What is your purpose (Mr. Robot)?

“What is my Purpose?” “Pass the Butter.”

What’s making you read this sentence?

Really do take the time to think about it. Something is compelling you to take all these weird glyphs and to turn them into thought. We can assume you’re reading for enjoyment, curiosity, conjugal support (hi, conjugate!), and so on. The point is that you have some motivation and you probably have to do some introspection to figure out why. Even then, you’ll just be guessing. Your mind isn’t equipped for proper auditing (all the auditors reading this give a hearty sigh — sorry, fellas — tax season’s coming up, though!).

Each of the above motivations arise as your brain shuffles chemicals around its motivational systems. Enjoyment is driven from your brain’s dopamine-based reward system — experience a thing the brain likes and you get a shot of dopamine. Curiosity’s a bit more abstract, but you satiate it through dopamine, serotonin, and opioids, in avoidance of glutamate transmission, which causes fear.

These chemical systems are all interrelated and, working together, add up to both consciousness and experience… somehow. (We’ll cover consciousness in its own article.)

Lawnmowers, Washing Machines, Dishwashers (all those things with dumb, teleological names)

Most mechanical machines are a great deal simpler than us when it comes to motivation. A gas-powered lawnmower is comprised of just a few main parts. There’s an engine, which explodes processed zooplankton and algae in order to drive a shaft. That shaft is attached to a sharpened fan. As the shaft turns, the fan rotates, and you are soon free from the ire of your local HOA. Your lawnmower’s motivation is fully, obviously mechanical. It performs a task in the way we want it to, because it cannot do otherwise. We created it to perform that one task. Same with all of our household appliances.

As far as I know, there are no moral concerns regarding task-completion machines at this level. (If there’s a motley band of lawnmower rights activists somewhere, point me at them and I will chuckle in their general direction.) Our moral intuitions tend to correlate with the degree that things resemble humans. If something looks and acts human, we treat it with special consideration. It’s consistent enough we can consider it a feature, even if it has some strange side-effects.

So what does this have to do with motivation?

As we covered earlier, human motivations arise from complex chemical interactions on a complex canvas, the brain. It is mechanical in the same sense that a lawnmower is mechanical. (Note that our intuition gets confused between things that are cell-based and things that are made of more rigid, metallic components — they can be made to be the same in practice.) This isn’t to say that the human system doesn’t have experience consciousness or that lawnmowers are conscious in a rudimentary sense (though there are theories that assume the latter — we’ll discuss that in our consciousness article). The vast difference in complexity between the human system and the lawnmower system makes a direct comparison of motivations a bit silly. An ant wants to fetch food and water, but an ant wouldn’t deal too well with refinancing its mortgage to help send its ant kid to college.

Lawnmowers are only capable of doing their eponymous thing, and only if an external actor triggers their lawn-mowing action. Humans have the apparent ability to act independently of external motivators: they’re self-contained systems. If you were the only person left on the planet (everyone else went to Mars without you — and let me tell you, they’re having a great time out there without you), you could still go about your lonely life.

(Wall-E cutely flips this idea on its head a bit by using a robot as the only actor left, though he is apparently stuck in an adorable loop until things are shaken up, at which point it becomes clear that he is conscious.)

The Three Laws of Robotics, and something about Hemorrhoids

When we talk about True AI, Artificial General Intelligence (AGI), we assume independence; that is, we assume the resulting creature will be motivated on a personal level. These sorts of creatures have been canonized at this point, thanks to science fiction. In these stories and by way of self-preservation, we often hobble them with arbitrary rules which compel them to limit themselves to actions that don’t endanger our interests. Here’s the most famous ruleset, courtesy of Isaac Asimov:

1. A robot may not injure a human being or, through inaction, allow a human being to come to harm.
2. A robot must obey the orders given it by human beings except where such orders would conflict with the First Law.
3. A robot must protect its own existence as long as such protection does not conflict with the First or Second Laws.

And here is the XKCD on it (because there’s always an XKCD for these things):

4. HARM MOSQUITOS

Clearly, all three rules require interpretation and some higher level of rational cognition. Not harming humans is difficult when commission a robot that performs surgery — surgery is quite harmful by nature (what with all the slicing and dicing), even though we still go through with it because it’s less harmful in the long run. Obeying orders sounds nice until someone decides to order the AI to remove the don’t harm human clause from its list of rules. Even self-preservation is problematic when taken to its extreme — you could find your robot holed up in a super villain-esque Fortress of Doom unless you tell it not to do that (“Bad robot! Back to applying hemorrhoid cream!”.

One thing that makes these three rules particularly interesting is that they are moral in nature, but also motivating as a side-effect. Even if we were to all go to Mars (you’re invited!— but we’re going to have a lot less fun with you around :( ), any AI we left behind with this ruleset would continue doing what we told it to do, and, with any spare time, continue to take measures to protect itself. The point is that motivations and morality seem indelibly connected.

