Killer robots can be taught ethics

(Credit: Signet)

Adherence to the Three Laws of Robotics as put forth by Isaac Asimov has been, until now, entrusted to whoever held the joystick. That may change.

A robotics engineer at the Georgia Institute of Technology has developed an "ethical governor," which could be used to program military robots to act ethically when deciding when, and whom, to shoot or bomb.

Ron Arkin has demonstrated the system using attack UAVs and actual battlefield scenarios and maps from recent U.S. military campaigns in Afghanistan. (videos)

In one scenario, a drone spots Taliban soldiers, but holds its fire because they're in a cemetery--fighting there is against international law.

In another, the UAV identifies an enemy convoy close to a hospital, but limits itself to shooting up the vehicles so as to avoid collateral damage to the hospital. The mindful bot would also house a built in "guilt system," which would force it to behave more cautiously, after making a mistake.

While the work shows promise, it also draws attention to the inadequacy of trying to program machines with morals, especially ones expected to perform in a complex battlefield environment, according to experts.

"Robots don't get angry or seek revenge but they don't have sympathy or empathy either," Noel Sharkey, a roboticist at Sheffield University, U.K., told New Scientist. "Strict rules require an absolutist view of ethics, rather than a human understanding of different circumstances and their consequences."

Arkin acknowledges that it may take a while before we can trust predators and other unmanned killers with life and death decisions.

"These ideas will not be used tomorrow, but in the war after next, and in very constrained situations." Arkin is quoted in New Scientist. "The most important outcome of my research is not the architecture, but the discussion that it stimulates."

[mbraswell3]

if area == "graveyard":
fire = False

Its not that hard

[zmjman08]

the difficulty lies in determining whether or not an area is a graveyard.

[CrimsonCantab]

howabout:

try {
if person.isSurrendering() || person.unarmed() || person.isBegging() ...
fire = true
}
catch MightBeSuicideBomberException
//too hard to figure out, just shoot the guy
//better safe than sorry!
fire = true
catch HatesAmericansException
//shouldn't deserve to live
fire = true
catch WearingCivilianClothesException
//meh, who cares if we take out a few citizens as well
//they might be a threat
fire = true


yea, that was pretty easy!

[fokkwp]

a) Naturally, there'll be an override button

b) The override button will be on the desk of the Justice Department under a red protective lid labeled "only in case of war." And on the same desk will be a small sign reading "The war on terror never ends."

[myles taylor]

And people are so good at following morals? In our quest to make sure robots don't become too logical, we forget that how many people are murdered every day? How many people are hurt or injured by humans? I'd say that even if they are barely able to get it right, the fact that they will always follow it makes them a step above humanity.

[shootfirst]

This is BS they are talking about the 3 laws, and one of them is to not harm humans. Even putting a touch of gray in that understanding opens up too much interpretation. Machines should not be able to make a decision period about dropping bombs or using live ammunition. Someone should watch that Star Trek Voyager where 2 factions were in a war and they built robots to fight the war with programming to kill the other side till the war was won. Basically everything but the robots died and they kept fighting each other, although their numbers were dwindling because they could not reproduce themselves until Voyager interfered. We have tons of scifi that already look at the complications and the repercussions of letting machines run rampant with rule sets that are incomplete or not sophisticated enough.

I also agree computers/AI isn't intelligent enough to decipher a simple picture and there is no way they should or can determine who is in the wrong and who is in the right.

[gsigas]

Computers are smart enough to decipher pictures if there are only certain things they care about (i.e. a constrained context) and if they have enough memory and processing power. For example you can train a computer to recognize the human form or what weapons look like and what the characteristics are (how do people move, how do they emit heat, what is a weapon made of, how is a weapon used, etc). A powerful enough computer can then make extremely accurate guesses about what in the picture is a person, what in the picture is a weapon, what in the picture is a vehicle or structure etc. So computers are very capable of identifying what is in the picture within a specific context. Given enough accurate and specific information about the context a computer would also be able to decide what is a valid target and what isn't. The computer would suffer from the same limitations as a human in its decision making in that it would be limited by the quantity and quality of accurate information (lower quantity and quantity the more likely a mistake) and by the scope of its objective (the larger the scope the more likely a mistake).

[yearstretch]

Asimov?s later works updated the 3 Laws to include a 4th Law (called the 'Zeroth Law' as it took precedence over the other 3). The Zeroth Law requires robots "Not to allow humanity to come to harm ..." in order that robots could protect humanity en masse even at the expense of individual humans. Such a law would have allowed the murder of Hitler for example which would be impossible under the original 3 Laws. It is that level of programming that is most difficult as it requires objective abstraction and a clear view of how much harm may be caused by one action (killing Hitler) in preference to another action (not killing Hitler). Of course the real moral dilemma here is that Hitler himself thought he was protecting humanity when he mass murdered the insane, infirm, gypsies etc. etc. etc. So ? how does one arrive at an algorithm that selects between the relative merits of human actions? Clearly it?s easy in the case of a clear and present danger where a robot may kill a man running at you with a knife but the farther away from the actual event one is the harder it gets. Imagine a robot knowing that a man was about to shoot from a mile away then the robot could easily be told to kill him but of course that would first require the robot to know of the action and the outcome, which may be more a matter of sensor and intelligence input to the robot than a moral dilemma. At a thousand miles a man pressing a button to release a missile is out of reach and anyway how do tell the robot that one button releases a missile while a different button pours coffee? You see the problem.

As one who has been playing with machine intelligence algorithms for more than 35 years now I can only say it?s a fascinating subject and each new step towards machines that are more intelligent, in a general rather than a specific sense, which we have already reached, than we are is to be encouraged as it should provide an objective measure that we as mere wetware humans can learn from.

[Curmudge]

The Zeroth Law opens the door to robotic mega-nanny who will protect you from yourself. Recall the old movie "Colossus, the Forbin Project." To prevent harm to "humanity", humanity itself will have to be taken out of the driver's seat.

[mbraswell3]

if Hitler.isPresent():
Fire = True
elif Graveyard.isPresent() or Hospital.isPresent() or School.isPresent:
Fire = False
else:
Fire = True

I modified my code a little

[willdryden]

There was an Outer Limits show like that. The robot put the family on a diet, locked each kid in their own room, and would not let the family out of the house because it might harm them. The first time a robot tries to tell me what to do, I'm grabbing a cutting torch.

[zeroplane]

This makes sense, Just look at what happened to RoboCop in movie 2? heehe.

[Dr_Zinj]

Define "humanity."
Define "harm."

For humans to evolve, they would no longer be "human" by most definitions. And insults to genetics or development don't fit the definition of human to a lot of people; which is why we allow abortions.

We deliberately cause harm to each other and our children every day. It's necessary to apply stress for learning and growth. People who exercise their bodies are causing harm. People who don't exercise their bodies are causing a different sort of harm. Harm is relative and situational and totally dependent on values. Because every person has a unique set of values, you will never have a 100% acceptable definition of harm or benefit.

Come to think of it, the definition of humanity is value-based also. So no concurrence in that matter either.

The Zeroth Law requiring robots "Not to allow humanity to come to harm ..." violates human free will. It's also a violation of the Federation Prime Directive.

[esiders]

This is hilarious, How can you teach ethics to a robot when the people teaching the ethics abide by them. It is a simple case of practice what you preach.