Three Laws Safe

This is an Economist article about the ethical obligations robotics programmers have.

There’s a lot embedded in it.

Starting at the beginning, the article refers to HAL’s “problem” in 2001: A Space Odyssey; to fulfill his obligations to keep the mission secret from the humans onboard and fulfill the mission, he decides to kill them. The blogger claims this shows that “Society needs to find ways to ensure that they are better equipped to make moral judgments than HAL was.” The HAL comparison and the later argument that robots may need to be able to act on better than pre-defined rules argue in favor of an heuristic approach, but there’s a good reason to doubt the value of heuristics: we’ve been trying heuristics (kind of, at least if you believe that even the pure reason Enlightenment philosophers still took The World as an experience input before they started reasoning) for centuries, and we don’t know what ethics to teach our robots.

Referring to HAL’s decision as a “problem” in his programming is unfair though (and makes any balance the writer shows later in the article moot). It’s clear that the writer believes that robots should have the same implicit standards — killing is bad, theft is bad, willful deception is bad — that we do, while one of the biggest advantages of machine intelligence could be its ability to think either without any intuitive biases or with a different set of intuitive biases. Oh, critics of philosophical thinking can never be outside of the philosophical system they hope to examine? That wouldn’t need to be true if we weren’t so obsessive about making sure robots thought like we do. The issue is of course that we’d teach robots sets of biases that, even if different, would take our own as reference points, so its alternative biases would be heavily filtered through our metabias (which I guess already means something in statistics). We could solve that problem with diverse programming groups in terms of occupation and ethical background and combinatorial ethical programming of thousands of robots.

The above assumes that we even ought to teach robots ethics. While teaching a robot that, to use another example, there are ethical complexities involved in destroying a house with three terrorists inside and one civilian (or any non-zero quantities of terrorists and civilians) may be difficult, it’s less unreasonable to expect a robot to understand political complexities. A robot may be better able to answer the question “Should we stop using robots that randomly kill children?” with something better than a snarky “Accidentally bombing children with our super army of automated missile firing bots may have worked at the start of the war, but conditions on the ground have changed.” With enough complexity programmed into its rules — a reasonable goal however amorphous “enough” may be, given the improvements in processing power, memory density, etc. since 2001: ASO — drones and other combat robots could equally likely determine that particular attacks aren’t necessary.

Let’s go back to the three terrorists and a civilian inside a house example. If, for example, the long-run security or strategic backlash of killing the civilian outweighed the risk of attack from those three terrorists, the robot would not attack. Better still, robots don’t condition their risk estimations in an environment of fear; with realistic probabilities programmed in, instead of the sweeping fear that every terrorist will blow up the White House tomorrow if we don’t kill him today, the drones would decide that most attacks just aren’t worth it.

This advantage would show up in two places. First, drones with clear decision rules would be better able to react to an enemy’s being away from civilians. Rather than “minimizing” civilian casualties, an intelligent robot could eliminate them. If it takes a certain consistent amount of time for a high-ranking enemy’s vulnerability to be communicated to those who can authorize an attack, enemies can plan travel between inhabited areas that takes no longer than that window. That strategy would leave shot-callers with a continuous choice between killing civilians and passing on the shot, and considering the surge in drone strikes during the Obama administration (it looks like we got the tough on terror president after all), passing on the shot is unlikely. It’s easy, then, with the human element, for enemies forcibly to erode whatever ethical mandate the US may have had at the beginning of an operation.

The second place the advantage shows up is the avoidance of stupid battles. Nevermind who fired first in Fallujah, the claim that drones/robot combatants “would not commit rape, burn down a village in anger or become erratic decision-makers amid the stress of combat” is important. A drone with decision rules would not have opened fire on a crowd of civilians in Fallujah because, as far as I can tell, there was nothing important in Fallujah. Then, with nothing important to defend, that same drone would not have entrenched and fought an eight month battle. During the eight month battle that the drone wouldn’t have fought, it also wouldn’t have tied up 21 Iraqis, blindfolded them, cut off their legs, executed them, and thrown them in a mass grave (again, if that is what occurred, but even if it is not, think of it as a stand-in for other atrocities).

I’m not thrilled that we’ve killed an important someone else by drone strike, especially because I feel like we kill Al Qaeda’s number two in Afghanistan once every three months. But if we’re going to continue to “fight terror” in Afghanistan, I guess I prefer doing so with five foot Hellfire missiles to stationing tens of thousands of trained killers from a different culture on the ground and asking them to please only shoot at bad guys.

The flip side of all of this, of course, is that robots with risk calculations for everyone in the world and who kill subjects whose risk to some stable order or particular security interest exceeds acceptable levels would be the most terrifying thing that has ever been. We’d basically set ourselves up for this:

But that doesn’t mean we can’t hope for this:

The blogger makes two (boring) recommendations for improving robot ethics. First, laws about accountability, programmers, etc., etc. Ok, we get it, accountability is big right now. Why we would introduce accountability in this ethical grey area when it’s noticeably absent in so many others (private equity, for instance, or US fopo) remains a mystery, but whatever, accountability is fine. Second, “where ethical systems are embedded into robots, the judgments they make need to be ones that seem right to most people,” you know, because usually when people agree on a thing it’s ethically and justified instead of just democratically justified. For apparently being into high-minded ethical rules, the blogger took the easy way out on that one.

What’s exciting about the blog post isn’t particular points though as much as the reality that we live in a world in which it’s important to think about what kind of ethics we teach our robots. Imagine an autonomous one of these following Aristotle around. Admittedly, it would probably have to follow around Žižek or Sandel today and would thus only learn how to be exceptionally angry or impressed with its own idea of distributive justice, but that would just be the beginning. Teach two autonomous robots Rawls and Sandel, two Badiou and Žižek, and two Dworkin and Nozick, leave them in a room, write down the synthesis they produce, and (go away, Gödel) suddenly, the question of what the result of such an attempted reconciliation of beliefs would be becomes answerable. I won’t say “knowable” because there are some obvious biases towards particular types of knowledge present, but even answerable would be a great leap forward.