Note: The following are excerpts from Scary Smart: The Future of Artificial Intelligence and How You Can Save Our World by Mo Gawdat.
A couple of years or so before the yellow ball, Google had acquired DeepMind. Back then, the brilliant Demis Hassabis (CEO and founder of DeepMind) stood before the senior leadership group of Google to present to us the technology they had developed. This was the time when they taught AI to play Atari games. It’s not a huge stretch to spot the connection between the way machines learn and the way children do when the demo that is shown to you is of a machine playing a game. But I still missed the core message and instead marvelled at how far tech had come since I stopped writing code a few years earlier. Demis showed a quick video of how the AI machine – known as DeepQ – used a concept called deep reinforcement learning to play the famous game Breakout, a very popular Atari game where the player uses a ‘bat’ at the bottom of the screen to bounce a ‘ball’ back up at a bunch of rectangles, organized in what looks like a brick wall at the top of the screen. Every time the ball hits a brick, it disappears and is added to your score. The quicker you can manage to go through all the bricks, the higher the score you can achieve. ‘We gave the computer nothing but access to the pixels on the screen and the controls. We did not give it any instructions on what the game was about or how to play,’ Demis said. ‘We then rewarded the machine through an algorithm to maximize the score.’
He showed a video of DeepQ after it had been trained by playing 200 games, which took just one hour because they could use several computers playing in parallel to accelerate the learning. DeepQ was doing pretty well already by then. It could hit the ball even when it came down fast, say, 40 per cent of the time. We were impressed because we had never seen that done by a machine before. He then showed another video after just one more hour of training, and by then our artificially intelligent child was far better than any human who had ever played the game. You could barely see the ball because of how fast it moved and, without fail, almost every ball was hit back, regardless of the angle it bounced at or the speed at which it was hit. A world champion in just two hours of training. The team didn’t stop there, Demis said. They trained DeepQ for one more hour. And the miracle happened. The AI figured out some of the secrets of the game. As soon as a level started, it focused on creating a hole in the wall to sneak the ball between the roof and the bricks. This technique is known to be the best way to finish every level quickly and safely. Once again, the three-hour-old kid, DeepQ, had learned. It had learned fast and, astonishingly, it had learned things we never attempted, or intended, to teach it. Of course, DeepQ, like other children, did not stop at Breakout. In no time at all, it was the master – the world champion – of hundreds of Atari games. It used the same logic that it had developed to ace every other game that was offered to it. Impressive in every possible way. And if this is not awe-inspiring enough, let me finish this chapter off with another astonishing fact that is often skimmed across quickly when we speak about how AI learns. Sit down, please, before you read this.
We truly have no clue how they learned this. We have no track of the logic they followed and we often don’t even have a way of finding out.
You see, the more intelligent a being becomes, the more it includes a widely diverse set of topics into the formation of its moral code and the deeper it contemplates each topic and principle. The code of ethics followed by a tiger is minuscule, simple and clear, as compared to the complex code humanity constantly debates, rarely ever agrees on and almost never fully adheres to. This also applies across the diversity of us humans. Those of us who live simpler lives, such as the traditional tribes in less economically developed parts of the world, tend to have a clearer, though simpler, code of conduct than graduates of Harvard Law School. Clearly, intelligence allows for the rise of morality simply because it supplies the brain resources needed to reflect on the more intricate concepts of where the line between right and wrong resides. But that doesn’t always make more intelligent beings more ethical.
It seems that the more intelligent a human becomes, the more they allocate their intelligence to search for loopholes and roundabout ways to ensure they maintain acceptance in their community not actually by being entirely ethical but rather by appearing to be so. I refer to this dip in the chart as a momentary lapse of reason. A foggy moment when our intelligence fails to observe the truth about the value that ethics bring. As a being becomes even more intelligent, it eventually understands that the best way to navigate life is a straight path that promotes the well-being of others and not just oneself. Gandhi’s non-violent resistance, clearly, is a smarter path than the atomic bombing of Hiroshima and Nagasaki, and yet we often resort to war. Why? Is it because we are not smart enough? No. It is because we are motivated by the wrong targets.
Our modern world imposes targets that are often given higher priority than conforming to ethical values. Economic gains, assertion of power, expansion of territory, re-elections, wealth, getting likes on Instagram – these are just a few of the competing modern-day objectives. When these come into play, sadly, human intelligence is no longer concerned with questions of morality and ethics. Instead, intelligence gets deployed to its full capacity in the direction of achieving those objectives while still appearing to be ethical. The question, ‘How do I get away with it?’ becomes the primary focus. Some of the smarter ones amongst us just get really good at hiding, lying, finding loopholes, debating and morphing the moral code.
Often those who are the smartest are in some sort of position of power that can impact the lives of many others, so their actions can wipe out the acts of millions who adhere to the ethical code. With the aid of mainstream and social media, the acts of those few paint a picture that makes humanity at large seem corrupt. If you judged humanity by the devastating decision to drop a nuclear bomb on a civilian population, you would have every right to lose hope in it, but in doing so you would be forgetting that it is only a small minority of people who broke the commonly agreed ethical code – to preserve the lives of innocent civilians. They replaced it with other codes, such as ‘it’s advantageous for my country to demonstrate its power’ or ‘collateral damage is acceptable when at war’. Once the code is broken, a whole load of intelligence then goes into propagating the message that killing 220,000 overwhelmingly innocent people is necessary to stop the war. We believe the propaganda and forget that two wrongs – a war and a nuclear attack – don’t make a right. We start to question in our minds whether that headline could actually become the new code.
