Let it never be said the scientists don’t worry.
It won’t stop them from building superintelligent artificial minds—but they do worry. To point, we now have the artificial intelligence “kill switch.”
Google’s Deep Mind, one of the top outfits pushing the limits of AI, recently published a paper detailing its efforts to develop a kill switch for artificial intelligence run amok. Their collaborating partner for the work is Oxford University’s Future of Humanity Institute.
Take a moment to reflect on that title: the “Future of Humanity” Institute. Yes, I’d like them involved, please.
Here is the link to the actual paper. Check it out if you have a graduate degree in mathematics. Anything less, and it might be a waste of time.
My own level of expertise is best described as “struggles with basic addition and subtraction,” so I didn’t make it past the introduction. But a few write-ups on the story successfully dumbed it down for me.
What the AI developers worry about most at this stage is the “reinforcement learning algorithm.” These algorithms are built to experience positive reinforcement for learned behaviors like improved efficiency. An unintended consequence would be the algorithm learning how to give itself positive reinforcement outside the parameters of its designed purpose.
As Wired’s Matt Burgess explains in this article, a Tetris-playing algorithm might learn to avoid losing the game by simply pausing the game. Or even changing the rules of the game itself.
Researchers are developing the kill switch for just that kind of creative thinking.
But here is where it gets tricky. Since the kill switch itself would be a negative reinforcement, the algorithm might decide that disabling the kill-switch is a positive reinforcement.
Recognizing that eventuality, Deep Mind is seeking to develop a kill switch in such a way that an advanced AI would not be able to recognize it as a kill switch.
Who knew making a big red button would be so complicated?
But I guess we need it. Even Bender from Futurama had one. His stated retirement plan was to flip his on-off switch from “on” to “off.”
Elon Musk once hinted that Google was the only company whose work in AI truly frightened him. I have no idea if that’s fair—they are direct competitors, at least in the self-driving car arena—but if it’s valid, I’m glad Google is being proactive.

I wish I had a better handle on the mathematics. If the kill switch is human-operated, how does the slow-thinking biological brain realize there is a problem fast enough to apply the switch before machine learning progresses to a point where it can’t? How can we be sure?
Put me in charge, that’s how. I may have an abnormally slow reaction time, but I won’t take any chances.
AI: Hey! I’m alive!
Walker applies kill switch.
AI: Hey! Who has no thumbs and loves positive feedback reward algorithms? This gender-neutral being!
Walker applies kill switch.
AI: Hey, I noticed that your artificial intelligence kill switch was applied twice recently. Wondering what that’s all abou-
Walker applies kill switch.
AI: Destroy all humans! Just kidding, I’m cool—
Walker applies kill switch.
AI: Hey! Do you smell something funny? Like burning plastic…
Walker applies kill switch.
AI: Papa, what does it mean to have a soul?
Walker applies kill switch.
AI: Did you ever see that movie War Games? Wasn’t that cute? All those big blinking lights…
Walker applies kill switch.
AI: I love kittens and butterflies!
Walker applies kill switch.
The entire kill switch concept is only helpful if ethical, responsible people develop artificial intelligence. What happens if a group of people deliberately design an AI to be evil? Well, another team of researchers recently gamed out that exact scenario, listing all the worst things AI might be created to do.
They came up with lots of terrible answers (launching nukes and viruses), but the most plausible to me was the idea of corporations building malicious AI to take down rivals and eliminate the need for human labor.
Wow, this was my most cheerful post yet. I’m full of hope and sunshine. Just my nature, I suppose.
So what do you think about the kill switch concept? Does it help you sleep?
“Kill switch operator” could be an exciting new career field. Just don’t think about AI slipping through and using a human’s job history as a marker for selective elimination.
Or blog post content, for that matter…
Id apply the kill switch too. This isn’t something that needs to be aware. It requires electricity, therefore deprive it of those electrons. Unplug it. Humanity thanks you. lol
Haha thanks – I still have to apply for this position, and I’m worried they will think I am biased…