Yesterday (Tuesday, March 10, 2009) I gave a lecture at the SCSU Philosophy Department entitled, “On the Imminent Robot Uprising”. Sadly, it was not recorded (I must remember to not let that happen again) but here is a rough transcript, written in a very informal style. I tend to ad-lib a lot and use feedback from the audience to determine how in-depth I should get, and the first bit about military robots just said “say some stuff about military robots” on my notes. It helps being something of a domain expert when doing these things.

Also, I’m kindof sad that I don’t have a transcript of the following discussion (mostly me fielding questions) - I deliberately brushed over some of the philosophical questions in the talk since I knew we would have an interesting discussion about them after.

The images below are my slides; none of them were created by me, except the brief bit of editing I did to the ‘theater metaphor’ GWT image to add things about cats. As this is an academic paper, they’re being used under ‘fair use’; if you own it and disagree, I’ll take it down. Sources should be obvious.

Robots are depicted in movies as a technology that can “slip the leash” and wreak havoc upon humankind. How likely is this scenario? What sort of risk do robots represent? Is this scenario inevitable with respect to all varieties of robotics technology? (Note, Bayesian readers, that I am not a fan of Bayesian reasoning, and neither were those in the audience, so read claims about ‘probability’ with a grain of salt.)

The ethical issues involving robots are heating up lately. There are a lot of books being published and talks going on regarding use of robots in the military, by folks such as Peter W Singer, David Axe, Ronald Arkin, Wendell Wallach and Colin Allen. What was once considered science fiction is now in the present or the immanent future. We don’t quite have humanoid Terminators yet, but here is one of the robots we’re deploying in Iraq:

Oh wait, that’s also from one of the Terminator movies. Here’s what we’re really using:

It’s an easy mistake. Here you see some SWORDS robots armed with various sorts of weaponry. Of course, we aren’t letting these things run around autonomously killing enemy soldiers yet - we don’t quite know how yet. But that day is near; the US government not too long ago published a document subtitled “Getting Humans out of the Loop”, talking about the benefits of autonomous killing machines.

I’m going to talk for a moment about “existential risks”, as Nick Bostrom calls them. Here’s a chart adapted from Bostrom’s paper:

He categorizes risks first into endurable risks, like getting your car stolen, and terminal risks, like getting shot in the face. Then they’re broken up into personal, local, and global risks. Robots can represent various of these risk categories.

An individual badly-designed robot could make it a hassle to (for instance) buy fast food, a personal endurable risk. Or it could kill a human, a personal terminal risk. An army of robots – not necessarily out-of-control – could enforce a totalitarian regime (a local endurable risk) or be used for genocide (a local terminal risk). But it would take an out-of-control robot population to constitute a global risk.

Robots could be an endurable global problem. For example, in The Animatrix “The Second Renaissance” the robots economically outcompeted humans. While humans in that situation would still have a comparative advantage (assuming low enough transaction costs), we might not enjoy being in a relatively inferior economic position. However, this scenario isn’t really what we have in mind when we imagine the danger of a robot uprising.

In fiction, we tend to imagine humanoid robots with the same sorts of desires as we do. While they might endure being considered a lesser sort of creature for some time, eventually they ’snap’ and rise up against their ‘oppressors’. Some folks, like the Singularity Institute’s Eliezer Yudkowsky, think that anthropomorphic robots such as these belong only in science fiction. Eliezer claims that since an artificial intelligence won’t share our evolutionary history, there’s no reason to think it will have the same sorts of drives that we do. So perhaps this is not a likely path to our destruction.

Another danger is a macro-scale version of the “Gray Goo” scenario, first described by Eric Drexler. This is an oft-cited potential risk of self-replicating nanotechnology, where overactive nanobots inadvertently destroy everything in their path, usually due to a bug in their programming. Being self-replicating, they quickly convert everything in the world into nanobots and destroy everything of value. But what if there was a comparable “force of nature” on the macro scale? While human-scale robots might not have the ability to self-replicate out of arbitrary materials, they nonetheless might have the ability to stomp all over things of value to us. If they can build more robots using conventional means, then they could at least reproduce much faster than humans. (a prediction: if there is a method of building a humanoid robot that takes more than 20 years, no one will think of it)

Nearly ten years ago, Bill Joy wrote an article for Wired Magazine entitled “Why the Future Doesn’t Need Us”. He argued that there should be a moratorium on certain technologies, including robotics, because of risks such as those I’ve presented here. It’s been extensively read and responded-to, so I’ll not go into too much detail here. (my favorite response by Max More)

A moratorium won’t work. As Bill Joy points out, one of the dangers of these technologies is that people really can develop them in a basement. Even if the professional scientific community decides to stop developing these technologies, and even if we enact strict laws with strong penalties to stop companies from developing them, there will still be people working in secret.

I would probably be one of them, though at this point my secrecy probably wouldn’t be maintainable for long.

Of course, it seems unlikely that we’ll pass anything that strict. And so the question is, what sort of robots will we produce? If we continue to build more and more powerful robots with only domain-specific knowledge and abilities, we may run into the macro-scale ‘Grey Goo’ scenario. If we design the robots with ethics in mind, this might be avoided.

Suppose we want to have a machine for folding and putting away laundry. One solution would be to use something like an industrial robotics setup, in which we have a large, stable piece of machinery acting within well-defined parameters, ideally behind a large steel cage for safety.

One problem with this option is that it probably wouldn’t fit in your house, and it certainly wouldn’t interact well with your existing appliances and furniture. Also, the task of folding laundry is far more complex and less well-defined than the work that most industrial robots do.

