Could we make a moral machine? Could we build a robot capa­ble of decid­ing or mod­er­at­ing its actions on the basis of eth­i­cal rules? Three years ago I thought the idea impos­si­ble, but I’ve changed my mind. So, what brought about this u‑turn?

First was think­ing about sim­ple eth­i­cal behav­iors. So imag­ine some­one not look­ing where they’re going. You know, look­ing at their smart­phone, about to walk into a hole in the ground. You will prob­a­bly inter­vene. Now, why is that? It’s not just because you’re a good per­son. It’s because you have the cog­ni­tive machin­ery to pre­dict the con­se­quences of their actions.

Now imag­ine it’s not you but a robot, and the robot has four pos­si­ble next actions. So, from the robot­’s per­spec­tive it could stand still or turn to its left, and that the human will come to harm, will fall in the hole. 


But if the robot could pre­dict the con­se­quences of both its and the human’s action, then anoth­er pos­si­bil­i­ty opens up. It could choose to col­lide with the human to pre­vent them from falling in the hole. And if we express this is an eth­i­cal rule, which you see here, this looks remark­ably like Asimov’s First Law or Robotics, which is that a robot must not injure a human or through inac­tion cause a human to come to harm.

So thus emerged the idea that we could build an Asimovian robot. We need to equip the robot with the abil­i­ty to pre­dict the con­se­quences of both its own actions and oth­ers’ in its envi­ron­ment, plus the eth­i­cal rule that I showed you in the pre­vi­ous slide.

Screenshot of a robot simulator showing several robots playing what looks like soccer

Image: Webots

In fact, the tech­nol­o­gy that we need to do this exists and it’s called the robot sim­u­la­tor. So, roboti­cists use robot sim­u­la­tors all the time to mod­el and test our robot code in a vir­tu­al world before run­ning that code on the real robot. But the idea of putting a robot sim­u­la­tor inside a robot, well, it’s not a new idea but it’s tricky and very few peo­ple have pulled it off. In fact, it takes a bit of get­ting your head round. The robot needs to have, inside itself, a sim­u­la­tion of itself and its envi­ron­ment, and oth­ers in its envi­ron­ment. And run­ning in real-time as well.

Three robots standing in a room, with one inside a bounded-off "danger zone"

So, over the past two years we’ve actu­al­ly test­ed these ideas with real robots. In fact, these are the robots. We don’t have a hole in the ground, we have a dan­ger zone. And we use robots instead of humans. We use robots as proxy humans. So let me show you some of our lat­est exper­i­men­tal results.

Here we have a the blue robot, the eth­i­cal robot, is head­ing towards a des­ti­na­tion. This is its goal. But it notices right here that the red robot, the human, is head­ing toward dan­ger. So the blue robot choos­es to divert from its path to col­lide (gen­tle col­li­sion) with the human, to pre­vent it from com­ing from harm. This is exact­ly the same thing but a short movie clip. You can see again, the blue robot is the eth­i­cal robot. Our red robot is the proxy human. Cute robots, aren’t they.

So, we also test­ed the same with an eth­i­cal dilem­ma. Here our eth­i­cal robot is faced with two humans head­ing toward dan­ger. It rather dithers, rather hes­i­tant, and of course it can­not save them both. There isn’t time. Ethical dilem­mas are a prob­lem real­ly for ethi­cists not roboticists.

So, how eth­i­cal is our eth­i­cal robot? Our robot imple­ments a form of con­se­quen­tial­ist ethics. In fact, we call the inter­nal mod­el a con­se­quence engine. The robot behaves eth­i­cal­ly not because it choos­es to, but because it’s pro­grammed to do so. We call it an eth­i­cal zom­bie. Our approach has a huge advan­tage, which is that the inter­nal process of mak­ing eth­i­cal deci­sions is com­plete­ly trans­par­ent. So if some­thing goes wrong, then we can replay what the robot was think­ing. I believe that this is going to be real­ly impor­tant in the future, that autonomous robots will need the equiv­a­lent of a flight data recorder in air­craft. An eth­i­cal black box.

So, what have we learned? Well, the biggest les­son, in fact the thing that caused my u‑turn, is this, that we do not need to make sen­tient robots to make eth­i­cal robots. In oth­er words, we don’t need a major break­through in AI to build at least a min­i­mal­ly eth­i­cal robot. We don’t need to build Data from Star Trek.

I’d like to leave you with a ques­tion about the ethics of eth­i­cal robots. If we can build even min­i­mal­ly eth­i­cal robots, are we moral­ly com­pelled to do so? Well, with dri­ver­less cars just around the cor­ner, I think it’s a ques­tion that we’re going to have to face real­ly quite soon. So thank you very much indeed for lis­ten­ing. Thank you.

Further Reference

Alan Winfield’s blog with follow-up post and slides, and his staff pro­file at the University of West England, Bristol web site.