As a practicing Software Product Manager who is currently working on the 3rd integration of a Machine Learning (ML) enabled product my understanding and interaction with models is much more quotidian, and at times, downright boring. But it is precisely this form of ML that needs more attention because ML is the primary building block to Artificial Intelligence (AI). In other words, in order to get AI right, we need to first focus on how to get ML right. To do so, we need to take a step back and reflect on the question: how can machine learning work for human flourishing?
First, we’ll take some cues from liberation theology to properly orient ourselves. Second, we need to understand how ML models are already impacting our lives. Last, I will provide a pragmatic list of questions for those of us in the technology field that can help move us towards better ML models, which will hopefully lead to better AI in the future.
Gloria Dei, Vivens Homo
Let’s consider Elizabeth Johnson’s recap of Latin American liberation theology. To the stock standard elements of Latin American liberation theology–preferential option for the poor, the Exodus narrative, and the sermon on the Mt –she raises a consideration from St. Irenaeus’s phrase Gloria Dei, vivens homo. Translated as “the glory of God is the human being fully alive,” this means that human flourishing is God’s glory manifesting in the common good. One can think of the common good not simply as an economic factor. Instead, it is an intentional move towards the good of others by seeking to dismantle the structural issues that prevent flourishing.
Now, let’s dig into this a bit deeper –what prevents human flourishing? Johnson points to two things: 1) inflicting violence or 2) neglecting their good. Both of these translate “into an insult to the Holy One” (82). Not only do we need to not inflict violence on others (which we can all agree is important), but we also need to be attentive to their good. Now, let’s turn to the current state of ML.
Big Tech and Machine Learning
We’ll look at two recent works to understand the current impact of ML models and hold them to the test. Do they inflict violence? Do they neglect the good? The 2020 investigative documentary entitled (with a side of narrative drama) The Social Dilemma (Netflix) and Cathy O’Neil’s Weapons of Math Destruction are both popular and accessible introductions to how actual ML models touch our daily lives.
The Social Dilemma takes us into the fast-paced world of the largest tech companies (Google, Facebook, Instagram, etc.) that touch our daily lives. The primary use cases for machine learning in these companies is to drive engagement, by scientifically focusing on methods of persuasion. More clicks, more likes, more interactions, more is better. Except, of course, when it isn’t.
The film sheds light on how a desire to increase activity and to monetize their products has led to social media addiction, manipulation, and even provides data on the increased rates of sucide amongst pre-teen girls. Going even further, the movie points out, for these big tech companies, the applications themselves are not the product, but instead, it’s humans. That is, the gradual but imperceptible change in behavior itself is the product.
These gradual changes are fueled and intensified by hundreds of daily small randomized tests that A/B change minor variables to influence behavior. For example, do more people click on this button when it’s purple or green? With copious amounts of data flowing into the system, the models become increasingly more accurate so the model knows (more than humans) who is going to click on a particular ad or react to a post.
This is how they generate revenue. They target ads at people who are extremely likely to click on them. These small manipulations and nudges to elicit behavior have become such a part of our daily lives we no longer are aware of their pervasiveness. Hence, humans become commodities that need to be continuously persuaded. Liberation theology would look to this documentary as a way to show concrete ways in which ML is currently inflicting violence and neglecting the good.
Machine Learning Outside the Valley
Perhaps ‘normal’ companies fare better? Non-tech companies are getting in on the ML game as well. Unlike tech companies that focus on influencing user behavior for ad revenue, these companies focus on ML as a means to reduce the workload of individual workers or reduce headcount and make more profitable decisions. Here are a few types of questions they would ask: “Need to order stock and determine which store it goes to? Use Machine Learning. Need to find a way to match candidates to jobs for your staffing agency? Use ML. Need to find a way to flag customers that are going to close their accounts? ML.” And the list goes on.
Cathy O’Neil’s work helps us to get insight into this technocratic world by sharing examples from credit card companies, predictions of recidivism, for-profit colleges, and even challenges the US News & World Report College Rankings. O’Neil coins the term “WMD”, Weapons of Math Destruction for models that inflict violence and neglect the good. The three criteria of WMD’s are models that lack transparency, grow exponentially, and cause a pernicious feedback loop, it’s the third that needs the most unpacking.
The pernicious feedback loop is fed by biases of selectivity in the original data set–the example that she gives in chapter 5 is PredPol, a big data startup in order to predict crime used by police departments. This model learns from historical data in order to predict where crime is likely to happen, using geography as its key input. The difficulty here is that when police departments choose to include nuisance data in the model (panhandling, jaywalking, etc), the model will be more likely to predict new crime will happen in that location, which in turn will prompt the police department to send more patrols to that area. More patrols mean a greater likelihood of seeing and ticketing minor crimes, which in turn, feeds more data into the model. In other words, the models become a self-fulfilling prophecy.
A Starting Point for Improvement
As we can see based on these two works, we are far from the topic of human flourishing. Both point to many instances where ML Models are currently not only neglecting the good of others, they are also inflicting violence. Before we can reach the ideal of Gloria Dei, vivens homo we need to make a Liberationist move within our technology to dismantle the structural issues that prevent flourishing. This starts at the design phase of these ML models. At that point, we can ask key questions to address egregious issues from the start. This would be a first for making ML models (and later AI) work for human flourishing and God’s glory.
Here are a few questions that will start us on that journey:
- Is this data indicative of anything else (can it be used to prove another line of thought)?
- If everything went perfectly (everyone took this recommendation, took this action), then what? Is this a desirable state? Are there any downsides to this?
- How much proxy data am I using? In general proxy data or data that ‘stands-in’ for other data.
- Is the data balanced (age, gender, socio-economic)? What does this data tell us about our customers?
- What does this data say about our assumptions? This is a slightly different cut from above, this is more aimed at the presuppositions of who is selecting the data set.
- Last but not least: zip codes. As zip codes are often a proxy for race, use zip codes with caution. Perhaps using state level data or three digit zip code levels average out the results and monitor results by testing for bias.
Maggie Bender is a Senior Product Manager at Bain & Company within their software solutions division. She has a M.A. in Theology from Marquette University with a specialization in biblical studies where her thesis explored the implications of historical narratives on group cohesion. She lives in Milwaukee, Wisconsin, enjoys gardening, dog walking, and horseback riding.
Sources:
Johnson, Elizabeth A. Quest for the Living God: Mapping Frontiers in the Theology of God (New York: Continuum, 2008), 82-83.
O’Neil, Cathy. Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy (New York: Broadway Books, 2017), 85-87.
Orlowski, Jeff. The Social Dilemma (Netflix, 2020) 1hr 57, https://www.netflix.com/title/81254224.
