External Harms: Imposing Adverse Conditions

Let us start with external harms. There is plenty of excellent work on the ways in which Big Data has been used to impose adverse conditions, so we will only provide an overview in this section (we are particularly indebted to the work done by Noble 2018; Benjamin 2019; and D'Ignazio and Klein 2020).

The first harm has to do with explainability. The issue here is that the more complicated the algorithm becomes as it is fed with variables and data, the less we will be able to understand or explain how it arrives at the answer—even if it is the right answer. To understand the potential harms involved, consider the example of job application screening. In 2016, it was reported that over 70% of applications are never seen by human eyes in the first round of the job selection process, but are instead fed into algorithms which predict the suitability of candidates (Mann and O'Neil 2016)—this percentage has likely increased in the years since this study. This means that a recruiter may not be able to provide an applicant with an explanation for why they were not progressed past the first stage in their application, since the algorithm that rejected their application may be too complicated to provide a human‐understandable explanation for how it arrived at its decision. The lack of explainability is one way in which harms are done through the use of Big Data, as reasons may not be available for some of the most influential decisions in people's lives.

Explainability closely relates to the issue of algorithmic bias: if the training data is biased, the algorithm itself will be biased. In other words, the principle of “garbage in, garbage out” applies. We want to highlight two areas of concern when it comes to algorithmic bias: biased decision‐making and biased information‐curation.

The first area of concern is decision‐making: because of the data that they are fed, Big Data algorithms can become biased toward certain variables as predictors of success. For example, they may be biased toward the variables of “white” and “male” for job candidates, because “successful” candidates who were hired in the past in many companies were, for historical and structural reasons, predominantly white and male. This bias is hard to screen out, as other markers such as educational background can represent race and gender by proxy. The algorithm might be trained to ignore “white” or “male,” but still weigh Ivy League educations highly in its calculations, which will function as a proxy for those markers (as those who attend Ivy Leagues have been disproportionately white and male—recall that Princeton and Yale only admitted women starting in 1969). Harms as a result of biased decision‐making can be widespread, as algorithms can replicate and perpetuate biases in major decisions about job applications, housing availability, credit ratings, insurance pricing, child custody, bail, and criminal sentencing.

The second area of concern when it comes to algorithmic bias is information curation. Negative, salacious, or false content is correlated with greater user engagement (Kaiser and Rauchfleisch 2018; Ribeiro, Ottoni and West 2020; Spinelli and Crovella 2020). The previous choices of end‐users to consume such content can therefore indirectly bias data sets, thereby training algorithms aimed at promoting user engagement to provide more false or outrageous content. Moreover, there are plenty of malicious actors online (e.g., bots), so there are many direct attempts to bias data sets by engaging with and creating false content. Problems then arise when Big Data algorithms are relied upon to inform narratives, as with the promotion of content at the top of search results and newsfeeds. There are risks of perpetuating misrepresentations and contributing to a form of epistemic injustice known as “hermeneutical injustice” (Fricker 2007), which is injustice that happens when significant aspects of one's experience are obscured from understanding due to prejudicial or oppressive norms, structures, patterns of behavior, or epistemic resources. One striking example in the literature was the discrepancy between Google Image search results of mugshots for the query: “three Black teenagers,” and wholesome portraits for the query: “three white teenagers” (Noble 2018). As this example suggests, algorithms reflect the data they have been trained on, which makes them largely a reflection of the society in which they are situated. This runs the risk of further engraining the biases of the society of which these algorithms are a part.

Another closely related issue is over‐surveillance, where harm is done through excessive and inappropriate data collection. To give “informed consent” in accepting terms and conditions for data collection, people need to understand what they are agreeing to. However, as expectations of privacy are often misplaced, violations of privacy and consent are common. This was brought out in a highly publicized case where the store Target used individual purchase history data in combination with a pregnancy prediction model, to predict a teenager's pregnancy—consequently sending her pregnancy‐related targeted advertising before she had even told her own father (Noble 2018). Data collection is also often disproportionate for certain groups in society, especially the most vulnerable—as this example of targeting a pregnant teenaged girl suggests. This means that over‐surveillance feeds back into the problem of bias. For example, certain neighborhoods, such as those with more people of color, end up being overly scrutinized when algorithms inform policing decisions (Benjamin 2019). This links to the issue of heightened security risks—data breaches will impact more on those vulnerable groups who have more data at stake.

