Whom does facial recognition recognize? Who gets recommended for a job by an AI hiring algorithm? How does bias shade the output of an image-generation model? Whose choices are reflected in all of the multifaceted decision-making that is part of generating and deploying AI models?

In an AI-saturated world, the issue of bias in artificial intelligence presents a new challenge for Catholic thought. We think that a helpful starting point for addressing this challenge is the idea of “multivalent structural sin” articulated by American Catholic theologian Kristin Heyer in her 2024 address titled“ ’Hearts of Flesh’: Structural Sin and Social Salvation.”

Bias is a manifestation of sin: A way of seeing and choosing and acting that denies the inalienable and equal dignity of individuals and groups of human persons. Artificial intelligence can entrench and amplify bias through many factors: the composition of the vast data sets required to train AI tools; the varied choices made in the process of training and deployment of such tools; the claims of objectivity that often accompany the use of AI; and the limited public understanding of the way in which AI models operate. Heyer’s work acknowledges the role of original sin and personal sin in creating social sin, but focuses on the way in which social sin takes on a kind of agency of its own: Sinful structures themselves act on persons. And Heyer argues that social sin is “multivalent”—constituted by three dimensions: structure, culture-ideology, and the interaction of responsibility and habitude.

In the case of bias and AI, Heyer’s notion of “structure” points us to the power of those who make decisions that shape the development and deployment of AI tools, especially given the absence of regulatory regimens (and the unequal power of those whose data is being used or who are subject to automated decision-making). By the category of “culture-ideology,” she prompts us to assess the narratives that claim that AI is objective or more discerning than humans. Finally, by “responsibility and habitude,” Heyer helps us see the ways in which bias in AI systems and stories may interact and act back on persons, fostering habits of seeing and acting that shape both developers and users of AI—as well as those who fall in neither of those groups but who are nevertheless impacted by algorithmic decisions.

Structure of AI bias

We began this essay with questions about facial recognition, algorithms deployed to make decisions about hiring, and ways in which bias might enter conversations with chatbots (including, for example, in requests for generated images). These examples illustrate structural factors in the problem of AI and bias.

Facial recognition is not perfect, and it’s not deployed in an unbiased world. As of April 2026, at least 14 people in the U.S. had been wrongfully arrested based on erroneous facial recognition results that had been treated as accurate by police; one of them, a grandmother, was jailed for a total of six months as a result of the misidentification (and failures in the process that followed). Facial recognition tools have been shown to produce higher inaccurate match rates for different groups of people—including women and the young or the elderly.

Hiring algorithms are not perfect, either. In May 2026, researchers from Stanford University in California released a study analyzing the records of 3.4 million people who had submitted 4 million job applications, all of which had been assessed by the same AI hiring software; the analysis showed “that these tools increase racial bias and shut the same people out of jobs everywhere they apply.”

A few years ago, a project titled “Models All the Way Down” investigated the LAION-Aesthetics dataset—one of several open-source datasets developed for the training of image-generating models that had been publicly released by a German non-profit organization. The project’s authors demonstrated that “[t]he concepts of what is and isn't visually appealing can be influenced in outsized ways by the tastes of a very small group of individuals, and the processes that are chosen by dataset creators to curate the datasets.”

Artificial intelligence tools, like all human creations, embed and reflect human choices. And beyond the immediate choices of their creators, they are also shaped in part by constraints that reflect prior choices made by humans long ago. For example, the architecture of the internet, and the way in which websites have long allowed for the scraping of content uploaded by individual users, enabled the pooling of vast numbers of images into the kinds of datasets that are necessary for training image-generation AI.

But some people are more likely than others to have access to the internet, more likely to have cameras, and more likely to upload their images; not just their choices about what is aesthetically pleasing, but their choices of whom or what to photograph, shape the output of the image-generating tools. Amid often-heard claims that AI is “trained on all of human knowledge,” there is a need to push back, and to ask whose knowledge, whose images, whose voices, whose choices might be missing or barely represented in frontier models that encompass so much.

And data itself is not commensurate with objectivity. As thoughtful researchers have argued, data is a construct, and the “process of constructing data builds social values and patterns of privilege into the data. Where those values and privileges are unjust, the injustice is then a characteristic of the data itself.” Think of arrest data, for example, and the choices embedded in it, about whom to arrest, and for which crimes. In the U.S., those choices are then reflected in algorithms used in the criminal justice system, which attempt to assess the likelihood that a person will re-offend, or determine the length of a person’s sentence.

Beyond data, model design and training choices shape the output of AI tools as well. For example, research has shown that even the length of training of a particular model can “disproportionately impact error rates.” As it turns out, “challenging and underrepresented features are learnt later in the training process... Thus, early stopping and similar… choices disproportionately and systematically impact a subset of the data distribution.” That subset reflects categories for which there is less data; in the case of humans, those whose identities, experiences, voices, etc. are less well represented.

Many researchers who work on AI models are quite aware of and passionate about this issue, and work to combat bias in their products. But large AI models are vastly complex systems whose success has been described by some computer scientists themselves as “alchemy,” and whose workings and shifts can surprise even their own developers. As new versions of models get released, advantages in some areas can be accompanied by retrogression in others: researchers looking at dialect discrimination in ChatGPT, for example, found that “GPT-4 improves on GPT-3.5 in terms of comprehension, warmth, and friendliness, but also exhibits a marked increase in stereotyping (+18%).”

The level of bias waxes and wanes. And when it comes to AI, bias doesn’t come in only through the data and design choices.

