People often fear the onset of a real life Skynet, and it’s easy to see why. When you have artificial intelligences (AIs) that can design their own AIs, understand the benefits of betrayal, and beat human players at complex strategy games, then a future of humanity being enslaved by machines is easier to imagine.
However, it’s worth pointing out that AI is still pretty dumb, and it often makes terrible and sometimes hilarious mistakes. Just try having a proper conversation with the AI in your phone, and you’ll quickly realize that a lot more work needs to be done in order to make it seem more akin to a human than a bot.
In response to what some have called “artificial stupidity”, a team at Google have conjured up project PAIR, which stands for “People and AI Research”. Led by two data visualization experts, the aim is to produce a blueprint for future AIs that will make them seem less error-prone and awkward, and improve human interaction with machine intelligence.
In effect, PAIR is hoping to come up with a new type of interface – or at least the design for one – that will smoothen out the interactions between AIs and their human masters.
The team note three focuses of their research: How to make AI easier for engineers to build, how AI can aid and augment professionals (such as in healthcare), and possible new applications for machine learning.
The equation to success here doesn’t just involve improving the AI. The expectations that people have towards their AIs must also be studied and pulled apart too.
That way, an AI could be designed in such a way, or presented in such a way, that encourages us to make requests that will get a positive response, rather than a muddled one. Importantly, this process will involve lifting the hood on AI development to the general public.
At this point, those that do not understand how algorithms band together to form a virtual intellect see it all as science-driven witchcraft, manufactured by those in tall towers far from the public eye.
That’s why it’s so unnerving to watch an AI correctly diagnose a disease, just as it’s so amusing to see one fail to come up with a non-threatening inspirational poster. AIs for many people are one of two things: learning children prone to making mistakes, or advanced aliens with unknowable minds.
PAIR wants to improve the partnership between AI and humans.
One rather simple-sounding way in which to partly achieve this involves the AI explaining its decision-making process as it goes, or at least pointing out the factors that influenced its conclusions or actions. That way, it becomes a more transparent, less mysterious, and ultimately less threatening process.
Without a doubt, the future will be built on the foundations of AIs. With this in mind, Google’s eagerness to make AIs more human compatible isn’t just a welcome initiative – it’s an essential one.
[H/T: Wired]
