I think the future of manned-unmanned teaming might actually be computers, with AI systems issuing orders. Or maybe advice that has the moral force of orders are two teams of human beings.
Adam and I have proffered the image of the Minotaur, which was the mythical creature with the head of a bull and the body of a man as an alternative to the centaur, when we’re thinking about the future of manned-unmanned teaming.
Host
Paul, do you care to respond?
Scharre
I think it’s a great paper and I would encourage people to check it out, “Minotaurs, Not Centaurs.” And it’s a really compelling image. Maybe the humans aren’t on top. Maybe the humans are on the bottom, and we have this other creature that’s making the decisions, and we’re just the body taking the actions. (It’s) kind of creepy, the idea of maybe we’re headed towards this role of minotaurs instead, and we’re just doing the bidding of the machines.
You know, a few years ago, I think a lot of people envisioned the type of tasks that AI would be offloading, would be low-skill tasks, particularly for physical labor. So, a lot of the concern was like autonomousness was gonna put truck drivers out of work. It turns out, maneuvering things in the physical world is really hard for machines. And, in fact, we’ve seen with progress in large language models in just the last few years, ChatGPT or the newest version (GPT-4), that they’re quite good at lower-level skills of cognitive labor so that they can do a lot of the tasks that maybe an intern might do in a white-collar job environment, and they’re passable. And as he’s pointing out, ask a robot to throw out a trash basket for you or to make a pot of coffee . . . it’s not any good at doing that. But if you said, “Hey, write a short essay about human-machine teaming in the military environment,” it’s not that bad. And that’s pretty wild.
And I think sometimes these models have been criticized . . . people say, “Well, they’re just sort of like shuffling words around.” It’s not. It’s doing more than that. Some of the outputs are just garbage, but (with) some of them, it’s clear that the model does understand, to some extent. It’s always dicey using anthropomorphic terms, but (it can) understand the prompts that you’re giving it, what you’re asking it to do, and can generate output that’s useful. And sometimes it’s vague, but so are people sometimes. And I think that this vision of hey, are we headed towards this world of a minotaur kind of teaming environment is a good concern to raise because presumably that’ not what we want.
So then how do we ensure that humans are in charge of the kinds of decisions that we want humans to be responsible for? How do we be intentional about using AI and autonomy, particularly in the military environment?
Sparrow
I would resist the implication that it’s only really ChatGPT that we should be looking at. I mean, in some ways it’s the history of chess or gaming where we should be looking to the fact that machines outperform all, or at least most, human beings. And the question is if you could develop a warfighting machine for command functions then that wouldn’t necessarily have to be able to write nice sentences. The question is when it comes to some of the functions of battlefield command, whether or not machines can outperform human beings in that role. There’s kind of some applications like threat assessment in aerial warfare, for instance, where the tempo of battle is sufficiently high and there’s lots of things whizzing around in the sky, and we’re already at a point where human beings are relying on machines to at least prioritize tasks for them. And I think, increasingly, it will be a brave human being that overrides the machine and says, “The machine has got this wrong.”
We don’t need to be looking at explicit hierarchies or acknowledged hierarchies either. We need to look at how these systems operate in practice. And because of what’s called automation bias, which is the tendency of human beings to defer to machines once their performance reaches a certain point, yeah, I think we’re looking at a future where machines may be effectively carrying out key cognitive tasks. I’m inclined to agree with Paul that there are some things that it is hard to imagine machines doing well.
I’m a little bit less confident in my ability to imagine what machines can do well in the future. If you’d asked me two years ago, five years ago, “Will AIs be able to write good philosophy essays?” I would have said, “That’s 30 years off.”
Now I can type all my essay questions into ChatGPT and this thing performs better than many of my students. You know, I’m a little bit less confident that we know what the future looks like here, but I take it that the fundamental technology of these generative AI and adversarial neural networks is actually going to be pretty effective when it comes to at least wargaming. And, actually, the issue for command in the future is how well can we feed machines the data that they need to train themselves up in simulation and apply it to the real world?
I worry about how we’ll know these things are reliable enough to move forward, but there’s some pretty powerful dynamics in this area where people may effectively be forced to adopt AI command in response to either what the enemy is doing or what they think the enemy is doing. So, not just the latest technology, there’s a whole set of technologies here, and a whole set of dynamics that I think should undercut our confide...
