I want to bring your awareness to what are the kinds of issues that can happen when you start using machine learning and AI modules. We've talked a lot about customer journey, we've talked a lot about all the different things that machine learning algorithm or AI software can do for you. But at the same time, I think it's important to understand what are the risks that are appearing as well. One example of the risk is understanding where the data is from. I'll give you two examples, but I think these examples, of course, you will see, is in some sense tip of the iceberg. These examples come from Amazon again, of course, Amazon is a big company. There are many other companies which also involved in some of these examples, but I just want to give you an example for Amazon, but you will see how general these examples are. One example is this idea of Amazon scrapping a secret AI recruiting tool. What was the problem here? The problem here was that Amazon, like many of the companies, perhaps have a lot of people who want to, in some sense, become employees at Amazon. Well, they had a lot of people submitting the resumes and in this case it was about, I think, software developers, if I remember correctly. Many people want to become software developers at Amazon. What did they want to do? They wanted to speed up that process. How did they do it? They basically said, wouldn't it be great if you could have an AI tool that goes through some of these resumes and tries to figure out which resumes perhaps have the highest chance of becoming successful, which people? What did they benchmark against? The benchmark against their current employees and their own resumes that they had, and then they started looking at how best to predict. But what was the issue? The issue was that the historical data that we had, which is the data of the current employees, were largely male software developers. What did they end up doing? Unfortunately, the AI tool ended biasing against women. Again, notice it's not the fault of the tool per se, it's the fault of the data. Thinking very carefully about where the data is coming from, how generalizable is that data? There are many, many other issues of that kind. Let me give you one more example. This is the example of Amazon and the racist algorithms that you see here. Again, obviously very controversial. Let's discuss this example and then we'll also talk about the controversy. Now in this case, Amazon was thinking about expanding their Amazon Prime service that were offering in the US. This example and this issue was especially in the Boston area. Now, what happened there was the following : they were looking at where to offer Amazon prime, they were looking at different ZIP codes, and then they basically did the following calculation; well, let's offer it in certain ZIP codes where there is a chance to become more profitable which makes sense. Once they started looking at that, now those of us who are from Boston they might recognize this, there's an area called Roxbury in Boston, which is a slightly lower income area. This ended up becoming what they call the Amazon problem, which is like a doughnut, which is in all areas around Roxbury, they were offering Amazon Prime, but for Rock Spring. But you can imagine this was a very controversial. On the other hand, you can also imagine that people in Roxbury who did not have many stores around them, those will actually be the people who benefit the most from Amazon Prime. Now, why did this happen? Because again, they were looking at data, but they didn't follow through to see what those algorithms are doing. First issue I want you to be aware of as you start using your customer data to predict the journey, shocking journey, whatever the case might be, is first think about how generalizable that data is. Is it touching all different customers that you have? Starting to think very carefully about how representative the data is. The next issue is around datasets and privacy. False customer privacy is a huge issue. In Europe, you have the GDPR. In California, you have the California Privacy Act. There are many, many other such Privacy Act that will be coming forth. I would advise you as you start thinking about using your customer data to then perhaps target customers, make the journey customizable, personalizing the journey, whatever the case might be, start thinking very carefully about ensuring customer privacy and what can you do to make sure that you don't infringe their privacy. Let me give you an example. This example comes from this company called Strava. Now, Strava is a company that in some sense was building a global heatmap, so to speak that's what they talked about. It's a big social network for people who are very active. If you're on Strava, you can put it up, you can say, well, I went ahead and jog five miles and so on, so forth. They had in some sense noted the fact that they are building a global heatmap of all the people who are very active. What happened? Well, their own data was anonymized and they made very sure that their dataset perhaps is not in some sense identifiable. But what did some people do? They ended up imposing a google map on top of that. What ended up happening? Well, in some sense they were able to find people, I believe, in Afghanistan and so on, who were the US Army that was stationed there. Clearly, that's what this headline talks about; the Strava heatmap and the end of secrets. What are the big idea here? The big idea here is your own dataset might be anonymized but increasingly, there are many, many people out there who are putting datasets together. That's something to be very careful about. Which is it's not just your datasets that should ensure privacy but think very carefully about how datasets coming together might end up taking care of privacy. A third issue is how are you using machine learning? This relates to the idea that we had brought up earlier about data itself but it's more than just the representativeness of data. It's the idea of how often your data is getting updated? Which is, are you using data from two years ago, a year ago? Look at the customer journey. Look at how frequently customers are buying your products. What's the journey like? Which parts of the customer journey are sticky? Which parts are changing? Because of course, as new technology comes in, as new competitors come in, all of that journey changes. Thinking carefully about how you're using your data, you're using your models, and how often is that data getting updated in terms of how the journey is changing is critical. Of course there's no data for free. What that means is, I think when you start thinking about using new technology, be it machine learning, be it AI, whatever the case might be, you need to have a proactive approach. Perhaps doing a little bit of testing and learning, collect some data, queue new models, see how well that new data that you've collected, how well your models are predicting that. The more you do that, the more you can be agile in terms of understanding how you machine learning and AI software is "getting updated" in terms of new data coming in and how those older models perhaps need to be tweaked as well.