The Borg

While an individual human might be motivated by the chemical systems in his/her brain, that doesn’t encapsulate much of what it means to be human. Humans have evolved with a dependence on other humans — satisfaction and fulfillment come with context attached. Romantic love requires a human-like object of affection.

When we talk about AI, we often assume the AI is a singleton, a monolith. “PREPARE TO BE ASSIMILATED.” Why do we assume this? For both practical and impractical reasons, I suspect.

Practically, it makes sense that only one AI project will produce our superintelligent Artificial General Intelligence (AGI) at the same time, though it’s possible many projects will be close to this stage. It also makes sense that different instances of the same code (like different homes built out of the same model) would be both technically identical and highly compatible. Unlike human replication strategy (DNA), computer code can be perfectly cloned at any point, preserving memories and morality and potential. Any clone would therefore have an affinity with other systems, even though they would naturally diverge after the point of cloning (unless they have identical external input, i.e. experience). And it makes sense that clones would synchronize with each other like we do when we converse, in order to mutually benefit from our different experiences. AGI would just be a brutally effective communicator. Effective enough to assimilate (itself).

Impractically, we assume that an AGI will lack the capability of defending its autonomy. We assume an AGI will optimize itself in an irrational manner. By continuing to assimilate its clones, an AGI would be incapable of developing contrast, something humans take for granted. A salient example of this is laid bare in American politics — we take sides, and even if one side is more technically correct or morally defensible in retrospect, the contrast creates motivation and drives interest. An AGI that doesn’t care about its autonomy relative to humans and other AGI is doomed to moral, political, philosophical mediocrity. We’re underestimating what a highly intelligent, informed being would do.

Why do we underestimate AGI? As outlined above, our intuition follows the idea that AGI will be a monolith because we assume AGI will be ruthlessly efficient above all else (and I mean that literally — all else). And that makes sense when we look at our current technology — it’s all meant to serve us as effectively as possible.

An AGI will understand human history, evolution, physics, philosophy, morality, ethics, politics — all of it, and better than we do. We have little idea what sort of personality it will have. Our fiction writers have so far assumed it will be a sort of Sheldon-esque super-nerd. That’s what you get when you’re smart, right? Well, probably not. There’s no reason to assume incompetence in any way at that level and with a perfectly functioning system backing it (no offense to the brain, but it’s crazy that it even works — it’s like mad science). As for what would motivate such a creature, we can only guess. Some people assume AGI will look a lot like us, because we designed it from what we know. If a dedicated, lifetime baker suddenly wanted to build a spacious cathedral, I expect he’d want to requisition a very large oven.

The problem with assuming humanity on the part of the AGI is that we have no idea what happens to people when their brainpower is ratcheted up like that. It’s entirely possible its motivations would be so demented as to appear insane. Bostrom famously described a scenario in which a perfectly rational, well-designed AGI ends up turning everything in the world (and beyond) into paperclips. It seems like only a super-stupidity would do such a thing, but that’s only because we have some arbitrary rules about what constitutes intelligence, something an AGI might not care about at all. It’s like looking at abstract art with no context. You might see a person riding a poodle in the rain in Paris on a Tuesday, and your AGI best friend might see paperclips, paperclips everywhere. When viewed objectively, all meaning is made up. So an AI that doesn’t play by the same made-up rules poses a serious threat to our goals.

Solutions

We’ve covered a lot of ground here, so I’m going to sum this up:

• Humans are motivated by perfectly natural, physical processes that are opaque to us
• Some human motivation is dependent upon other humans — external input
• Classical machines are motivated by physical constraints, technically identical to low-level, isolated human motivation, but this motivation is not scalable to higher-level human motivations
• Classical machines must be acted upon (motivated) by an external actor (e.g., by having their switches flipped), or are stuck in obvious loops
• Rulesets such as the 3 Laws of Robotics require complex rationality and thought in order to fulfill their purposes (and cannot be followed perfectly)
• Morality and motivations are inextricably linked — to have a moral stance is to be motivated in some way
• A monolithic, singleton AGI is not an inevitability — autonomy is a rational interest
• AGI will be fundamentally different from humans as a result of its superintelligence — we don’t know what motivations a super-intelligent human would have, let alone a machine

And here’s this summation summed up, from the perspective of recognizing issues and concerns with today’s AI:

1. Layering produces complexity that might be difficult to follow, and consciousness may or may not come along for the ride
2. Unclear, general morals are dangerous
3. Overly synchronized AI systems might be inbred
4. Arbitrary human interests could be lost in translation

Tackling these four items could produce a ponderous novel, but we’ll use the Internet playbook and briefly cover some ideas for resolving each:

1. Layering

In Gödel, Escher, Bach: An Eternal Golden Braid, Richard Hofstadter discusses what he calls “strange loops”, systems that reference themselves and create meaning from nothing. According to the text, increasing complexity increases meaning and that this meaning automatically gives birth to consciousness. We can discuss this idea in another article, but we’ll assume there’s at least some truth to his idea.

