Years ago when I was an analyst I sat in a presentation where a data scientist gave a talk about an algorithm he developed and implemented within the slots franchise. The slots franchise was a portfolio of virtual casino like slots games that can be downloaded and played on your phone. Even now, years later, 10 out of the top 50 highest grossing games in the world are all slots games. Slots games in and of themselves make hundreds of millions of dollars in revenue on a quarterly basis. The gameplay isn’t very complicated, a user spends virtual in game currency to pull the lever of a slots machine. If the correct symbols line up, you get more in-game virtual currency. If they don’t, well, you loose the in-game currency you just spent. As you play and rack up wins you get bumped up to “high roller” tables with progressively costlier entry fees. The animations for progressively costlier tables get more intricate, making the wins feel more grand. While the gameplay itself is straightforward, the progression system built around the gameplay is robust with layered penalties for not playing on a daily basis. Pretty soon the daily currency you get isn’t quite enough to play at these high roller tables, so you either have to save up your currency for a chance to play or you can spend just a few real currency to get more virtual currency. Much like real world casinos, the mobile franchises employ psychologists & economists to ensure that the rewards and win rate are perfectly tuned to keep a player engaged. Unlike real world casinos, its probably worth mentioning that there is no way to convert virtual in-game currency into real world money. If you take the time to accrue millions upon millions of virtual currency you do so for the love of the game. Multiple millions of people log onto these slots game every day but very few people actually make the leap to paying real world money for virtual currency. Less than two percent of people pay on a somewhat regular cadence. The top 5-10% of the two percent (~.1% of the entire playerbase) make up the bulk of the entire games revenue. They spend tens of thousands of dollars per month buying virtual currency in slots games and experiencing those sweet, sweet winning animations.
The presentation focused on the .1%, otherwise known as the VIPs. The data scientist had built a fancy machine learning algorithm to predict when VIPs. players would “churn” or leave the game. Losing a single VIPs player meant loosing thousands of dollars of revenue on a monthly basis for the slots franchise. He went on to explain how a series of tests he ran with his algorithm utilizing the VIP management team. The VIP. management team consisted of between 5 – 10 customer service workers and was traditionally used to check up every once in a while with these VIPs. Essentially give them a call, see how they were doing, congratulate them on a recent win. Since a small group of people had to manage thousands of VIPs, not every VIP would get a call every day. Through testing, the algorithm was able to identify players with the highest risk of churning and put them at the top of the VIP call list. The algorithm worked and subsequent reachout worked. It prevented a meaningful amount of people from churning who otherwise would, and thus led to a few hundred thousand dollars of additional revenue over just a six month period. The presentation ended to applause, with many of the people in the room walking towards the data scientist to personally congratulate him on this success.
I remember sitting in the back of the room quietly in thought. I wish I could say that I saw through the complex model, the revenue results and saw the truth behind what was actually happening to the players. The whole goal of the game is to acquire as many users as possible in the hopes of finding users with the propensity to become gambling addicts. Mobile casino games offer the ability to take the key dopamine release provided by in real life gambling and make it accessible anytime, anywhere. The slots franchise spends a fortune on acquiring new players with the goal of netting VIPs. Although some VIPs are wealthy, the vast majority of them are not. The most common profile of a VIP was a female homeowner, between the ages of 55 and 60, with at least some college education and an annual household income of more than $55,000. In addition to paying large sums of money into the game, becoming a VIP involves spending hours upon hours of your day playing different slots machines. As they become and sustain their VIP status they get increasingly frequent calls from the VIP management team. Many of the VIPs would form deep personnel connections and relationships to the VIP management team. I knew a few people on the management team, and they would often talk about how many often the VIPs would comment that the management team were the only real conversations they would have on a consistent basis.
There were many reasons why people stopped paying into a specific slots franchise. One of the main reasons was they simply didn’t have the money anymore. The truth of what was happening to the players was that this algorithm would often times find them at their weakest moment. Their personal slots budget had run out, and they were no longer able to play the one thing that brought them some kind of dopamine release on a daily basis. At this moment, a person who they had developed a deep personal relationship would reach out to them, check in on them, and quietly mention that they were at risk of losing their VIP status. Loosing your VIP status meant no more in game bonuses and, more devastatingly, no more regularly scheduled calls from the management team. This would prompt a meaningful number of them to open their wallets, refinance their home, or borrow money in order to keep the VIP status going. I truly wish as I was sitting there I could have seen that this data scientist had developed a system in which to exploit the addiction and loneliness of people who were my mom’s age all to extract a few hundred dollars. But I didn’t. Instead as I was sitting there my head was buzzing with ideas of how we could probably just cut out the management team entirely by smartly connecting the VIPs to each other. Social pressure, after all, holds way more weight than the tepid words of an overworked 24 year old VIP manager in San Francisco. With the money the franchise saved by cutting the management team they could probably host a couple happy hours.
