During the past months, I came to realize that my interests revolve around the study of the complexity of how organizations behave and adapt. This interest is quite broad because organizing is one of the broadest research fields. Governments, firms, teams, brains, cells, and nuclei organize. The chemical compounds that lead to life are the output of an organizing process filled with standard operating procedures, roles, and resources. Organizing is central to most of our science.
Thanks to Karl Friston, Richard Feynman, and a myriad of other scholars who would never call themselves organization theorists, we now know that the free energy principle is the central struggle of any organization. The free energy principle involves the day-to-day fight to accumulate enough resilience to account for the eternal push of entropy and change. A push that leads every system to its demise at one point or another.
The free energy principle is central to Ashby's good regulator theorem, the idea that "every good regulator of a system must be a model of that system". More precisely, every good regulator must contain a model of that system. Only in this way can a system build the requisite variety needed to account for change. How organizations build these internal models is my central interest, and there is never a better tool to study this beauty than ethnographic studies.
Ethnographies enable the careful observer to understand not only dull main effects but also the nuance and subtle externalities, the second paths, and the processes that lead organizations to accommodate change. To explain what I mean, let me present two examples taken from a Research in Progress webinar that I had the pleasure of organizing. In it, Bianca Crivelini Eger and Sabita Khagdi Sørenson presented the ethnographic studies that comprise their job market papers. The studies revolve around the study of common ideas. In Bianca's case, the development of identity. In Sabita's case, the development of workflows. Tons have been written about both, but the complexities they discover are not trivial.
Let me start with Bianca's study. The second self is the first finding from Bianca's study. A second self is a separate identity that a worker can build to explore new ways to as Hoffman puts it, present one's self at work. This second self encompasses a broadly inappropriate, primarily inconsequential, but absolutely non-absurd way of presenting oneself in society.
Let's break this last sentence apart. The second self, in this case, is a broadly inappropriate presentation of oneself at work. In Bianca's case, it would involve employees coming to work dressed as drag queens. As much as our pursuit for queer rights has progressed, the drag art form tends to lie outside of what we consider appropriate. Similarly, if a person cares not for appropriateness but only for getting ahead and accruing the consequences of their actions, it seems ludicrous to imagine that for a cis-gender man, the number of hours spent in acquiring wigs and learning how to walk in high heels would be of any consequence. Theory tells us then that the only other option is that this specific second self can only appear as a form of absurdity, a form of camp in the words of Susan Sontag and RuPaul Charles. Yet, absurd drag is not.
Drag might appear illogical. Indeed, we don’t have a logic to encompass its performativity. Yet, Bianca shows us that it is, in reality, a highly appropriate, consequential, and sensible action for the pursuers of the art form. We might not have a canonical "logic" to place it, but only the most naïve among believe that organization theory is remotely close to an exhaustive science. Thank God for the ethnographers, the discoverers of our grand canyons of ignorance.
The Crivelini-Eger canyon, for example, has its own unique ecosystem. It is a special place where one gains first hand experience vicariously, as crazy as that might sound. Where one does not learn online nor offline but offsite in some other role while employing different hats and presenting a different identity to society. Be it a banker, baker, or stock broker, within the Crivileni Eger canyon, a second self can drag you out of a slump and change who you are.
As someone who gains most of their insights from JuPyter notebooks, environments like this intrigue me. Computational modeling enables you to do a lot of things. I could build on March (1991) environment where individuals learn from the second self from a drag-code that is somehow correlated to the universal knowledge but also broadly anticorrelated to the org-code. In this environment, an agent socialized in the drag community, and the org community might be better off. But I would never do this because I lack the incentive.
What’s more, how could I sell this idea? How could I explain that some environments are very uncorrelated to the organization yet, correlated enough so that learning from both would help an individual learner? It would seem arbitrary. A reviewer would see it as a trick. They’d claim I just set a Goldilock parameter to a value high enough that it helps the learner but low enough so that no trouble comes about. Not so in an ethnography. Doing drag has a core stigma behind it. It is hard work, and it is not something you do at work. I can’t imagine Bianca will get a reviewer talking about Goldilocks.
The Goldilocks question is, nonetheless, important. Namely, when does a second self benefit one’s work life? Would Patrick Bateman’s second self be equivalent to doing drag? Killing people did help his career at the start of American Psycho. But I would side with Bret Easton Ellis and claim that being a serial killer would – eventually – make your work worse off.