A further harm done through the use of Big Data is in the imposition of adverse working conditions. Companies have leveraged Big Data algorithms toward a pursuit of efficiency at the cost of wellbeing, as workers are subjected to surveillance and judged according to the resulting data. For example, in the course of perfecting near‐instantaneous delivery, Amazon fulfillment center and delivery workers pay the cost by being at increased risk of miscarriages due to their working conditions (Gurley 2021) or being “forced to urinate in plastic bottles because they cannot go to toilet on shift” (Drury 2019).

There are also environmental costs associated with the physical presence of data and the training of Big Data algorithms, which use a tremendous amount of electricity and contribute to the acceleration of the rate of climate change. Such climate change impact is disproportionately borne by those in the Developing World (Dhar 2020).

Lastly, all these harms interact with the “technological halo effect”: the tendency to trust algorithms because we think they are impartial, and a more reliable guide than humans (Benjamin 2019). This perceived trustworthiness can further reinforce the biased conditions under which algorithms are trained and deployed, contributing to hermeneutical injustices as narratives shaped by algorithms are believed over stories told by people about their own experience. The technological halo effect also allows people to escape responsibility. It is easy to blame the algorithm or “cold, hard math” when harm is done in the above ways.

Internal Harm: Characterological Damage

We have discussed harms which involve the imposition of adverse circumstances, inhibiting opportunities to access and own resources, explanations, and narratives. We turn now to a second kind of harm: the internal, characterological harms sustained by end‐users and technologists through the use of Big Data.

Advantage over Justice

According to the Judeo‐Christian tradition, the human tendency to try to blame‐shift and escape responsibility runs up through the beginning of human history, starting in Genesis 3, where the man blames the woman for giving him the fruit from the tree, and the woman blames the serpent. The scholastic theologian Duns Scotus, in his theological anthropology, builds on Anselm's work on the will and posits that we have two basic affections: the affection for advantage (affectio commodi) and the affection for justice (affectio justitiae) (Hare 2001, 55–59). The affection for advantage is a tendency toward one's own happiness, whereas the affection for justice is a love of the intrinsic goodness of things for their own sake. Where the affection for advantage does not compete with the affection for justice, it is good—but when the two are in competition, it is sinful to rank the affection for advantage over the affection for justice. Human sinfulness consists in our having a kind of “default setting” or innate tendency to consistently rank the affection for advantage over the affection for justice.

Our temptation with the technological halo effect is to offload decisions onto the algorithm, along with responsibility for any harmful outcomes of those decisions. As we have a disordered tendency to rank the affection for advantage over the affection for justice, algorithms pose a very real challenge to human agency and responsibility, because of the kind of escape route they provide. Indeed, an updated retelling of Genesis 3 might be “The algorithm which you gave to me told me to do it, and I ate.”

Failure to Recognize Haecceity

We can also add a theological understanding to the failure of epistemic integrity on the part of both end‐users and technologists. We have suggested that in using people's data in the exploitative ways that they often do, technologists can fail to pay “loving attention” to those around them. We have also suggested that end‐users in “filter bubbles” can place higher evidential standards on out‐group members, thereby ignoring their concerns. Both cases constitute a failure to perceive others in the right way.

Returning to Murdoch's notion of “loving attention,” a helpful expansion comes from adapting the scholastic concept of haecceity, which is the property that individuates us and makes us unique. In its theological form, haecceity involves an appropriate recognition of how a particular individual is loved by God in all of their uniqueness (Hare 2015, 145−46). Indeed, the philosophical theologian, John E. Hare (2001, 77) associates this unique‐making or individuating property of haecceity with the promise in Revelation that God will give each person a white stone upon which is written “a new name that no one knows except the one who receives it” (Revelation 2:17, NRSV). Since “God's call to us is to grow into this individual character” (Hare 2001, 77), “loving attention” is the capacity by which we recognize how another individual is loved in all of their uniqueness by God, and how they are coming to enter more fully into this uniqueness.

Returning to the use of Big Data with these hamartiological resources, one might say that ignoring the concerns of outgroup members (in the case of end‐users) or treating people as mere data points (in the case of technologists) are not just instances of mere oversight. They represent a failure to perceive others as God perceives them and intends for us to perceive them, and so come under a kind of failure of “joint attention with God,” which is constitutive of sin (Stump 2012).

The Sins of the Fathers

Finally, the “sins of the fathers” concept (Milgrom 2001; Krašovec 1994) provides a model of how we are all bound together, end‐users and technologists alike, such that our “sin” promotes online structures that encourage others to sin.

