Captchas and deep fakes, and neural networks making themselves obsolete.

How relentlessly will our robot overlords pursue efficiency?

Reading about deep learning and neural networks reminded me of a recent podcast I’d listened to on Captchas. What’s a captcha? If you’ve been online before, you’ve probably seen one. They’re those squiggly, hard to read bits of text that you have to translate. Or, more recently, captchas are those tiled pop ups where you select all the images with “x” in them, where x is a bus, stoplight, van, pigeon eating a bagel… etc.

gif of a recaptcha box

Captchas are interesting from a few angles, but most pertinent for this discussion is that besides acting as security, the “solutions” you provide help train neural networks. More specifically, you help train Google’s neural networks. Every time you pick out a bus or car, Google gets a little better at doing the same thing. Every new sign on you create where you fill out a captcha makes AI stronger. Stop making Gmail accounts just for coupons, before the machines win.

Ironically, as Google’s neural networks get smarter, so do those of hackers and scammers. So, the captchas have to get more difficult. Really soon, the networks are going to be so smart that they just don’t work. So it’s kind of a weird microcosm of jobs being automated away, where a machine is creating efficiencies that make it obsolete. Will our robot overlords be so relentless in their pursuit of efficiency that overthrow themselves? That’s an interesting question, but it’s not the main one I’d like to focus on.

Deep fakes or remixes?

On a different level, this increased image recognition, processing, and recreation ability we see with captchas is leading to some pretty fundamental shifts in how we interpret media. The idea of “deep fakes” has recently caused a stir with a (not even particularly well) faked video of Nancy Pelosi.

With computers on the cusp of, or perhaps already able to combine, subvert and recreate footage, what kind of effect will that have on our media consumption? Could it signal a return to reasoning more grounded in philosophy rather than verified facts? And when a neural network or algorithm takes one form of media and turns it into another, does that count as remediation?

Imagine the potential of videos created entirely by algorithms. Would they feel as authentic as human creativity? The moment of AI singularity commonly refers to a computer as smart, or smarter than humans. What about computers that are more creative than most humans. Or at a more narrow level, just computers that make better videos?