I wonder whether a solution to the serial killer / drag queen question is the idea of a pre-existing logic of professionalism. To follow Ibarra and others, we could imagine that all of us have provisional selves. We carry these contextual selves around as hats in a bag and place the correct hat where and when it is appropriate. Drag then might well be a route for designing and trying out new hats. A way of broadening the multitudes we have inside us and helping us learn by facing contradictions, as Walt Wittman would say. Drag is work at the end, and thus there might be an underlying logic of professionalism that unites the drag scene and the office. Bateman was not being remunerated for his killing, and thus his actions might not share the logic of professionalism. The logic behind professionalism seems an understudied topic, or at least one for which I fail to find institutional-level citations.
As a final thought, I would push back at the separation of the selves. At the office and at the drag show, there is just one body, one brain, one gut that learns and processes emotions. Adrenaline, oxytocin, and dopamine run through one set of veins all the same be it in drag or in suits that look like the ones our grandparents used to wear. The difference between wearing a tie or a wig is that the latter requires bravery.
Bravery is a characteristic seldom present in cis-gender men. It is an action that is risky and fills your body with emotions. Yet, the reason why our history books are filled with brave humans is that bravery is valuable. Bravery is scary, but there is no greatness in being spineless. Indeed, the Confucian saying that we all have two lives and the second one starts when we realize there is only one, is a clue to the importance of bravery. A parable that portrays the ossification of our spine and the birth of an authentic, wholehearted and brave self. Bravery is a transformational action. How much of the unique sights in the Crivelini Eger canyon could be attributed to the transformational effects of a cis-gender Italian man being brave for the first time in their life? Could surviving the adrenaline shock be enough to change one’s work life? We might remind ourselves that adrenaline is addictive.
Similar to Bianca, who opens the box of how people develop their work identity, Sabita opens a further box. One that tends to be overhyped and whose books literally write themselves. Namely, the idea that technology will change the way we work. This idea has been running strong for almost a quarter of a millenium. Ever since the Spinning Jenny started the programming revolution, people have been worrying about how many of us lose our jobs by automation. Its longevity is a testament to how the industrial revolution has changed our lives. Yet, often we forget how ancient automation is. Every decade we idolize the new algorithm of the day. Just today, I had to add chatGPT to the list of banned words on my Twitter, along with Elon Musk and Donald Trump. Who’d said I would become a Luddite?
One could argue that everything one could write about machines at work has been written. As some say, when you add a machine to a work process, it can lead to the combination of performing either: better than the human or the AI alone, better than the human but worse than the AI, or better than the AI but worse than the human, or finally worse than both the human or the AI alone. We can write a lot about how these four outcomes can come about, but more than four there are not. Similarly, we can write about what happens to the humans who run the machines.
Funnily enough, what we tend to lack is empirics of how all of this comes about. As in the case of business ethics, a vast majority of the AI at work focuses on theoretical developments or the study of biases that deceived people's experience. But it does not have to be that way. As Sabita shows.
Sabita studies the use of algorithms at the Danish tax agency, an agency that is in charge of collecting around three-quarters of every Danish citizen’s income (half as income, one-quarter as VAT). In terms of revenue, it has to be the biggest firm in Denmark.
As a Latin American person, I grew up listening to hero stories of how audacious entrepreneurs beat down the tax agency and evaded taxes their whole lives. How no entrepreneur ever made a profit. How when audits came, nothing would happen. I can remember the pride on their faces when they retold their victories. Yet, I did not live in an entrepreneurial household. My dad’s income was taxed as the source. There was not much to do. So we never perform the heroic actions of defunding our state.
Time has changed me. Leaving Costa Rica and understanding how a well-functioning tax agency can actually build a more fair society changed my mind. But, I still have the instincts of how to evade taxes or how to go away from paying what I should. Therefore reading about how taxes are really collected hooked me from the first time I heard of Sabita’s project in 2019.
