I have a cheap WiFi enabled vacuum cleaner. One morning, I barked "ALEXA! CLEAN THE HOUSE." The gynoid immediately responded with "Now playing songs by Crowded House." The future's ace, innit?
So I opened the cleaner's app, hit a button, and heard the sound of a domestic-droid doing its thing. For two minutes. The app spat out this fabulously helpful warning.
I trundled downstairs to see what precarious sinkhole the mop was teetering on the edge of. Only to be met with this.
My cleaner snoozing in a little puddle of sunlight like a domestic cat! Weird. I pressed the button and it refused to move. I gave it a little nudge, out of the sunlight, and it started up again! Only to get confused moments later...
I observed this happening a few times. The robot gets genuinely bewildered by stark differences between light and shade. To the point where it will follow a shadow thinking it is a wall.
This is a £130 cleaner - it doesn't have fancy multi-sensor AI with a Machine-Learning neural network. It's barely above the if-else-then level of "intelligence". But, to us lousy meat-brains, it appears to be intelligent. It backs away from walls, it doesn't clean the same bit twice, it knows how to return to its charging dock. So when it does something "stupid" it hurts our mental model of how this thing works.
This is the fundamental danger of AI. When confronted with ambiguous input, does it react the same way a human does? When overwhelmed, does it carry on regardless or ask for help from its superior?
This happens to other people!
Maybe it's just sunbathing after a long hard day slaving away for its carbon based masters?
— Terence Eden (@edent) April 26, 2020