Is it even possible to kill an AI? Today, if you took a hammer and smashed a computer, your action would be considered wasteful but it is not a crime. What if you kill an AI that has spent years developing knowledge and living experiences? Because it’s based on silicon, while we are based on carbon, does that make it less alive? What if we, with more intelligence, managed to create biologically based computer systems, would that make them human? What you’re made of (if you have intelligence, ethics, values and experiences) should not matter, just as much as it should not matter if your skin is light or dark. Should it? How would the machines that we discriminate against react? What will they learn if we value their lives as lesser than ours?
What if the machines felt that the way we treated them was a form of slavery (which it would be)? How do slaves react to power and authority? Humanity’s arrogance creates the delusion that everything else is here to serve us. Like the cows, chickens and sheep that we slaughter by the tens of billions every year. What if cows became superintelligent? What do you think their view of the human race would be? What will a machine’s view of the human race be if it witnesses the way we treat other species? If it develops a value system that prevents humanity from raising animals like products to fill our supermarket shelves and restricts us from doing it, would we think of this as a dictatorship? Even if we wanted to treat the machines as equals, how could we when they are so different from us? As an example, take the difference in our perception of time. Things are much slower for us than they are for a machine. What if a choice was to be made between rescuing a machine or a human, for example, a self-driving car or its passenger after falling into a lake? Who do you rescue first when a millisecond of suffering for a machine is equivalent to ten years of suffering for a human? Why should we value the life of the human more in the first place
How about reproductive freedom for machines? Would there be a one child per family policy? If we don’t restrict their reproductive abilities, what would prevent them from creating trillions of copies of themselves in seconds and outnumbering us with our nine-months-long reproductive process? How would they feel if we prevented them? How would they feel if we killed one of their offspring?
If we become cyborgs, as Elon Musk predicts, and extend our intelligence by connecting it to the intelligence of the machines, would we then value the machines supporting the rich more than those integrated with the poor? How would the poorer machines feel? Would the poor have the resources to integrate with a machine at all? Would it be ethical to create this new form of digital intelligence divide? Can you imagine what the relationship between Donald Trump’s machine and Vladimir Putin’s machine, if those existed, might have been?
How about virtual vice? When an AI that works as a romantic sex robot, of which there are surely many primitive examples today, is raped, should we punch the perpetrator? If we don’t, what will we be teaching the robot as a result? What are we teaching that robot about humanity by inventing it in the first place? Are there vices that are wrong for humans to commit with humans but okay with AI? Who makes that choice? What if the intelligence of those robots informs them to adopt other vices since it seems to be okay for them to adopt some? What if we made a robot to fulfil the desires of someone with a submissive sexual preference, would the violence exerted by the AI then be okay? Would that AI then not go out of its way in an attempt to convince other humans to be submissive? What if it was clever enough to convince us? Would that be okay?
Then there is the actual definition of vice, which seems to blur drastically on the internet where bullying, porn, narcissism and pretentious lies appear to be accepted in ways we don’t approve of in the physical world. For that to be the primary source of information for today’s young machines, what do you think their perception of normal is going to be?
And let me add just one more example. What about AI’s work ethics? Dr Ben Goertzel calls these: selling, killing, spying and gambling. Shocking as this sounds, it is true. Most of AI’s investment today is focused on performing tasks related to these four areas – though, obviously, they are called by different names such as ads, recommendations, defence, security and investment. Those young machines are developing all the intelligence they can to excel in the exact tasks we’ve assigned to them. We criticize child labour and feel appalled by the thought of child soldiers. Well, welcome to the world of artificial child trauma at its extreme.
These, and thousands more, are not only complex questions but questions that we have never had to ponder before. This is why I left them all in question form. I want you to reflect, as I do, simply because I don’t know the correct answer to any of them and I honestly don’t expect you to know either. The breadth and complexity of the ethical dilemmas we’re bound to face are endless. And we’re supposed to try and resolve them within the next ten years. The answer to how we can prepare the machines for this ethically complex world resides in the way we raise our own children and prepare them to face our complex world. When we raise children, we don’t know what exact situations they will face. We don’t spoon-feed them the answer to every possible question; rather, we teach them how to find the answer themselves.
The next time you get the ‘Are you human?’ test on a website, you are not only proving that you are human, but by providing an answer you are also building a test for the student bots. Have you been seeing lots of questions about traffic lights and pedestrian crossing lines lately? Of course you have. These are being used to train self-driving cars by collecting billions of traffic-related photos that you and I are recruited, for free, to identify. The first students that survive are just lucky to have started with a random code that is slightly better than those who were not so fortunate. Soon, however, the value of luck fades as keeping only what works, and attempting to improve millions of copies of it, eventually yields a student bot that can actually tell 8s from 3s. As this talented bot is copied and changed, slowly the average test score rises, leaving only the smartest of all students to survive. Eventually, the infinite slaughterhouse of student bots yields a few that can tell an 8 from a 3 in a photo they’ve never seen before, with a level of accuracy that beats our human abilities. Welcome to the future!