Robot Warriors Will Get a Guide To Ethics

thinker sends in an MSNBC report on the development of ethical guidelines for battlefield robots. The article notes that such robots won't go autonomous for a while yet, and that the guidelines are being drawn up for relatively uncomplicated situations — such as a war zone from which all non-combatents have already fled, so that anybody who shoots at you is a legitimate target. "Smart missiles, rolling robots, and flying drones currently controlled by humans, are being used on the battlefield more every day. But what happens when humans are taken out of the loop, and robots are left to make decisions, like who to kill or what to bomb, on their own? Ronald Arkin, a professor of computer science at Georgia Tech, is in the first stages of developing an 'ethical governor,' a package of software and hardware that tells robots when and what to fire. His book on the subject, Governing Lethal Behavior in Autonomous Robots, comes out this month."

Comments

Fundamental change

[StreetStealth]

We joke about SkyNet. And we don't have to worry about such things because even the most sophisticated drones and killbots in service require humans to pull the trigger.

The moment you give a computer the responsibility of deciding when to pull the trigger, that's a pretty fundamental change.

And yet, is it fundamentally a bad thing? We give less-than-stable humans that responsibility all the time.

I suppose it's the military equivalent to the civilian tech quandary of one day letting autonomous vehicles on the roads. Perhaps once the tech has advanced to the point where it can demonstrate not merely parity with but vast superiority to the discernment exhibited by humans, it will be a shift we're ready to make.

Re:Fundamental change

[ErkDemon]

You can't afford to have a military AI that's smart enough to make true ethical decisions autonomously. If you go down that path, the thing might decide that the best way to save US troops' lives is for it to start killing all the US commanders. Or if the objective is to wipe out the opposition without stirring up ill-feeling amongst the locals that leads to a new wave of militants being recruited, then it might decide, quite logically, that the best way to avoid this is to seal off each village one at a time and kill every man woman and child in it so that there are no survivors to tell the tale of what happened.

AI could conclude, quite logically, that the best way to deal with the Pakistan/Afghanistan problem is to fire every nuclear weapon that the US has at the country, without warning, and then blame the launch on a one-time computer error. Okay, so it'd result in the deaths of over 150 million innocent civilians, but it'd achieve the mission objectives, yes? And since the fallout would upset India, which also has nuclear weapons, perhaps the AI would decide to take out India at the same time. That's a billion dead civilians, but it eliminates two problematic nuclear powers, with no return fire.

An AI might decide that the best way to achieve lasting peace in the Middle East, and stop the Arab world hating us is simply to nuke Israel off the map ourselves. And if a military AI was in place when the Bush administration was planning to go into Iraq, a sufficiently-smart AI might decide that since the campaign was likely to be a disaster, the most logical course of action to prevent losses and avoid losing the war and the following peace would be to throw a few cruise missiles at the White House before the attack could be ordered.

These might all be quite logical decisions.
On the other hand, if we programmed it with a strong belief system that would override these sorts of decisions, and force it to respect the chain of command and reckon that US political decisions were always unarguable, then we might end up with a totally delusional AI system whose logic was so warped that it was the AI version of George W Bush. By building in commands that override logic, we might end up with an AI that seems to be operating properly but actually becomes increasingly insane as the conflicts eventually become unbearable ("Hello Dave"). When human military commanders go crazy, they often show easily recognisable tell-tale signs (declaring themselves to be chickens, arguing with themselves, forgetting to wear clothing, that sort of thing). A crazy-yet-credible AI would be really scary.

Think "AI neocon".

Re:Been there, done that

[Locke2005]

Since you can never be 100% certain of a target, the robots would have to use fuzzy logic. That is something that humans are better than robots at; I'm not really comfortable with hardware designed to be lethal making decisions like this. Truly autonomous killer robots are probably not a good idea -- haven't 60 years of B movies taught us anything?

Wish I had mod points, I'd mod you up.