In Weapons of Math Destruction Cathy O’Neil does a great job illustrating how black box algorithms have enabled the large scale spread of unequal treatment across various columns in society. She demonstrates how in key industries and social functions algorithms do not offer a fabled bias free decision making system based on all powerful “math”. Instead, the use of complex algorithms serves to codify pre-existing discriminatory beliefs and practices in different, more opaque ways. O’Neil details the use of mathematical models to scale injustice across various realms including criminal justice, predatory loans, college admissions, and low wage employment. The key power of the all powerful algorithm is its mathematical density & opaqueness. It allows those who wield it to use it as a shield in order to mask clearly discriminatory and harmful outcomes.
Listen man, I’m a numbers guy. I put this model into production, rev goes up, I get a bonus. You want to know how it impacts people go talk to the “cares about people” guy.
That to me is the key takeaway from this book. With the continued advancement of data collection and compute power, data scientists are able to scale the rent seeking initiatives demanded by those in power to a level they could previously only dream of. Data scientists are starting to take over what was previously in the realm of ghoul ass management consultants. The ones who would storm into a business after a private equity hostile takeover and start slashing the business left and right until they can easily suck what little assets have been left. The ones who get called in by city governments to reinvigorate the city and offer innovative solutions such as buff up the police force, kick the undesirables out, and gentrify baby. This brings me to my biggest issue with the book as a whole: the solutions proposed to combat WMDs.
Although there are brief vignettes of solutions (or rather attempts at solutions) layered throughout the book the conclusion & afterword sections of the book lay out O’Neil’s key propositions.
- Make data scientists take some kind of moral hippocratic oath
- Make sure our representatives pass legislation which limits the damage that unfettered algorithms can cause
In my view, these solutions are pretty naive and shows a misunderstanding of why these algorithms are pursued in the first place. For the first suggestion, I can picture the conversation between the DS and his boss in the earlier anecdote.
DS: Hey boss I’ve discovered this algorithm basically preys on older, lonelier people and pushes them to spend beyond their means to keep playing. I don’t think we should be using this to convince them to come back into the game.
Boss: So you’re telling me you want to throw away hundreds of thousands of dollars in revenue?
DS: Well actually, I think we should instead be using this algorithm to do the right thing, to identify people who are clearly addicted and have to stop. We can reach out to them and make sure they get gambling addiction rehab.
Boss: You want to throw away hundreds of thousands of dollars in revenue and instead insure that our most valuable cash cows never pick up the game again?
DS: Well ya, it’s the right thing to do.
Boss: Dope I hear you. Hand your model off to our stats intern I’m transferring you to another team.
DS: Are you kidding me? But why?
Boss: Because I’m going to use the quarterly bonus from the increased rev to buy a beach house in Santa Cruz you idiot.
Of course, this conversation would never happen. The DS knows he would get a bonus himself, and he needs to make sure that he can cover his kid’s tuition at Athenian. Maybe I’m biased, but tech workers looking past their material circumstances to keep the greater good in mind is more unlikely than hitting a home run on any of the Big Three. Besides, their entire job is predicated on finding the most scalable way to get to the objective function (be it revenue, # of people off the streets, etc…) in the quickest, easiest way possible. Unless the driver behind objective function is eliminated, those tasked with developing the WMDs will never be able or willing to change their destructive effects. For the gaming example above, it is the existence of easily accessible free to play games that is the root of the problem, not the individual WMDs that are incorporated inside them. Transparency in these models showcasing the destructive effects will have no sway over those in power whose very paychecks depend on ignoring the destructive effects. If you truly want to address this particular issue, we as a society need to outlaw the concept of hyper exploitative free to play games. This holds true for every exploitative industry covered in this book, from the predatory loan department to the for profit college system.
For the second point, getting lost in technocratic debates about a particular algorithm’s “fairness” moves the source of power to change things away from the people being affected by them. In O’Neil’s words:
Challenging WMDs will require a movement of people who refuse to bow down to the algorithmic gods, who band together, collect evidence of their harm, and demand better laws from policy-makers.
Nowhere in that solution is there room for the those being directly effected by the WMDs themselves to assert their point of view and challenge them directly. In fact, the only real substantial progress with regards to criminal justice has come in the past few months as everyday people have risen up and made their voices heard with the protests over George Floyd’s death. The Minnesota police department was disbanded and forced to abandon its use of Compstat to harass citizens not because of some pencil pushing data scientist showing some graphs, but because the citizens made their voices heard and burned down a damn police station. Likewise, unionization movements starting all across Amazon warehouses are the only fundamental organization standing in the way of increased workplace tracking and surveillance. Demonstrations and the organizing of worker power are the only real tools we as people have against the ongoing creep of WMDs into our everyday lives. My key point here is that change doesn’t come around waiting for some Ivy graduate DS who has seen the light to have a heart to heart with a few Congressmen. The power to combat the creep of WMDs comes from organizing in your own workplace and not being quiet in the streets when you see injustice happening. These solutions are hard, and they force us out of our comfort zone, but these are the only solutions that tackle the root cause of the proliferation of WMDs.