Righteous Robots: The Ethics of Cognitive Computing

The Quick & Dirty

Theorists, philosophers, and writers have worried of a cognitive computing doomsday from the very beginning.

Would a cognitive computer in charge of a dam system sacrifice human lives to uphold its duty?

Would a robot asked to optimize paperclip manufacturing turn the entire planet and its inhabitants into a near-infinite number of paperclips?
 
 Can we program human values into a computer?

The Bigger Picture

The world met the word “robot” on January 25, 1921, at the premiere of Czech sci-fi writer Carel Kapek’s play R.U.R. (Rossum’s Universal Robots). The word comes from the common Slavic-language root for “work” or “labor,” so the word “robot” was meant to denote simply “workers.”

In the play the roboti (that’s the Czech plural for robots) are synthesized from organic matter, so they’re more like clones than what we think of as robots. And like any good dystopian sci-fi would have it, everything goes swimmingly with the roboti. Until it doesn’t. Followed shortly by the extinction of the human race.
 
 Which means that as long as the concept has been around, it’s been married to foreboding premonitions, not unlike The Monkey’s Paw or Pandora’s Box.

But whence do these fears arise? Other than for economic reasons, why would people who hate their jobs tremble so portentously at the thought of a machine doing it for them? Are these fears counterproductive, or do they point to a real potential threat?

Brad Allenby, writing for Slate, put forth a clear-eyed summary of what it is we actually fear in his article The Wrong Cognitive Measuring Stick:

…much of the dystopian hysteria around A.I. reflects the fear that it will act as humans act (which is to say violently, selfishly, emotionally, and at times irrationally) — only it will have more capacity. In essence, much of what we fear is a much more competent [evolved intelligence, i.e. human intelligence].

Kapek described robots more as clones, though our current definition is more like sophisticated automatons or capacious and perceptive computers.

Much hooplah was stirred when Elon Musk & Co. published their open letter and formed the Future of Life Institute to guard against, in Musk’s words, the “existential threat” of artificial intelligence.
 
 But while they express worries over the impending material threat of cognitive computing, their main concern is not the violence wreaked by autonomous assholes but the terrible misallocation of R&D toward defense over the myriad other sectors that could improve our lives and the likelihood of an arms race for automated weapons, leading to a black market boom.

Stephen Hawking, in a recent Reddit Ask Me Anything, stressed a different concern about building a specialized cognitive computer that’s better at it’s job than humans:

The real risk with AI isn’t malice but competence. A superintelligent AI will be extremely good at accomplishing its goals, and if those goals aren’t aligned with ours, we’re in trouble. You’re probably not an evil ant-hater who steps on ants out of malice, but if you’re in charge of a hydroelectric green energy project and there’s an anthill in the region to be flooded, too bad for the ants. Let’s not place humanity in the position of those ants.

This is a valid worry that addresses both the excitement and the good ol’ Kapekian terror in the face of building something that’s better than humans.

But in some ways, it presupposes (or merely warns of) the mistake of foisting some serious responsibility on technology before passing through a decent interval in which man and machine work together.

Thomas H. Davenport and Julia Kirby, in their June 2105 essay “Beyond Automation” in the Harvard Business Review, seem to address this worry while also consoling those would-be-replaced workers:

The strategy that will work in the long term, for employers and the employed, is to view smart machines as our partners and collaborators in knowledge work. By emphasizing augmentation, we can remove the threat of automation and turn the race with the machine into a relay rather than a dash. Those who are able to smoothly transfer the baton to and from a computer will be the winners.

Similar to Hawking’s worry, an article by Edward Moore Geist at The Bulletin of Atomic Sciences, discussed an example culled from Oxford philosophy professor Nick Bostrum’s recent book Superintelligence, which is both fun and horrific to consider:

In one particularly lurid example, Bostrom paints a picture of how a superintelligent, but otherwise unconscious machine might interpret a seemingly innocuous goal, such as “maximize the production of paperclips,” by converting the entire earth, as well as all additional matter it could access, into office supplies.

The “intelligence explosion” is the crux of Bostrum’s doomsday theory: when we succeed in creating a machine that can learn hundreds, thousands, millions, billions of times faster than humans, then in a matter of hours, or at most days, it could likely evolve into humanity’s ultimate rival while possibly remaining the ally of just one or two people who are inevitably jerks.

What seems obvious is the need for building a value system into cognitive computers, one that prioritizes human life and environmental issues and takes its tasks and its own existence as subsidiary.
 
 But could we make a computer that prefers human life and the greater good over duty?

If you don’t agree with Bostrum’s technological optimism, however, then you won’t agree with his endgame pessimism, which is where the argument might break down for lack of a serious audience. Geist counters Bostrum’s optimistic pessimism in that same article:

But over the course of 60 years of attempts to create thinking machines, AI researchers have come to the realization that there is far more to intelligence than simply deploying a faster mechanical alternative to neurons. In fact, the history of artificial intelligence suggests that Bostrom’s “superintelligence” is a practical impossibility.

If the contention over superintelligence is accepted as a given to make room for a different contention, the question then becomes: Can we instill human values in a computer brain?

But are human values limited to a set of behaviors, or does it go beyond into the realm of empathy and “common sense,” whatever that is? Can we program our robots to open the door for people instead of barreling ahead? What about waiting for passengers to get off the subway before it gets on? Can’t we just feed it Dale Carnegie and Emily Post and hope for the best?

A broadly functional superintelligence would need the ability to desire before it developed the capacity for the kind of relentless pursuit Bostrum describes.

For example, AlphaGo can whip butt at Go, but it doesn’t know what it’s doing and doesn’t have the capacity to care if it wins, loses, or forfeits in the middle.

Which leads to three possibilities: Superintelligence with

1. No desire
2. Ethically hierarchical desire (i.e. value human life first, environment at close second, defer to human decision makers depending on the clarity of a decision’s impact either way, etc.)
3. Amoral desire

Only #3 would lead to a Bostrum psycho-verse because it would resist being overridden, but is such a beast even theoretically possible? Is it reasonable to assume we have a 33.3% chance of the end of the human race or are the above options not equally likely?
 
 Whatever the answer, I just hope I’m on the Singularity’s good side.

Join us next week as we dig into the future of Natural Language Processing and the Writing Robots