The “sins of the fathers” model suggests that the sinful decisions of previous persons can incline present individuals toward repeating the same vicious behavior. To clarify, a seeming tension in Hebraic hamartiology is that at times it appears that the “sins of the fathers” refers to how children are punished for the iniquity of the parents (e.g., Exodus 20:5; Ex.34:7), while at other times it is written that “A child shall not suffer for the iniquity of a parent … the wickedness of the wicked shall be his own” (Ezek. 18:20). The apparent tension is often solved by pointing to the inescapable nature of the cycles of generational sin in question, such that the children suffer for their own freely committed iniquity, but the cycles of sin trap them such that they will ultimately repeat the sins of their parents (Milgrom 2001, 461; Krašovec 1994). Indeed, other Hebraic texts express how the children are only punished for the sins of the fathers “if they hold the deeds of their fathers in their hands” (b. Sanh. 27b), and the Sipra qualifies the verse “Our fathers sinned and are no more; and we must bear their guilt” (Lamentations 5:7, NRSV) with “whenever they adhere to their fathers’ deeds” (Behuqotay 8:2; see also Milgrom 2001, 461). Robert Altar comments on the underlying logic of the hamartiological concept of the sins of the fathers by explaining that “it is often the way of the world for sons to follow the path of their fathers” (Alter 2018, 568). In other words, the sinful behavior of the ancestors sets up structures and schemas that constrain the behavior of subsequent generations, nudging them to replicate this sinful behavior.

With algorithmic bias, there is an inevitability to the cycles we are trapped in: if previous users clicked on biased or harmful sources of information, and these get ranked at the top of our feeds, we will ultimately (freely) click on those sources, further perpetuating the cycle for subsequent users. The biases and behavior of subsequent users are then captured in data sets and used to train further algorithms, showing the inescapability of these cycles.

“The sins of the fathers” concept reveals how responsibility and moral constraints come together. This model also suggests that individual moral agency is constrained in ways that cannot be escaped by current machine learning approaches—the “patterns of sin” simply become more engrained. Indeed, an updated rephrasing of the Hebraic proverb, “The parents have eaten sour grapes, and the children's teeth are set on edge” (c.f. Jer.31:29; Ezek.18:2) may be, “The parents have clicked on fake news, and the children's Google results are set on edge.”

In sum, although our tendency is to avoid blame for the harms done through the use of Big Data, our involvement at any level (whether as user or technologist) constitutes us freely entering into cycles of sin which can serve to entrap us, and others along with us.

Pursue Radical Dispossession

Our first recommendation is to pursue radical dispossession, which involves being set free more generally from the control of harmful Big Data algorithms. Dispossession, in its theological form, works by relinquishing one's own desires that conflict with the desires God intends for them to have, and that conflict with the desires of those one is oppressing—instead placing one's attention upon God and others. Through this redistribution of attention through radical dispossession, one's desires for God and for the flourishing of others undergo a kind of intensification (Coakley 2013; 2015). In other words, before “loving attention” and a respect for the haecceity of others is possible, one must first lay the groundwork through which one can more fully turn one's attention to God and others. Dispossession means laying down privileges and benefits and comes in many forms depending on one's place in the ecosystem.

Consider how this would work for end‐users: some end‐users will not use algorithms that make their lives easier (e.g., e‐commerce recommender algorithms, which promote acquisitiveness), those with unhealthy relationships with technology (e.g., pornography or gambling addicts) will redouble their efforts to divest, and so on. Radical dispossession will not be easy, and may well involve developing “burdened virtues” that in some ways run counter to individual flourishing. “Burdened virtues,” drawing on work from Lisa Tessman (2005), are traits and skills which are virtuous in the sense that they enable resistance to constraints, but are burdened in the sense that they are not fully morally right (in an unqualified sense) and come at a cost. To provide a couple of examples, you could develop burdened virtues of dishonesty and sensitivity (for a deeper discussion of these examples, see Robertson and Johnson 2023). One example of such dishonesty is putting in false details whenever you are using online platforms. This helps to preserve your self‐efficacy—you are holding back information according to your commitment to privacy, and you know that any true information entered in could be used to influence you (and others like you) in targeted advertising. However, this dishonesty challenges epistemic integrity, and it also challenges self‐unity (as motivations and behaviors for self‐efficacy and epistemic integrity are brought into conflict). Another example would be committing to being sensitive to the way things are outside of your own “filter bubble,” and to bearing this truth within your own group. This commitment supports epistemic integrity but could again challenge self‐unity. Sensitivity to others’ suffering is painful and may come with difficult realizations about yourself, and your standing in your own group may suffer as a result of referring to views outside of the group. In other words, resisting constraints in the course of pursuing dispossession may be hard work, and may involve doing characterological damage to oneself in other areas.

