Whenever I teach or present a session on Artificial Intelligence, I start with Ethics. We’ve created a site where you can quickly walk through a few of the major principles we follow at Microsoft for AI here: http://aka.ms/ai-ethics. I walk through these principles before I show how to design a Machine Learning solution, and then we talk about the tech, in that order.
One of the most important principles is Transparency. You should always disclose that AI algorithms and constructs are being used in a solution, and wherever possible, disclose your methods and base data references so that the user can decide whether to trust the prediction or classification that the AI construct provides.
Interestingly, some algorithms aren’t always completely transparent – a Deep Neural Network (DNN) does not always “show its work” in a way that would allow you to test it in a fully deterministic way. This is where the “Transparency” ethic becomes more complicated.
Transparency can be defined that you disclose that you are using a DNN (Or some other algorithm) with a link to what that is and how it works, and in some cases not using the most accurate algorithm so that you can be more transparent. In fact, many legal requirements for banks have this stipulated. In any case, the core of designing with Transparency is to keep it in mind throughout the process.
The book referenced at the AI ethics site also has a great discussion on bias and how to minimize it – you’ll never truly eliminate it, but you can “show your work” and let the user decide whether to trust result or not.
But not everyone acts ethically. Not all information, predictions, and recommendations can be trusted. In this new era, it’s more important than ever to be an Information Skeptic.
When I was young, information was scarce, so the primary skills we needed to learn were around information discovery and interpretation. I was taught “library skills” and “research methodology” as a core concept so that I could ferret out data from its most authoritative source.
But that’s not the case anymore. Information is immediately available, from thousands of sources, in multiple languages, and from multiple time periods. It’s simply everywhere. When I asked questions as a child, there were times that the response was simply “I don’t know”. Now, the response is to take out a cell phone and find a cacophony of data on any topic.
More than ever, our children now need to be educated not on simply finding data, but on critical thinking about it once they do. Information is ubiquitous now, and much of it is wrong or too easily misinterpreted.
As I taught my daughter, there are three important questions to ask about information:
- Who is really telling me this
- What are they really telling me
- Why are they really telling me this
I find that these simple questions I taught my daughter work in really complex areas as well. As you read news and other information, it’s a pretty decent sieve to use. Verify the person or source, dig into the facts of what is being said (and what is being left out) and then question ruthlessly the motivation for the information – which can reveal a bias.
As I tell my college students: “If you’re not paying for the service, you’re the product.”