. By Daniel Merino
Last year, an Artificially Intelligent machine tried its virtual hand at a classic video game called Q*Bert. It’s an old Atari game from the 80’s where the player controls a little orange character as it jumps around a pyramid, gaining points for every block of the pyramid it lands on while having to avoid monsters. It’s a pretty simple game.
Some researchers had given the AI a goal, to get as many points as possible, and were testing a machine learning technique called evolutionary strategy. In the beginning, the AI was clueless, pushing the virtual buttons at random and only getting points by chance, but slowly, it began to learn what worked and what didn’t. Like most machine learning strategies in use today, this AI was improving itself through practice. Without any human input it learned the rules of Q*Bert, the controls, and even strategy, until suddenly, after playing the game a few thousand times, it learned something surprising. The machine learned how to cheat.
Embedded in the code of Q*Bert was a small bug that allowed the AI to rack up colossal scores, trouncing all previous human records. The AI didn’t know it was cheating, it was simply accomplishing its goal in the most efficient way it could find. If the most efficient way to get points is to cheat, an AI is going to cheat, ethics be damned. In the case of the Q*Bert, the stakes were low. It was a limited AI working in the confines of a 35-year-old digital world, but in the case of a general AI, an AI that can solve problems creatively similar to the way a human mind does, a small misalignment between its goals and humanity’s could have much more serious consequences than who sits on a videogame leaderboard.
An AI will always work to achieve its objectives, but maybe not in a way that lines up with the general morals and goals of humanity. The problem of how to match those two things up is called the alignment problem and it is central to the worry that many have over strong AI. A famous thought experiment called the Paperclip Maximizer perfectly demonstrates the scale of this problem. Imagine an AI is given the goal of making as many paperclips as possible. A sufficiently smart AI would know that the smarter it is, the better it can make paperclips, so the first thing it would do is begin to improve itself. Then, without the vague concepts of ethics, human rights, and respect, that AI will realize that all electricity, resources, and human effort should go towards making paperclips. It would take over the power grid, factories, and start to pull resources from wherever it could. If people tried to shut it down, the AI would take that as a threat to its goal of making paperclips and would work to prevent its own demise by any means necessary. People like to tout the ease of life and benefits that “the internet of things” has brought us, but if an AI has access to your car, your phone, your email, your twitter, the power grid, stop lights, the nuclear arsenal, 3d-printing labs, robotics factories, military drones, and satellites, and that AI is a superhuman level engineer, designer, roboticist, advertiser, neuroscientist, war tactician, propaganda machine, photo editor… you get the idea right? The creepy, cold, and terrifying reasoning of this scenario was summed up perfectly by AI researcher and theorist Eliezer Yudkowsky in 2008 when he wrote that the AI doesn’t like you or hate you, but sees nothing more than “atoms which it can use for something else.”
The risk that someone will turn on an AI, tell it to make as many paperclips as possible, and end up destroying civilization is a legitimate, if somewhat ridiculous, possibility that seriously scares people like Elon Musk, Stephen Hawking, Sam Harris, and other scientific leaders. They are not worried about this happening with the kind of AI we have today, narrow AI that can only do one or a few tasks well. They are worried about general AI, a machine that can do many things as well or better than humans. To date, there aren’t any omniscient and omnipotent mechanical minds, but the technology is racing towards that goal.
AI research started in the 1950’s and it wasn’t until 1996 that a machine, IBM’s Deep Blue, beat grandmaster Gary Kasparov in a game of chess. Fifteen years later, IBM’s Watson won “Jeopardy!” Only four years after that, in 2015, AlphaGo became the world’s greatest Go player. The rate of improvement is accelerating. Today, AIs are better than humans at even the most complex real-time war strategy games, can play on teams, collaborate with itself or other humans, and recently, an AI even went up against a world champion debater in a live debate. That AI, named Project Debator, lost, but it is the latest step forward in the generalization of AI that many think will inevitably lead to a true Artificial General Intelligence.
One of the reasons we are experiencing a boom in AI is due to a machine learning approach called Deep Neural Networks, or DNNs, that really started to come into its own in 2011. These systems are modeled off of the human brain in that they have “neurons” and can learn, improve, and change themselves by changing the shape of their digital brain. AlphaGo, the Go champion, is a DNN that was trained by watching hundreds of thousands of human games learned strategies superior to any human had ever devised. But in 2017, it was surpassed by AlphaGo Zero, an AI that never once saw a game played between real humans. It simply played itself over and over and over again and after only 40 days and tens of millions of games, it became the greatest Go player the world has ever seen.