Technologists will have to let go of exploitative approaches to data. Computing design needs to go further upstream to optimize for user integrity, aiming at promoting and measuring quality rather than quantity of engagement and user satisfaction. Although initially this may seem to come at a financial cost to technologists, a potential upside is that the technologists who prioritize quality may be seen as more trustworthy by end‐users, which could in turn contribute to financial benefit.

There may also be the possibility to adapt theological principles to move toward more radical forms of dispossession. One possibility is applying the logic of the Jubilee year. The Jubilee year is commanded in Leviticus, and happened every 49 years (or possibly every 50, according to certain interpretations)—occurring at the end of seven cycles of Sabbatical years (which themselves occurred every seven years). The Jubilee prescriptions primarily concern property and labor rights. In this year, slaves and captives would be released, debts forgiven, and property returned, as described in the book of Leviticus: “You shall … proclaim liberty throughout the land to all its inhabitants. It shall be a Jubilee for you: you shall return, every one of you, to your property and every one of you to your family.” (Lev. 25:10) In other words, if someone “falls into difficulty” (Lev. 25:25) and sells a piece of property, or their home, or even themselves and are unable to redeem it or themselves—all of these things are returned in the year of Jubilee.

The application of Jubilee principles is especially salient in the tech ecosystem because technologists are predicting that with a move to the metaverse, some people's full‐time employment would involve spending time online and providing their personal data to companies. Consequently, personal data is very much viewed in economic terms. Applying Jubilee principles would involve personal data being returned, by being deleted. This would also allow us to break free of the digital structures that hold us captive, as it would hit the reset button on all of the ads that previously drew us into the kinds of behaviors we were trying to avoid. The gambling addict, rather than being bombarded with advertisements promoting online gambling, would be faced with different advertisements and have the chance to chart a way out. Of course, if they fall back into old patterns of online gambling, the advertisements they are faced with will then center upon gambling again, but this just shows the necessity of Jubilee years occurring regularly—we need second, third, fourth, nth chances.

We acknowledge that dispossession may be difficult. The burdened virtues involve characterological harm, and regular deletion of personal data could impact on carefully curated personal profiles and the efficiency of recommender systems. Nevertheless, in dispossession, joy is found, as Jonathan Tran argues (2021, 238): “Joy without dispossession is escapist. Dispossession without joy is sadist. The two together order the Christian life. At least the one offered by Jesus ‘who for the joy that was set before him endured the cross, despising its shame, and has sat down at the right hand of the throne of God’ (Heb. 12:2, WEB), and the one racial capitalism provokes.”

Of course, this form of radical dispossession is not easily reconcilable to the current global capitalist order in which the tech ecosystem is situated (to which Tran alludes in the quote above). Considerations of the bottom line fundamentally drive how technologies are designed and deployed. Still, even within this order, there is room for tech leaders and tech companies to practice radical dispossession. In fact, Tran (2021) provides one such example to support his account, through his ethnography of the California Bay Area tech company, Dayspring Partners. Dayspring charts an alternative path through the current global capitalist order, by focusing on relationships and the well‐being of all involved stakeholders, rather than on maximizing financial margins for the company. Through significant investment in and service to their community, their financial profitability arises from their prioritizing relationships and well‐being over profits, as their approach increases consumer trust. Additionally, they enjoy other, nonfinancial goods, such as high employee well‐being, strong relationships with their community, and joy through their practice of radical dispossession.

Finally, recapitulatory atonement models provide resources toward another way to escape harmful cycles and enter into the practice of radical dispossession. On these models, human nature fell through the iniquity of the first man and woman, which constrained human beings into committing subsequent iniquities. The fallenness of human nature is perpetuated and engrained through the sins of all subsequent human beings. Each person is free in any given moment to not sin; however, because they have a fallen human nature, there is no possible world in which they will be able to live their entire life having successfully chosen to not sin in every decision—they suffer from “transworld depravity” (Plantinga 1974). Recapitulatory atonement models suggest that Christ, by joining Himself to human nature and living a sinless life, restores the potential for human nature to live without sin. Through some degree of participation in Christ in the present, believers begin to take on this redeemed human nature and have a greater possibility of avoiding sin—although complete freedom from iniquity will not be achieved until there is full participation in redeemed human nature, in the eschaton.

