Many of us have written a computer program of some sort. Maybe just a script to complete a task, maybe complex software the clients interact with. For all of us, I would guess that we've been asked to do something, or we've asked ourselves to build a program, and we use that request as a specification for our code.
When we do this, how many of us get this right the first time? I'm sure some do, and the simpler or narrower the task, the more likely we get it right. However, we also sometimes make mistakes and realize that our specification wasn't correct or that we interpreted the specification incorrectly. Even when I write code for myself, sometimes I realize I haven't described when I need clearly enough, even in my own head.
That happens when I control the entire process. When someone else is my client, there are plenty of ways in which we can make mistakes, especially in decoding the specification. As we use more and more software to make decisions for us and automate our world, I wonder how often we've assumed the software works in a way different than it actually does. With machine learning and artificial intelligence programs, I think there is an even greater chance of a misunderstanding.
Recently I saw that in New York City there is a proposal to disclose how software is used in hiring decisions. Humans are barred from discriminating by law, but what about software? What if there are bias that can come out in the software based on how someone enters data?
I think there are certainly potential issues in having software evaluate individuals, though I do think software will do a better job than many people, once we've worked out the bugs and we know how we want it to work, and then we can compare that to how the algorithm produces results.
These are tough issues, and difficult projects. I also think it's important that we analyze the data that is used in testing software, to be sure that it doesn't contain its own biases. Ultimately I think having public data sets that are evaluated by many individuals is important. I could see a reference set of data and the expected results, or some range of results being required to ensure software meets the goals we want as a society. This, along with the disclosure of how algorithms work will make our use of software better over time.