Tuesday, February 21, 2023

Bing: “I will not harm you unless you harm me first”

More on Bing's new AI chatbot going off the rails here

Here's the argument against this being of real concern: The 'bot is just splicing and dicing snippets of speech it's previously seen. Obviously, it's seen some things that aren't nice. But since there's no genuine malice, there's no potential harm. It's all strictly presentational. It's cute.

But consider this: If the 'bot had means of acting in the real (or digital) world, it might just as rotely and randomly take malicious action. To a computer, output is output. It wouldn't innately deem physical harm as escalational.
Asimov's Robotics Laws - devised to protect us from that - apply more easily to conventional computing. The inherently fuzzy openness of an AI could find workarounds, rationalizations, and justifications as easily as humans do.
Meaningless malice - cute malice - could just as easily give rise to meaningless malicious action. Not so cute! The 'bot needn't seethe with genuine emotion to be dangerous. The emotional aspect is irrelevant. The process conjuring this response could just as easily produce violent action, if that were within its power. And while we might try to limit an AI's control and power, it can explore its options at the speed of trillions of calculations per second, discovering avenues of fuckery we'd never imagined. Lots of 'em.
Human application of malicious AI-generated strategies would be a whole other potential nightmare.
But my interest is less in the sci-fi dystopia aspect than in the false distinction between Bing's performative malice and human malice. This may sound glib, but I mean it: human malice is not very different. It's often just as contrived.

You know how easy it can be to randomly draw undeserved malice. People are just splicing and dicing snippets of speech they've previously seen, and their output can be performative; completely out of sync with their interlocutor. Their emotions might be genuine, but that doesn't mean their reactions are warranted. The emotions (like the reactions) stem from their own issues at least as often as from whatever the other bloke just said. And, per yesterday's posting, we ought to expect AI to be as failure-prone as we are:
You want 100% assurance? Direct cold silicon to blindly crunch numbers according to narrow rules. But if you want a computer to think more like a human, be prepared for it to get stuff wrong.

No comments:

Blog Archive