Scientists: We can’t control a superintelligence
How could we, when we can’t even control social media?
This past week has seen a lot of buzz about an article in Journal of Artificial Intelligence Research with an ominous conclusion bang in its headline: Superintelligence Cannot be Contained: Lessons from Computability Theory.
The core of the article is the Control Problem, which has occupied many an AI researcher. To be fair, most of them are busy trying to build something approximating intelligent behaviour in a limited sense, but some are gazing further ahead. Nick Bostrom, philosopher at Oxford University, is one of them. He made a splash six years ago with the book Superintelligence, which considers how a general AI surpassing the intelligence of mere mortals could pose an existential risk to mankind.
If, some day, we construct an artificial general intelligence that learns on its own, scouring the internet for resources, recruiting assets outside of the net to act upon the physical world, we may quickly find ourselves in uncharted territory. If the AI can improve its own mind through programming skills vastly superior to ours, we could see an intelligence explosion that leads to what Bostrom calls superintelligence.
With superintelligence comes a big question: How can we make sure such an intelligence doesn’t decide to wipe out humanity?
Also a disquieting answer: We can’t.
While this may seem self-evident to most people who take five minutes to think about what a superintelligence would look like – and realize we can’t really know – the authors of the JAIR article frame their argument in terms of Alan Turing’s Halting Problem. It’s an interesting excercise, although their conclusion should surprise no one: A superintelligence cannot be contained. Why? We cannot simulate the consequences of all the actions it might take, hence we cannot prevent it from possibly causing harm to humans.
The argument is sound but it is also futile. As the paper ends, the authors stumble on the reason why, without realizing it. Under the final section, “Discussion,” the authors comment on the state of the world in terms that will leave anyone who follows the news puzzled:
Today, we run billions of computer programs on globally connected machines, without any formal guarantee of their absolute safety. We have no way of proving that when we launch an application on our smartphones, we would not trigger a chain reaction that leads to the transmission of missile launch codes that start a nuclear war. Indeed, in 1965 Arthur C. Clarke wrote a short story (Dial F from [sic] Frankenstein) warning us that, as soon as all the computers on the Earth were connected via telephone, they would take command of our society. Yet, today, we still use our smartphones everyday, and nothing has happened.
“Nothing has happened” is as strange a take on technology’s impact on the modern world as I’ve seen. The billions of phones and computers connected through the internet today have fundamentally changed our world, and our lives. They have done so for better and for worse. While the good may outweigh the bad, no one could argue that, for instance, social media has done no harm. The harm is palpable; at best the positives outweigh the negatives.
The reality is we did not understand the potential harm of social media back when Myspace and Facebook were started. Nor did we understand it as social media matured and Facebook, Twitter and Youtube became global platforms, running on algorithms designed to manipulate the behaviour of their users. Today we can clearly see some of the harm, for instance when these platforms are used to spread disinformation and conspiracy theories, to lie and to bully and to incite deadly mobs as in Myanmar and in the US. But we still have little idea of just how harmful the algorithms of social media platforms are. Or how we can make them less harmful.
This is the real argument here: We can’t even control our current technologies. We obviously can’t control a potential technology far more powerful and inscrutable, per definition beyond our powers of understanding. Discussing whether it is theoretically possible to control a superintelligence is a fun academic exercise, but it has little bearing on the world we live in. Even so, hey, super-powerful AI that may wipe out humanity – it sure makes for clickable headlines.
On a final note, it is funny that the paper mentions Arthur C. Clarke’s story Dial F for Frankenstein, if only to dismiss it as a warning of something that did not come to pass. What they fail to mention is that Clarke’s prescient story was in fact an inspiration for Tim Berners-Lee, as he conceived the World Wide Web.