By Kim Rees, head of data experience design at Capital One
As a practitioner of data visualization and the design of data systems, I appreciate how the choices we make about data collection, curation, extraction, and transformation impact the outcomes of our analyses and algorithms. The way we choose to frame a question can sometimes yield vastly different results. With these considerations — and their far-reaching implications — in mind, it becomes of critical importance that they are factored into algorithmic development to help set the vision for, and more accurately inform, ultimate intended outcomes.
As we look at the future of AI, machine learning (AI/ML), and autonomous systems — and their increasingly outsized role across businesses, public agencies, educational institutions, and society broadly — we should be aware that models are trained on our past behaviors, not on our values or ideals. For instance, say an airline creates an algorithm to upgrade passengers trained on past upgrades given by gate agents. We could view those agents’ choices as an expression of the company values. While the airline doesn’t have an explicit value of a certain type of customer, perhaps gate agents’ unconscious bias led to more affluent passengers getting this perk.
The “problem” with AI and machine learning.
The problem with AI/ML isn’t that it "gets it wrong" or that it’s built with malice, it’s that the technology has gotten really good. Most of the negative headlines we read about algorithmic outcomes or bias are referring to the fact that the machine actually got really darn good at its job. The machine being a machine, this means that it excelled at executing a very narrowly defined set of tasks based on prior actions and decisions (for example, a more robust set of training data on a certain set of passengers may yield preferential treatment for that group on future algorithmic outputs). The hard part about what I think of as “responsible AI,” is that it forces us to take a hard look at ourselves and recognize that our data is not neutral, because we aren’t neutral.
To me, this begs the question of whether we use the word "bias" too liberally. There are many ways algorithms can single out a particular group of people without having biased data. For instance, the mere quantity of data available for a specific person in any given use case may cause an algorithm to rank them higher than someone else. People with a more robust digital presence may enjoy favorable treatment simply because the machine knows more about them.
In order to ask the machine to make the right decision, we can’t just give it our data and have it learn from our actions. Just as a mom, I can’t let my kids learn from me in a vacuum. Because I’m not perfect, and I let my exasperation come out while driving in rush hour traffic. I need to course correct and tell them I made a poor choice and how it could have been handled better.
Humans are naturally flawed.
Presuming we won’t suddenly start making flawless and fair human decisions, so the models can learn our values properly, we need to start by admitting that our past human decisions are flawed. From there, we can perhaps go back to clarify what our values look like when we make certain decisions and start embedding the appropriate levers in the models and algorithms to ensure that those values are reflected throughout the entire system, outputs included.
To help us uncover our flawed past and understand the possible long-term ramifications if proliferated through AI, we need a diverse approach. A diversity of perspectives, whether the perspective stems from socioeconomic, racial, gender, physical ability, professional discipline, or other, is essential to exploring the potential futures created by this technology. And to avoid the possible negative outcomes.
A humanity-centered approach.
It will take a decidedly human-centered approach — or even humanity-centered approach — to ensure confidence in our machine-generated decisions. I fear many of our technological advances will be wasted if we don’t look at human well-being first. If we take a humanity-centered approach, we can start with a crisp view of an ideal outcome and align our algorithms and models to serve that end state. If we simply start with the data at hand, we’ll merely promulgate and intensify the inequalities we already have.
Taking a “slow food” approach with AI, and being mindful of every ingredient, source, and construction, will serve us far better than the “move fast and break things” approach. As machines make more and more of the decisions in the world, we simply can’t accept ill-designed, myopic systems bred in a vacuum of values.
This post is sponsored by Capital One. | Content written and provided by Capital One.
from SAI https://read.bi/2QoHkwd