nevermind the a.i., beware of human

With easy, cheap access to cloud computing, a number of popular artificial intelligence models computer scientists wanted to put to the test for decades, have now finally able to summon the necessary oomph to drive cars and perform sophisticated pattern recognition and classification tasks. With these new probabilistic approaches, we’re on the verge of having robotic assistants, soldiers, and software able to talk to us and help us process mountains of raw data based not on code we enter, but the questions we ask as we play with the output. But with that immense power come potential dangers which alarmed a noteworthy number of engineers and computer scientists, and sending them wondering aloud how to build artificial minds with values similar to ours and can see the world enough like we do to avoid harm us by accident, or even worse, by their own independent decision after seeing us as being “in the way” of their task.

Their ideas on how to do that are quite sound, if exaggerated somewhat to catch the eye of the media and encourage interested non-experts in taking this seriously, and they’re not thinking of some sort of Terminator-style or even Singularitarian scenarios, but how to educate an artificial intelligence on our human habits. But the flaw I see in their plans has nothing to do with how to train computers. Ultimately an AI will do what its creator wills it to do. If its creator is hell bent on wreaking havoc, there’s nothing we can do other than stop him or her from creating it. We can’t assume that everyone wants a docile, friendly, helpful AI system. I’m sure they realize it, but all that I’ve found so far on the subject ignores bad actors. Perhaps it’s because they’re well aware that the technology itself is neutral and the intent of the user is everything. But it’s easier just to focus on technical safeguards than on how to stop criminals and megalomaniacs…