What you need to know about AI

This piece was originally published at Forbes.com.

Lately, when clients come to me as a consultant, Artificial Intelligence (AI) usually comes up in our conversation. And when I’m asked, “How do I use it?” that tends to actually mean “What is it?” Let’s reach a basic understanding of AI so when you’re ready to explore what AI can offer, your discussion can be as productive as it can be.

WHAT IS ARTIFICIAL INTELLIGENCE?

As the name implies, AI is the intelligent behavior of machines. Most companies could utilize AI to interpret complex data. Here’s how it works (in a very simple way): The AI model asks a set of data a question and returns an answer. To accomplish this task, an AI model needs to understand the data it’s interpreting. So, for AI to deliver accurate, useful information, an AI model needs to be trained on the data it’s given. We’ll get into that soon, but first let’s talk about that data.

LEARNING FROM DATA

For AI to work, it needs to learn from specific kinds of data. With organized, not random, data, an AI platform can learn what it needs to. Let’s say you want to train an AI model to identify dogs in images. Organized data would consist of animal images, including dogs, to help the AI discern what is and what is not a dog. Random data (in this case, images of tables, lawnmowers, anything not reasonably close to our concept of a dog) doesn’t help the AI distinguish dogs from other animals. Know what’s in your data and you should be able to avoid any randomness.

At this point, I should clarify that AI isn’t actually telling you what something is — AI tells you what something probably is not. This is determined through a prediction percentage — you’re not teaching an AI model to know an animal is a dog but rather training it to tell you it’s a certain percentage confident it recognized a dog. If you’ve been feeding your AI model with images of only dogs and cats, then introduce an image of a squirrel, your AI will be less certain of what it sees. But, once you teach the AI model what a squirrel looks like, the AI can discern with more certainty.

HOW TO TRAIN YOUR AI

Machine learning, the ability for computers to learn without being programmed, can take place a couple different ways. One way is supervised learning, where an AI model infers a function from labeled training data. Those images of animals I mentioned earlier would be labeled “dog,” “cat,” “hippopotamus,” etc. to help the AI learn. The other machine learning method is unsupervised learning. This is where machine learning draws inferences from data sets consisting of input data without labeled responses — you provide a bunch of pictures of animals with no descriptions and let the AI figure out what is and what is not a dog. Remember to, with either approach, provide organized data that your AI will learn from.

Google provides a lot of data sets and pre-trained AI models for purchase. But, they’re all about object recognition and will only do a good job of recognizing things that existed when the model was trained. So, a Google AI model may have learned what a bike and pogo stick are at some point, but a newer invention, like a Segway, could confuse it.

BEYOND CATS AND DOGS

I’ve led or been on some teams that have worked on training data for tasks other than image recognition, like classifying audio samples or, during an innovation session, recognizing what restaurant is delivering food by identifying food service workers and recognizing what was being delivered. The latter is an excellent example of what can happen when you train very specific models. In this application, the AI model learned from uniforms, not beverage cups, cars or even people. The data set provided in this case included shapes of what delivery drivers could be carrying and the clothes they wore.

The accuracy of your AI model is really about the bucket of data you give it. You need a lot of pictures, and they need to be of a similar quality. The more similar they are, the less training you likely need.

WHAT COULD GO WRONG?

Okay, our hypothetical AI is up and running. What is at stake if it’s wrong? Within your own organization, determine the impact of a false positive or false negative. AI that determines what is and isn’t email spam could afford to let some spam through but could create problems if it marks something important as spam. Imagine the consequences of AI providing results of x-rays with a false negative cancer diagnosis.

Now, if you feel more comfortable with how AI works, you can begin the challenging task of figuring out where it fits in your organization.