The Genie Problem

Posted on 2018 November 25


the genie problem johnny-automatic-genie-s-lamp

It would be much better to call a halt in material progress and discovery rather than to be mastered by our own apparatus and the forces which it directs. — Winston Churchill

In the debate about Artificial Intelligence — will it be a good thing? Will it bring destruction? — there’s a big divide between optimists and pessimists. Optimists believe AI will usher in paradise for us all. Sure, there’ll be a bump or two along the way, but basically AI will solve all our problems, and the doubters are scaredy-cats. Pessimists fear AI will usher in doom. AI will be so vastly smarter than the rest of us that we’ll have no way to control it, and when its work conflicts with our other goals, the work will prevail, perhaps at the cost of humanity itself.

That sounds way too pessimistic, right? How could a computer system destroy humanity? We wouldn’t let that happen! As one optimist claimed, “We can always unplug them.

…Or can we?

My views fall neatly into the center of the debate, where hope for a wonderful AI future is balanced by worry that its immense power will veer off course and go horribly wrong. Already AI systems have created benefits in driverless vehicles, medical diagnoses, ophthalmology, robot surgery, investing, airport management, product fulfillment, Siri, Alexa. The worst dangers, on the other hand, haven’t yet shown up.

Our challenge involves “Superior AI”, the kind that is so smart — vastly smarter than people — that it can begin to solve major problems that have stymied us. The trick is to get AI to do its work without doing damage. We must program it from the outset to avoid harm. But how can we achieve this without hamstringing the process? “Hey, AI, do what we ask without causing harm.” The AI complies and promptly shuts down. Huh?

Every activity causes at least some modicum of damage. When you walk down the sidewalk, you crush untold billions of microbes. To eat requires agriculture, which slaughters animals or deforms natural habitats. Merely going to work or socializing leads to emotional bumps and bruises. How is an advanced computer going to navigate such moral quicksands?

Even if we could solve this, we’d have to correctly program the AI on the first try. Otherwise the system will quickly outrun us mentally, so that we may have trouble corralling it to fix any problems. The AI needn’t be malevolent but merely literal in obeying its programming to wreak accidental ruin that we can’t stop.

More worrisome: a lot of AI research is for the military, which would like nothing better than battalions of robots programmed to kill the enemy. The uber-AI that commands those androids might discover it needs to spread out across the Internet to improve its work. Suddenly malevolent AI clones are everywhere in daily life, causing unstoppable peacetime damage.

Many corporations are competing to be first off the mark with Superior AI. Imagine the power a company could wield over its competitors — or the government itself — were it to possess the greatest intelligence ever known. And let’s not forget the lone hackers, one of whom might someday launch a rough version of Superior AI whose flaws quickly spiral out of control.

In short, humans basically have one chance to get Superior AI right. Any mistakes will be engraved in the AI’s software, and in our future, forever.

These dilemmas can be summed up in what I call “The Genie Problem”. It goes like this:

One day, on a hike in the desert, you find — half-buried in the dirt — a teapot-shaped lamp, grimy from decades of exposure to the elements. You pick it up and dust it off with your hand, rubbing its smudged silvery surface. Suddenly smoke pours from the lamp and a genie appears, turban atilt, muscular arms folded. (He looks like Mister Clean except with better haberdashery.) The genie announces in a foreign accent that he will grant you three wishes.

This is great! First, you wish for ten million dollars. The genie nods, twitches his nose, and claps his hands. A minute later, while idly counting up the ways you’ll spend your newfound fortune, you receive a message on your phone — great reception out here! — announcing that your parents and siblings have died tragically in a boating accident, and you are entitled to a life insurance payout and an inheritance that add up to … you guessed it … ten million dollars.

This is not so great. Stunned, but hoping for good to come of the genie’s promise, you make your second wish: “I want Peace on Earth.” The genie does the nose twitch and hand clap. You look around: nothing has changed. You shrug, put the lamp in your knapsack, and trudge back to your car. The genie walks at your side.

At the trailhead you find your vehicle and drive off. On the way home you notice there’s no one talking on the radio — music here and there, but no deejays. When you reach the highway, you see no traffic, but instead cars perched at crazy angles; a couple have rolled over. You stop to investigate: there are no injured or dead, because there are no people in the cars. Back in town, the roads are empty except for the odd collections of stalled vehicles. On Main Street no people are outside and none are in the stores. You get home and walk in: “Hello? Anybody here? I’m back!” No one answers.

Frantic, you search everywhere, make calls, check Facebook, but no one responds. Finally you realize what has happened. It was your wish for “Peace on Earth.”

The genie has followed you everywhere, awaiting your final command. You know what your third wish will be — to undo the first two. You hope, you pray, that you will ask it correctly, because …

You have one chance to get it right.