About AI Safety

So one of the current hot topics these days about artificial intelligence seems to be AI safety: How can we make sure that AI will actually be beneficial?

Some top thinkers from science and technology (namely Elon Musk, Stephen Hawking, and Bill Gates) have expressed their concerns about the dangers of artificial intelligence. If those three guys are worried, shouldn’t we all be?

Well, it first of all depends on what you’re talking about: current and near-future narrow AI systems or full-fleshed AGI systems (for a short distinction of these two things see my old blog post).

If we talk about the second one, then one can see in many science fiction movies (think of “Terminator”, “Matrix”, or “I, Robot”) that superintelligent AI could be a threat to us. Just as the fate of apes lies now in the hand of us humans (who are more intelligent), our fate could at some point lie in the hand of intelligent machines. And there is no guarantee that they will treat us nicely (just as we ourselves don’t treat animals that nicely, or at least we didn’t for quite some time).

So the question that arises there is basically the following: How can we make sure that any superintelligent entity will still share our goals and values and will not turn against us? If the idea of superintelligent machines is too abstract for you, think about a potential drug that would make a human 10 times more intelligent (like in the movie Limitless): Before giving that drug to someone, what can you do to make sure that they won’t use their improved intelligence against you? Once they’ve taken the drug, they’ll be able to easily outsmart you, so this is a tough problem.

If you’re interested in these kinds of thoughts, I can highly recommend the book “Superintelligence” by Nick Bostrom which discusses these issues in great detail.

Okay, but now back to my original distinction between current narrow AI and future AGI: Clearly, with respect to future AGI, there are some problems down the road. Although these dangers are not imminent, yet, it makes sense to start thinking about them.

On the other hand, we’re already dealing with narrow AI systems today – and they can also cause trouble (although probably not the “extinction of humanity” kind). Take for instance the incident with Tesla’s autopilot: A man using Tesla’s drivers assistance system died because this system did not detect a tractor trailer in front of it – the car crashed into the trailer and the driver died. AI is already in many places today and if an AI system makes a wrong decision (based e.g., on a software bug), the consequences can be dramatic.

In my personal opinion, narrow AI can only do as much damage as we let it do: If we give an AI a certain amount of power (like driving our car, doing trades for us on the stock market, filtering job applications, or controlling nuclear war heads), we should always be aware of the possibility that this narrow AI system malfunctions. In order to detect this (so we can react in time and “fix” a wrong decision), we need to closely monitor these systems to make sure that nothing goes wrong. As we all know: “Everything that can go wrong, will go wrong.” So the moment we give an AI system a certain amount of power, we have to be aware of what can go wrong in the worst case and how we could deal with that.

There’s this other quote from the Spiderman movies: “With great power comes great responsibility”. I think this also applies to artificial intelligence: If we give power to AI systems to do stuff in the real world and to influence our lives, it is our responsibility to make sure that we can stop the system in time if things start to go wrong.

Leave a Reply