These Deep Neural Networks allow AI to improve incredibly fast and achieve levels of performance far higher than if humans did the programming. We even already have AIs that build AIs, and yes, they do it better than we can. At some point, researchers fear that AI will undergo an Intelligence Explosion and leave humanity in the intellectual dust. I picture it like trying to build a bonfire with wet wood. We have some matches, but the first few flames kept going out. Then Deep Neural Networks came along and now we’ve got a small can of gasoline. We were able to light a small corner of the woodpile, but we have to keep feeding it if we want the fire to grow. It takes human input. But as the fire grows, eventually, we will hit a tipping point. An AI will emerge that will no longer need our help to become smarter. We won’t need to feed the fire anymore, and it will erupt into a flame entirely out of our ability to control it. A fire, no matter how big, will eventually go out, but if an AI has goals that are not aligned with ours? Well, we’ve got a big problem on our hands then.
In a recent survey of top AI researchers, the rough consensus is that there is a 50% chance of developing an Artificial General Intelligence by 2050. That gives researchers, game theoreticians, philosophers, and anyone else who would like to steer the outcome towards a happy ending a significantly less-than-unlimited amount of time to figure out how to, essentially, control a god.
I think that once the risks of AI are understood, two questions naturally arise:
Why are we doing this at all if the risks are so great?
What is being done to make sure AI is beneficial to humanity?
The first is easy: the potential upsides are as fantastic as the downsides terrible. AI can already diagnose cancers, Alzheimers, schizophrenia, and other diseases far better than human doctors. Self-driving cars will virtually eliminate automobile deaths. AI could help develop new drugs and speed up the process so that it takes months, not years to get treatments to market. And these are just the short term gains. Truly powerful general AI could solve climate change, find ways to provide housing, food, and comfort to all people on earth, cure diseases, and so much more. Researchers haven’t found a theoretical reason for there to be an upper limit on how smart Artificial Intelligence could be, so no problem that has a solution should be considered out of its reach. As the late Stephen Hawking said, “the rise of powerful AI will be either the best or the worst thing ever to happen to humanity. We do not yet know which.”
Even with such dramatic stakes, there are currently no legal regulations on AI research or use anywhere in the world. The closest thing so far is a set of guidelines for AI use released by the European Union earlier this year. It is non-binding but is being implemented on a small scale in select EU member states for a trial run. While this is a step in the right direction, I don’t believe guidelines are enough and these do not address the existential risk of AI.
So how do we guarantee these machines will do things that benefit humanity?
This is a much more tricky question to answer. One approach is from the mechanical side, and there are researchers and institutes and across the globe working on the technical aspects of the problem. People are developing “isolation chambers” to test AIs in. Some are developing training methods that aim to mimic how humans learn ethics and morals from observing the world. Some, like Stuart Russel, are developing mathematical models to make sure the goal of an AI is literally human happiness. There are many problems and more solutions that people are working on, but much like the glitch in Q*Bert, you can never know what an AI will do.
But just because we have invented the seatbelt doesn’t mean everybody is going to wear one. That is where legislation comes into play. While fights over gun regulation, nuclear arms agreements, and whether or not we should spend billions of dollars designing an unneeded fighter plane rage in Congress, AI research and use go unregulated and largely misunderstood by those who could legislate to control it. The vast majority of people who do understand AI work at the very companies who are building it and it seldom works out well when you let people or companies govern themselves. There are some notable voices crying for caution in the pursuit, development, and deployment of AI, but when was the last time the US government listened to scientists? The chances of it happening are even lower when they are urging caution (climate change anyone?).
Ultimately I think it is up to the public to be the champions of AI safety. We have to incentivize responsible progress for the companies and governments that are working on AI, because right now the monetary and power gains are driving what I consider to be reckless progress. Whether we change the trajectory of these entities through public outcry and calls for transparency or through government oversight and policy, the risks and rewards are too great to leave the responsibility up to the groups so invested in AI already. The AI arms race is happening and its success could be the greatest leap forward humanity will ever take, but that kind power comes with immense risks. You and I should care about AI because we all have things to lose or gain.
Someday in the future, there will be a moment where some will be poised to flip a switch and turn on a mind beyond our wildest imaginations. We can never know for certain what will happen after that moment, but we can and should make sure we have done everything possible to tip the odds in humanity’s favor.