The very choice of what to target as a problem to be tackled with AI, too, can reflect bias. Consider the concept of “predictive policing”—the effort to predict where crime will happen, so as to allow law enforcement to respond more promptly when it does. As a critique, the developers of a project called “While Collar Crime Risk Zones” used machine learning to predict the most likely location for financial crimes in the U.S., training their model on data about incidents of such crimes since 1964. The point, of course, was that governments don’t usually aim predictive policing at white collar crimes, and not for lack of data.

In the past decade, governments in multiple countries have implemented algorithmic tools in an effort to root out welfare fraud (often with massive harmful consequences for the vulnerable people who rely on welfare); we are, however, much less likely to see algorithmic tools deployed to identify political corruption or loopholes in the tax systems that allow those most affluent to pay less.

Culture-Ideology of AI

Although there is growing awareness about the limitations of AI and the complex interactions among AI tools and diverse societal contexts in which they are deployed, the use of such tools has been accompanied by certain narratives and beliefs which continue to hold sway. For example, too often the deployment of such products has been accompanied by an assumption that they work as intended—often disproved only after many people were harmed. It has also been accompanied by the claim (and hope) that AI is objective: “data-based” decision-making has long been contrasted to decision-making based purely on intuition or personal bias, and machine learning processes data through multiple layers of math, which, again, connotes objectivity. As noted above, however, subjectivity enters AI development and deployment at many points, which explains why some critics have referred to the claim of AI objectivity as “bias laundering.”

And new claims continue to enter the discourse: for example, that “synthetic” data (generated by AI) will be able to accurately mimic human data for various purposes, including research about how humans think, communicate, or collaborate; that the deployment of AI is an unstoppable force; that such deployment will ultimately lead to human flourishing; or that AI is becoming more like humans and thus might itself deserve dignity and rights.

The issue of AI bias, however, forces us to consider all such claims through the lenses of justice and the common good.

Habitude in a world increasingly shaped by AI

AI indeed shapes our habits and ways of seeing and choosing. But, in Heyer’s model, such shaping does not preclude the possibilities of choice and responsibility. And there are important efforts underway to respond to the problem of AI bias as a matter of habitude.

While biased AI tools now shape the lives of many people around the world, growing awareness of the problem of AI bias is also leading to increasing efforts to reshape the AI systems themselves. There are multiple parallel efforts to bring more of the world’s knowledge into datasets needed for AI tools: for example, the Mozilla Foundation has been building a “Common Voice” dataset which now includes more than 100 languages–data collected with the support of more than 400,000 volunteers. There are also efforts to develop entire models better aligned with different parts of the world: one example is Latam-GPT, which aims to be “a model specific to Latin America and the Caribbean, aware of the cultural requirements and challenges that this entails, such as understanding different dialects, the region’s history, and unique cultural aspects.”

Of course, any such models and datasets will come with their own biases–but also with a built-in critique of the claim of AI objectivity.

In a book titled Fairness and Machine Learning, researchers Solon Barocas, Moritz Hardt, and Arvind Narayanan detail the nature and intricacies of the challenge of making AI more fair, but add that “the turn to automated decision-making and machine learning offers an opportunity to reconnect with the moral foundations of fairness.” They argue that machine learning “has the potential to help us debate the fairness of different policies and decision-making procedures more effectively.” Still, they caution that work on fairness in AI will not lead to easy answers.

Indeed. As of now, bias is still a key issue, and it is only part of the unfairness stack. If we want to have more fairness in AI, we should also push back against the notion that individual creators’ work and all of our personal communications are fair game to be “harvested” as training data for large models. We should demand better payment for data labelers and contractors who are now generating snippets of code or text to be used as better training data. We should demand that any AI systems that in any way impact the social safety net be tested carefully in pilot programs and independently audited before they are deployed at scale. We should reject the targeting of algorithmic tools in ways that make the criminal justice system less just. We should ask that the infrastructure costs created by AI-optimized hyperscale data centers be borne by the companies that are benefitting the most from the AI boom. We should also demand careful the location of any new data centers, so that vulnerable and less politically powerful communities don’t have to bear an unfair burden in terms of energy and water consumption and public health implications that come with such development.

Conclusion

If the collection of some types of data sets already risked reducing human beings to data points unless done with great discernment, the processing of such data by large language models might push us even further toward the stripping away of human agency, rights, and dignity. And the deployment of automated decision-making in certain aspects of society might propel us toward what Pope Francis called the greatest sin of our time: the “globalization of indifference.”

Aside from indifference, a sense of powerlessness in the face of entrenched problems such as bias in AI (and not just in AI) can lead to the related threat of apathy. In “Hearts of Flesh,” however, Heyer also offers a timely three-pronged prescription for responding to apathy: “First,” she writes, “shift methodological foci from parsing complicity toward taking responsibility; second, allow disorientation(s) to tenderize our hearts; and third, center the protagonism of those marginalized.”

What is being described as the “age of AI” offers us all plenty of disorientation. Such disorientation might be perceived by some as acause of apathy, but Heyer chooses to see it as a means of pushing us out of existing structures and habits. Her prescription is a call to action.

It’s been ten years sinceProPublica journalists published an article called “Machine Bias”: “There’s software used around the [U.S.] to predict future criminals,” they wrote in 2016, "[a]nd it’s biased against blacks.” Many people have become more aware, since then, about ways in which AI can perpetuate and amplify inequalities, but there is much work still to be done.

Authors

Cover Image Credits: The practice of requiring employees to wear cameras to record their activities is becoming increasingly widespread. | Photo: R Satish Babu/AFP/Getty Images | Source: La Razón (Published on 3 July 2026)