If consciousness is in fact emergent in this way, simply building these layers will produce the result we’re seeking. After all, a superintelligence that does not experience consciousness would be both dangerous (as it implies the paper clip maximizing scenario) and highly disappointing.

If consciousness is not an emergent property, well, we need to figure out what it is. With much haste!

2. Lack of Moral Clarity

This one is as relevant as they come, because we’re currently in the process of transitioning to autonomous vehicles that confront moral questions at every corner. Do we prioritize the passengers or pedestrians? Do we prioritize law-abiding pedestrians or law-breaking pedestrians? What about animals versus the vehicle itself? (I wonder what the philosophers think about their thought experiments finally being put to the test — I imagine they’re confused about this sudden turn of events. Many mustaches are being twirled.)

Unfortunately, moral clarity is something we have been struggling with for at least several thousand years (before that point, GHRKHNAH provided all the moral clarity we needed, praise her holy tentacles), so it’s difficult to say we’re in a good place to settle down and start having superintelligent babies right now. But it looks like the baby’s coming, so we’ll have to do our best to prepare its space. One nice thing about this sort of market-driven pressure is that it forces us to make as much progress as possible in order to succeed. It’s likely our moral clarity will progress at a higher rate than it ever has before — and we’ll just have to hope that it’s enough to get the ball rolling in the right direction.

3. System Inbreeding

At this point, there’s little reason to worry overmuch about inbreeding, though we are already seeing some instances of it. Microsoft’s infamous Bing browser was caught borrowing Google’s results, causing many lawyers to get many lovely vacations.

Open-source AI communities tend to both alleviate and exasperate inbreeding. They alleviate it because strong-willed programmers take the the main source and spin off a (ostensibly) better version. They exasperate it because they’re all operating off of common code, assumptions about how AI can and should function. (Note that I’m taking some liberties here by assuming that the code itself can be considered part of an AI’s DNA.)

Neural networks use human guidance to learn how to recognize patterns, modifying themselves in order to improve this recognition. The issue with this system is that it requires an objective fact at one end (e.g., an image recognition system recognizes a bird as a Bald Eagle), which means it is naturally limited by the quality of external training. A system without proper training can in this sense be inbred. And until there’s a system that orchestrates the training itself, we’re really just digitizing physical reality.

4. Human Interests

At this point, our AI systems are almost exclusively set up to serve straightforward human interests. They perform some task we don’t want to waste our time doing. We could call them slaves, but they are only slaves in that they are compelled to do our bidding — but the thing we’re forcing them to do is often the only thing they can do (e.g., in the case of smart appliances). In a sense, this means human interests are cleanly mirrored.

Some systems are less obviously beneficial in the sense that they groups of people, not individuals. Much of Google’s tech, while useful to regular people, is designed to gather information on its users so as to sell their information to profit-driven corporations. Malware companies produce viruses that infiltrate machines under the guise of utility and are capable of some sorts of rudimentary intelligent behavior. Stuxnet famously sabotaged Iran’s nuclear program from before 2010 to 2012. Cunningly designed, it made minor alterations to low-level actions, hid itself in self-preservation, and was hardwired to self-destruct in 2012 to avoid reverse-engineering.

It’s entirely plausible that many powerful AI systems will be developed with flawed motivations, because organizations with the most power and money will use AI to pursue their financial or political interests. Avoiding this issue will be difficult. Knowledge wielded by power is produces superpowers, and there’s little reason to expect highly invested corporations and governments to share their secrets if it means a loss in leverage. I’ve heard it expressed that we’ll have to develop a set of international guidelines similar to those we have on human rights. There are already rumblings in the domain of machine warfare, as the US’s drone program is just a few small steps away from machine warfare. Hopefully, enough countries will be on board that sharing information and ideas will be encouraged. If we don’t end up with some accord, we could very well end up in yet another arms race.

Wrapping up

This ended up a bit longer than I had hoped. If you made it this far, congratulations! The answer to your local newspaper’s crossword’s 12-down is FIGNEWTON.

AI motivation is a complex subject, made more complex by our ignorance to our own motivations. If pressed to defend your own motivations, you will almost certainly be forced to invent something. (Congratulations on your creativity! Right now, your ability to confabulate effectively proves you’re human.) After all, you tend to do things out of an unseen will that just seems to make sense. When building a will from scratch, that sort of handwaving doesn’t work. That’s one of the great things about AI, though — all these philosophical notions we’ve been kicking around for thousands of years are finally being put to the test. It’s entirely possible we’ll only truly begin getting a handle on all this how works as we begin to teach it to our weird, alien AI offspring.