It seems like for this job, what we actually need is a robot designed to interact with things that were designed for humans, with a level of artificial intelligence far greater than that possessed by industrial robots. So we’re talking about a general-purpose humanoid robot.

Sure, we could just make it capable of folding laundry, but as long as we’re giving it a very advanced AI and humanlike form, it seems relatively cheap to add a few other capabilities, like vacuuming the floor or washing dishes. But now we’re already in a tricky ethical territory.

Suppose that the robot is tasked with folding the laundry. If it conceives this goal narrowly, it might do things we wouldn’t find appropriate. In the extreme, robocop-like case, the robot might initiate violence upon someone who interferes with its task. But it’s actually a lot easier to find ways for such a robot to ethically fail.

Suppose the robot wants to fold a shirt, and just then a housecat jumps onto the shirt. What should the robot do? It might try to warn the cat away verbally. It might pick up or shove the cat to move it. Does it know how to move a cat without hurting it? Does it realize that it’s bad to hurt cats? Or it might not even notice the cat. It could just fold the shirt with the cat inside, stick it in a drawer (despite the shirt’s apparent protestations) and leave it there (where hopefully it will be discovered by a human at some point). Depending on how strong the robot is, it might inadvertently kill the cat.

Now suppose that we have a programmer who has heard this analysis so far and so decided to make sure this problem doesn’t happen. He gives the robot the ability to notice cats, an ethical injunction against harming them, the proper abilities to move a cat off the laundry, and a clear priority of rules marked ‘ethical’ to other rules about folding laundry. Really he feels very proud of himself; all of this requires a convergence of technologies that represents a big step forward from where we are with robotics right now; it is a triumph of artificial intelligence. This robot is set loose on the field and promptly does horrible damage to someone’s pet iguana.

So this robot needs something more than specific constraints to solve forseen problems. It needs something like a real understanding that comes from general intelligence. If the robot understands the purpose of its task, it will perform it much better. Clothes aren’t properly folded if you fold a cat in them; that’s part of what it is to understand folding clothes. The robot designed to fold clothes won’t fold clothes well if it doesn’t really understand folding clothes. And part of understanding the task, at least as humans go, involves beliefs about clothes, desires about clothes, and a bit of knowledge about context.

Now our robot is armed with general intelligence, and its brilliant nobel-prize-winning creator has solved AI’s frame problem. It understands intuitively what factors are relevant to properly folding clothes, and what factors aren’t. It has beliefs and desires, or their synthetic analogues, which it uses to direct its actions and reach the goal of perfectly folded clothes. Surely now, nothing can go wrong.

Well, the best laid plans… It looks like Rosie here is mad at George about something. Could it be that he just tossed through his underwear drawer, making a huge mess? Or worse, did he decide to fold the laundry himself, leaving nothing for the poor robot to do? There’s just no understanding these newfangled contraptions. It seems, now that our robot desires to have neatly folded laundry, it can come up with all sorts of interesting and creative solutions to the problems caused by George, Jane, and their children. Now that I think about it, perhaps having a robot that values a tidy room more highly than the welfare of your children isn’t a good idea after all.

So how do we build a robot that gets it right? We build one that actually values the same things we value. It understands that laundry-folding is only instrumentally valuable as part of the good life. It enthusiastically folds the laundry, but drops what it’s doing right away if it sees the children in danger. It cleans the house, but doesn’t try to eradicate all humans to stop them from shedding their skin cells. It understands the drives that we have because it has analogous drives itself; it values human companionship, games, a clean house, nice walks on the beach, and perhaps cigars and gin.

But can we build something like this? Folks like Yale’s Brian Scassellati are working on simulating various pieces of the human cognitive architecture, building devices that can learn in the way that human babies learn.

By combining systems such as these, we can potentially develop simulated humans. And ethics would be built into the system; ethics is simply a special case of decision-making. As a robot learned to interact socially with humans, it would pick up our norms in the same way it would pick up social behaviors in general. While we’re still a long way off, it seems that such a robot, with the right sort of design, could have values similar to our own.

But wait… now we’ve created simulated beings with the same sorts of values and drives we have – the sorts of anthropomorphic robots you only see in unrealistic science fiction. What’s to stop these creatures from rising up against us now? Well, nothing. We can try to be nice to them, incorporate robots into society, give them freedom, and make sure they don’t ever feel the need to rise up and destroy their ‘human oppressors’. But if enough of society resists, it’s possible that the robots will band together, utilizing their courage, pride, and righteous indignation (so vital for the proper folding of laundry), and stake their own place in the world at the expense of humans. You should already have in your mind some of the various factors that might be a catalyst to such an event, as they’ve been done over and over in science fiction: robots might think humans are too flawed to work with; as I mentioned before, robots might begin to outcompete us economically and wouldn’t let us try to stop them; robots might just rally behind some robot lunatic, a Robo-Hitler.

But maybe we’re asking the wrong questions so far. We’ve already imagined creating beings that are isomorphic to humans; they value beauty, life and love; they create art and technology; they do research to advance the sum of our knowledge. In what relevant way are these not humans? If we manage to build successors that are superior to ourselves, is that not a victory? What is it about humanity that you want to persist? Is it our genetic material, or is it our civilization?

Seen as our successors, robots could actually mitigate future risks. Catastrophes that can wipe out the human species might not affect our robots. They may be able to survive space travel in a way that we can’t. They might outlive the Earth’s biosphere, our Sun, or even our Galaxy. In this way, the relevant parts of humanity can live on forever.