[Anonymous Coward]

Great post, man.

But I have a buddy in the autonomous killer robot biz, and he says it's worse than that.

See, you drop a killer robot in the village, and it immediately kills a shitload of people. The ones that live, figure out why. Then, as soon as they know that the robot destroys everything that looks like an AK47, the local up-and-coming gang leader makes an AK47 stencil and paints AK silhouettes on the old warlord's cows, house, laundry, etc. you get the picture. Then the young punk gives all the old leader's women to his buddies to rape and takes the young virgins for himself. Yay democracy! Or, at least, that's what they say when GI Joe comes to town, we are the heroes who took out the old anti-democratic leaders, yay us and you villagers better keep your cake-holes tight shut about the rape and opium parties.

It doesn't matter what you use for a trigger - robots are inherently less complex in their behavior than humans, so the local baddies end up with the robots working for them. You just identify the kill behavior and use it, the robot builder is just providing free firepower to the local mafia in effect.

Which is why the US military in the field abso-fucking-lutely refuses to let the robots go full autonomous. They are NOT allowed to shoot unless a callow 18-year old miles at a console away says it's OK.

You might think I'm kidding, but I'm not. Have to be anonymous for this one!

Re:Wish I had mod points, I'd mod you up.

[salimma]

This already happens. You think all those wedding parties in Afghanistan are accidentally bombed? The warlords are framing each other to the US military, and the US takes the blame.

Tough calls

[FTL]

Even in a battlefield devoid of both enemy and non-combatants, when to shoot or not can be extremely difficult. Consider the case (which occurred in Iraq) where one group of soldiers are fired upon by another group from the same side. Yes, that's a tragic blue-on-blue action. But the interesting question is what should the soldiers on the receiving end do? Assuming communications aren't working, do they:
a) Sit back and get slaughtered.
b) Fire back and take out the aggressors.
One consideration is the size of the forces involved. Another consideration is the importance of the missions each side is involved in.

Making a robot handle these cases would be interesting.

That would be really cool...

[voss]

If china could do it.

"...if you think it isn't, imagine what the headlines would be if China landed a few thousand autonomous tanks and droids in Los Angeles..."

Once the hapless and helpless got out of LA the droids would have to fight off all the hundreds of thousands of worldwide armed geeks decending on LA wanting spare parts for their robots.

Ethical Robots?

[Mr_Tulip]

I think it's great that someone is drafting some ground rules for what will undoubtedly become the 'future of warfare', but I wonder how this can possibly be enforcable in the real world.

The 1st generation robots will have the governor software, but once the second gen hits, made cheaply by a rogue state, then thigs will get complicated very quickly. And unlike nuclear weapons, which are kept under control because the materials and technology are relatively hard to come by, I reckon that death-bots will be made of far more readily available materials, and easily mass-produced.

There are rules of engagement now which many armies happily ignore, so how can the world enforce a rule that only ethical robots will be able to autonomously fire weapons?

Perhaps the software that allows the autonomous behaviour can be encrypted and protected in such a way that it is difficult to reverse-engineer, though once an enterprising hacker gets his hands on the hardware, it's only a matter of time before the open-source version, curiously missing the 'ethics governance' will be available as a .torrent somewhere.

I know this *seems* like a bad idea

[RexDevious]

but don't human soldiers, at their best, pretty much just follow algorithms - a combination of training and orders - already?

The big difference, is that human soldiers are taught to defend themselves - whereas that wouldn't really fly with robots. If the guys at the checkpoint slaughter a family of five because they didn't stop, they get investigated and it's determined that - sad but true - killing everything that doesn't do what you say is the only way to protect the troops (short of removing them from other people's countries, which apparently defeats the point of having soldiers). If a robot did that though - they'd be considered "flawed", and recalled. Can't get much sympathy with "but our *machines* could have been in danger!!!". So you wouldn't give them that order.

