We need to talk about dueling artificial intelligence.

If you want to make your AI smarter, why not just have it battle with another AI? It sounds like a pretty straightforward idea, right?

For example, have one AI tasked with creating a hyper-realistic image while another is challenged to determine whether that image is real or artificial. This is just one domain in which artificial intelligences can duel, but any domain modeled by computers such as voice, video, etc., can host what is known as a “generative adversarial network” or GAN of competing AIs.

And while this sounds like fun, consider the vulnerability created by dueling AIs to online verification programs or systems intended to distinguish human from machine input, such as CAPTCHA. These typically serve as a way of thwarting spam and automated extraction of data from websites. Dueling AI can easily conquer these kinds of gateway software platforms, and may one day have to be protected by something akin to blockchain networks.

What happens when AI is too smart?

Of course, there are some potential upsides here. The rapid, unattended self-enhancement of competing AIs may one day lead to dramatic advancements in transportation, technology, and medicine. A number of e-commerce industry leaders have taken steps to research how competing artificial intelligence can improve their business models since they have the potential to be applied to all problematic domains.

That said, companies seeking quick fixes or solutions should not neglect the essential role that humans play in protecting computers, data, and networks from attack or unauthorized access, and continue developing breakthroughs on our own.

The idea of imbuing dueling artificial intelligences with something akin to a sense of imagination without reliance on humans could end up as dangerous tools for online fraud and digital fakery. We could find ourselves in a digital war of AI foot soldiers directed by far-flung human generals.