AI

AI Doesn’t Know Your Intentions

April 13, 2019

We hear about AI in the news every day. “AI is coming for our jobs”, or “AI is going to take over the world and murder as all”, or even “AI company now valued at $100 Billion and getting ready for IPO.” I’m personally incredibly excited about the future of AI and how it can help us essentially solve every problem we could ever have. But that day could fall anywhere from 10–100 years from now.

Even today, AI lacks the most basic skill that we humans possess that render it significantly less effective than a 6 year old: it can’t fully understand humanintention. I’m not going to get into all of the amazing things AI CAN do today (you can find plenty of research and writing on that) but I’d like to focus on just this intention limitation AI has today to show just how far we have to go to fully realize the full power of AI technology.

Here’s a simple example: budgeting software Tiller (if you haven’t heard of them yet, check it out, it’s great) helps you categorize your purchases across your credit cards to understand your spending habits. They have a product that suits the utilization of AI technology perfectly. Plenty of data, ability to give recommendations, and the opportunity to automatically label transactions.

Here’s the issue: even with sophisticated software and a feature to help auto-categorize transactions, they simply can’t understand the intention of the purchaser.

For example, if I set up the categorization engine to put “Trader Joes” or “Kroger” into a “Grocery” expense category, Tiller can do that, and will do that each time. But if I go to Whole Foods and buy cookware, or a coffee, or something non-grocery related, there’s no data captured that shows my intention. What if I want it categorized somewhere besides the “Grocery” expense section? All the software (or AI) would see is the vendor I purchased from, but can’t understand my intention of shopping there. And herein lies a hairy problem and an interesting crux of the “Doomsday” or “AI is going to save us all” arguments.

Courtesy of Unsplash.com

An AI today can guess what I’ll probably buy (based on spending habits) or where I’ll shop or what category it thinks my expenses should fall in, but it can’t understand subtle changes in my intent to go somewhere or my intentto buy flowers at a grocery store that I don’t want in my “Grocery” budget.

This seems like a mundane detail, but in order to get an AI we hear about in the media that can solve all of our problems and bring us to a utopia we couldn’t reach ourselves, it will need to perform at an unfathomably higher level than us humans. This means understanding human intent, whether verbally or physically, that any one of us (well, most of us) can do today.

Courtesy of Unsplash.com

As another example of this limitation and where we need to go in terms of AI research: think about when you shop on Amazon. It can show you the recommended products and give you insights into things you might like. And for the most part, it’s spot on.

But what their algorithms and machine learning software cannot do: understand WHYyou’re on Amazon today.

Are you shopping for a gift for your sister’s birthday? Are you just browsing out of boredom? Or are you looking for a specific piece of furniture? Their software and AI platforms can complete incredible tasks across devices by tracking what you search or where you are or your typical spending habits to take a BEST GUESS at your purchasing intent. But we are still a long way from Amazon (or any other company) understanding that you had a rough day at work so all you want to do is buy three bottles of red wine and a pint of Ben & Jerry’s ice cream from Amazon Fresh.

While this might seem like “mind-reading” to you, or even technologically impossible, the point stands that in order for AI to solve problems and become the all-powerful force in our lives that the media and other AI companies promise us is coming, we need to first solve some of the most basic problems in AI revolving around understanding human cognition and, in particular, human intention.

If you know of any companies or projects that are working on solving these “soft” or “creative” AI skills, I’d love to hear about them!

Leave a Reply

Your email address will not be published. Required fields are marked *