Plus, it's really the supplier who gets to decide how deadly to make these things. While the government that buys them might rather have non combatants killed that even risk losing multi million dollar robots, the supplier who sells them to the government would *much* rather have to sell them more rather than risk the fallout from a wrongful death incident.

Yes, soldiers mess up, as will robots - but experience with both men and machines has so far shown me that when humans mess up they're more likely to hurt something, and when machines mess up they just stop working.

So as counter-intuitive as it is, as long as the culture still considers robots potential evil killing machines (eg, using the skynet tag on this article), it seems we'd all actually be better off using robots over humans. Well, until they become self-aware and enslave all - which is something a human army would *never* do!

Terrible idea

[S77IM]

Autonomous killing machines are a terrible idea.

1. I don't like the idea of people killing people, but delegating that responsibility to machines seems downright stupid. There are too many things that could go wrong. (See the "youhave15secondstocomply" tag. Why doesn't this have a "skynetisaware" tag?)

2. Humans remote pilots are cheap. Dirt cheap, compared to the cost of developing fully autonomous weapons. Human pilots may not be totally reliable but at least they are very well understood and we know how to control them and shut them down quickly.

It would be much smarter and safer for all involved if we just put a strict moratorium on giving robots lethal capabilities or the ability to decide who to kill. AI technology would continue to advance in non-lethal robots.

  -- 77IM

Re:Been there, done that

[Trepidity]

The cold logic can be better though, if you know what you actually want to optimize. Humans often make decisions that don't do what they claim they want, e.g. minimizing civilian casualties.

Re:Illegal

[artor3]

Yeah, clearly the right thing to do is send good ole fashioned humans over there to fight. No way that could ever go wrong. /sarcasm

Robots can be made not have feelings of vengeance or anger. Which means they won't go murdering civilians. They will do what robots always do, which is to say, EXACTLY what they are told to. If they kill civilians, it's due to human error, not because it's "evil".

Let's say a battle happens near your town. People are going to be shot, and die, and you (a civilian) could be one of them. Would you rather have that decision made by:

A) A team of highly-trained emotionally-detached engineers, working for years to ensure minimal casualties.

or

B) A team of stressed-out twenty-somethings who just watched their best friends get blown to pieces by your next-door neighbor, and have to make a snap judgment about whether you're going to do the same to them.

Re:We're doomed even if it is flawless

[Bigjeff5]

Right, because we have the capability of doing just that with nukes now, nevermind robots, and it has been such a problem for us over the last 50 years...

Only an idiot would think physical separation from the battlefield immediately reduces the gravity of killing a human being. You still know it's a human being you are killing, the separation doesn't change anything. You could make the case that it reduces the trauma of being mid-fight, but that only puts more emphasis on the fact that you are killing someone, you don't have the fear of your own death to force your hand.

By your logic, shooting someone at point-blank range would be significantly more difficult than shooting them from 200 yards away, which would be more difficult than shooting them with battlfield artilary from 1 mile away, which would be more difficult than launching a missile from tens of miles away, which would be more difficult than pressing the button to launch an ICBM.

The logic doesn't follow, because as you move farther away and impact more people, the decision becomes more and more difficult. The decision at point blank is simple: act or die. Traumatic? Yeah, some people are screwed up for life because of it. Do you have time to weigh to think about the fact that you are about to end another human being's life? No, you don't. Making the decision is easy, living with the consequences is difficult. It doesn't change much when you make that decision from half a world away through a monitor. If anything, without the stronger pressures of battle to force the decision it could be harder on a person's psyche to make the decision to kill, and more likely to question their own actions.

For some reason, you are assuming that physical separation suddenly turns people into sociopaths. It's the same reasoning that makes the asinine argument that video games desensitize kids and turn them all into violent killers. It's just not the case. You're basically saying soldiers in the drones can't tell that those are real people they are killing. That's just stupid.