This was the case for many reasons. The first is that Sabita’s tax collection problem is very similar to many of our models of learning, contextual N-arm bandits, to be orecise. Yet, completely and beautifully different. In a bandit, you tend just to try to learn which option, A, B, C, or D gives you the most value and try to update your beliefs about it. In Sabita’s case, there is an abstraction, as one cannot “pull the same arm” every time. Indeed, auditing the same person every year would be akin to torture. So there has to be some element of abstraction to build categories of people. And here is where the magic starts.
How would you play this game? Clearly, you cannot do it with a single value. Commission and omission errors make just partial sense here. Imagine that you have a category of people who are in 2022 dirt poor and honest. There are millions of them; would you check on them in 2023? There might be some thousands that fit the dirt poor and dishonest category. Wouldn’t it be better to audit them? What about the super-rich, a well-intentioned rounding error might pay off much more than auditing the poor? So should we just audit the rich? Who knows! You cannot stop auditing the poor honest people because there is a chance I move to Denmark and corrupt their hearts. You better not stop auditing the dishonest poor nor the rich! And that is the genius part, you need to audit optimally with a limited budget.
Expanding a simple screening task to a contextual bandit would be enough to make the “from arms to trees” call to arm a reality. But Sabita goes the extra mile which makes everything more interesting. As Hutchins and Barley asked before, Sabita asks how does technology adapt organizations? They are clearly occasions for restructuring, but how does it happen? As an ethnographer, Sabita experienced the use of a plethora of algorithms at many stages of the tax process. There is no single AI to rule them all, there are dozens, used in myriads of ways by the different tax agents. No better or worse framing, ensembling is way more complex than what armchair theoreticians might imagine.
Two aspects remain in my head from Sabita’s work. The competence of the Danish tax agents and the materiality of the algorithms. Hutchins (1991) is a crazy read. He puts you in the shoes of a captain who needs to navigate a ship after the automation falls apart. This mammoth of a captain managed to divide labor and aggregate effort to successfully navigate the ship through a turn before stopping and avoiding a collision. It is an absurd paper and well worth a read.
What I find unique, though, is how exceptional the captain is. Take away my Google Maps, and I cannot even walk in my neighborhood, and this person managed to avoid an “Ever Given” collision without a computer! I my mind, I kept this captain as an exception. I had good reason to do this. As Friston and Feynman explained well, everything from quarks to governments follows the least action principle. We all try to do our least when we can afford to. So it surprised me to see how hundreds of tax agents in the Danish agency work at a kind of impressive level of performance.
Going back to my Google Maps analogy. I often have no idea when Google is wrong, I do not know what biases it has. I do not know if it is good or not. My phone has it installed. I use it. It collects my data. I get my burger. We are all happy. Not so in Sabita’s case. The tax agents are able to understand the blindspots of the algorithms, adapt the goals, and develop new ones. The algorithms are just systems that provide reliability, but the tax agents have a lot of discretion and skill. This was uncommon for me.
Thankfully for me, Sabita provides a reason for this form of exceptionalism. As Luciana D’Adderio has pointed out, the algorithms in the tax agency have become things the agents use as they perform their routine work. In this sense, the routine work enables them to adapt their actions and perform a kind of double-loop learning in which they adapt to the peculiarities of the case while at the same time getting a mountain view that enables them to adapt to the limitations of their tools. Routines are central to our views of organizations and their development, and one would argue that the decision processes are routines. I wonder what this paper could get from engaging a routine dynamic lens.
Independent of this, the nuance with which errors are observed in this paper is a beautiful step away from the electrical engineering view of errors from Sabita’s coauthors' prior work. Mesmerizing as decision structures might be, bit-wise errors are bland. Not so in Sabita’s case. Every case and process has different views and interpretations. They are works in progress, outputs that help paint a picture of the state of an audit or the risk presented by a tax declaration. This is an important notion for the screening function school of thought, as error matter, but how they are contextualized matters more for the organization's adaptability. At the end of the day, it’s the higher loops of learning that lead to the longevity of the firm. Errors and algorithms, come and go.
Looping back to the free energy principle. The two papers provide us in depth notions of how very normal things happen in organizations. Things I could model but would not because I could not imagine how they really happen. And even if I did no one would believe my code. It is clear why when I look in my Mendeley for the papers I admire the most, few of them have regressions and many of them tend to be book-sized ethnographies that ASQ published. Papers like the ones Sabita and Bianca presented yesterday at our Research in Progress webinar.
Comments