A computer scientific analogue of restored or redeemed human nature could involve synthetic data: artificially generated data that represents fully virtuous online behavior. Synthetic data can be fed into data‐sets in order to train out the bias from algorithms. With algorithms that contain less bias, users will be surrounded by higher quality information, which will in turn enable them to cultivate and exhibit a higher level of virtue online. Their online behavior will feed into subsequent training of the algorithms, which will enable virtuous behavior in further end‐users, and so on. Instead of relying on gathering personal data and consigning users to cycles of algorithmic bias, synthetic data provides one pathway toward co‐liberation. Indeed, if algorithms trained through this kind of synthetic data could “nudge” human behavior away from cycles of algorithmic bias and vice, and toward radical dispossession and virtue, they could contribute to healthier online lives, deeper human connection, and joy. Technologists and regulators will have to think carefully about what restored human nature looks like, however. This is an area in which dialogue with philosophers and scholars of religion would be particularly useful.

Pursue Embodied Relationships

Alongside pursuing radical dispossession, our second recommendation is to pursue embodied relationships. There are twin dangers with radical dispossession: one danger would be to think of the call to dispossession as involving a kind of denial of the body and detachment from the world, including the tech ecosystem; another would be to embrace the move to the digital world by accepting increasing levels of abstraction and disembodiment, such that consequences for others and ourselves seem small. We are arguing against both of these views, for a form of radical asceticism that involves a move more deeply into our embodiment, and into an integration between online and offline life.

Embodied relationships with others make claims on you as a whole person—not just upon some of the offline parts. When you have a concrete, embodied individual in front of you, they make claims upon you as your friend, your child, your lover, your student, and so on. They do not just want a part of you, and they certainly do not just want some data points. In order to be fully known, there must be integration between all parts of you: the online and offline parts. Consequently, while the majority of the algorithms that structure our online lives view us in reductive and instrumentalized ways, our offline, embodied lives involve friends who aim to view us in all of our haecceity and uniqueness—or to put it theologically, as God views us. We would do well to remember who we are—not just in our offline lives, but in all parts of our lives, including online.

This principle of seeking embodied relationships can also inform the design of algorithms. Technologists need to recognize that algorithms, in one sense, perform similar activities to humans when they learn. Human beings gather data, develop models of the data that are used to predict events, and then adjust these models where those predictions were shown to be inaccurate. Algorithms are just able to accommodate, process, and predict much more data and much more quickly than we can. Theological approaches can inform the boundaries and directions for human learning: there are forms of human learning that are beneficial, such as medicine and art; there are also forms of human learning that are dangerous and forbidden, such as torture. When we approach algorithms, we should apply these same categories. Furthermore, there are better and worse ways of learning. For instance, if we were to treat our teachers in instrumentalizing and reductive terms, we might still learn the information we want, but there will be some kind of deep moral and pedagogical remainder, because it would not have been obtained in the right way. Imagine coming to class and taking no interest in the personal life of your teacher, and just “downloading” the information and leaving. There would be deep human goods, such as friendship, opportunities to exercise service, and even joy, that would be denied to both you and your teacher. There could also be character deficits that arise, as you could become a more self‐centered, small‐hearted person.

If technologists develop their algorithms through treating end‐users in dehumanizing and objectifying ways, treating people as clusters of data points to be manipulated for their own ends, there are goods and character traits that technologists deny themselves, and there are goods denied to the end‐users in being treated this way. Algorithms must be designed with something like haecceity in mind: they need to be optimized to promote things that will help each individual flourish and grow into their individual character. In all this, we need humility about the limitations of algorithms, appreciating that there are some things that Big Data does not currently capture. We need to keep looking to lived experience and to those people whose narratives are usually left out or misrepresented.

In summary, the work of tech ethics will require being in the right kind of listening, caring relationships to concrete, affected individuals—and continually adjusting to the concerns of these individuals. It may not be practical to be in these kind of relationships to concrete, affected individuals—carefully listening to and adjusting to affected stakeholders breeds inefficiency, which is especially challenging for an industry driven by the dictum “Move Fast and Break Things.” Nevertheless, this reorientation is needed for the tech ecosystem to support joy and deeper human flourishing, for it is in a deep alignment within and between our whole selves, others, and the world that joy can be found.