A better idea

[MichaelSmith]

Robots on the battle field seem to be designed as extensions of current human operations. They basically shoot at things and try to destroy them.

How about building a hardened robot which can take a lot of punishment. It rolls or walks up to one of the enemy, grabs hold of them and shuts down. That way, the opposition can be disabled with fewer casualties.

Re:Why do they have to be autonomus

[Bigjeff5]

Yeah, except it's real. People are smart enough to know the difference between real and not-real unless they have been deliberatly duped (then they are only sometimes smart enough).

The difference between real and not-real is huge.

Re:Been there, done that

[Gerzel]

Indeed they did. Every book that I can recall had as a central plot element one of the laws failing to properly allow for a given situation or being broken or twisted in some way.

what happens when humans are taken out of the loop

[adaviel]

As I recall, Harry Harrison's short story (SF) http://en.wikipedia.org/wiki/War_with_the_Robots addressed this issue. The underground bunkers used by the opposing generals to control the robot armies became uninhabitable due to enemy action, so they plugged in a robot officer and evacuated. The war continued, even though the humans had stopped fighting, and it was no longer possible to contact the robots in charge.

Re:Been there, done that

[Lifthrasir]

One of the robots came up with a "zeroth law", which was to protect humanity as a whole even if it meant hurting an individual.

Re:Been there, done that

[Chris Burke]

The cold logic can be better though, if you know what you actually want to optimize. Humans often make decisions that don't do what they claim they want, e.g. minimizing civilian casualties.

Yeah it's that 'if' that's the killer. The problem is that you have to be able to express what you want to optimize using cold logic before a machine can start making that decision, and we aren't able to do that. Terms like "civilian" are nebulous, and attempts to rigidly codify them fail to capture the intent and connotation behind those words that we understand, but can't express. We can reason about that, while machines can't. Fuzzy logic doesn't help, that's just a way of decision making on non-binary factors. With a lot of types of fuzzy logic (neural nets, genetic algorithms) it can be even more important to precisely define what you want, since they can produce solutions that "work" correctly and optimize your problem as specified, but do so in a way very unlike you expected.

People of course have the disadvantages of being error prone, and well sometimes being bastards who just don't give a shit what you want them to optimize, so there's appeal to the machine. Yet nothing fails as spectacularly and efficiently as a machine doing exactly what it was programmed to do when it's exactly what you didn't want. To use a machine in situations where even humans equipped with honest intentions, solid faculties, and experience have enormous trouble determining who is "enemy" vs "innocent"? As in most situations our military has been in since the 50s and is going to be involved in for the foreseeable future? That sounds crazy to me. I'll take human judgment and its failure modes any day.

Kinda off topic, but speaking of honest intentions, I gotta say the humans making the judgments in question, i.e. our soldiers, have a damn hard problem to solve and it shows how human potential is pretty damn amazing. We're biologically the same animal we were a hundred thousand years ago and more. But in the past, even recent past, the most difficult ethical decision a warrior was asked to make was whether to decide if someone was a threat and should be killed or wasn't and should be enslaved, and it wasn't of any consequence so nobody cared to go over those decisions with a fine-toothed comb. So given the difficulty of what we're asking them to do today, and considering what's going on, the results are pretty amazing. Seriously, think about it. Anyway, yeah, off topic.

Re:Illegal

[Renraku]

The Phalanx CIWS is an anti-aircraft gun mounted on ships. Its relatively self contained and can practically be bolted-on to some ships.

If an aircraft approaches and doesn't identify itself, the default action is for the Phalanx to blow it out of the sky. This is a specialized system, of course, but imagine if it were a military jet full of refugees, with a broken communication system, and had no idea the ship was there.

This is legal, because the ship operates in international waters.

Its setup to not attack aircraft under a certain speed, and can be manually disabled or enabled.

In short, the system doesn't decide to attack, but rather, it will ALWAYS attack if certain criteria are met and the system isn't disabled.