1 Introduction
The evolution of artificial intelligence (AI) is advancing, and yet, algorithmic systems are no longer a prospect of the future but have already become an integral part of the lives of many. Their reach has expanded from mundane applications (e.g., systems recommending music) to systems that make potential life-changing predictions about delinquents' recidivism rates [
2]. In the
financial sector, for instance, robo-advisors move significant financial assets. In human resource management, algorithmic systems support personnel selection; and in cybersecurity, they are used to detect cyber-attacks. Given the wide-ranging pervasiveness of algorithms, discussions about how to ensure that these systems serve the welfare of society at large are increasing. Such discussions include the desire for them to be fair, so they do not systematically favor or disadvantage certain groups or individuals on the basis of inherent characteristics [37]. Yet, algorithms have produced biased output and have systematically discriminated against specific groups of people [41, 43]. To understand the emergence of such biases and to work toward their mitigation, interventions have been developed (e.g., in the form of legislation [14]), and intensive research has examined various sources of bias and their mitigation—either technical [43] or by raising awareness of potential sources of bias [13]. Several scholars and practitioners are debating whether a “code of ethics” should be programmed into algorithmic systems to ensure that the generated output is aligned with human values [59], thus giving rise to a growing strand of research focusing on machine ethics, e.g., [50]. Bringing the issue of values in AI to the table marks an important starting point. However, we argue that, in many cases, regulation and guidelines are not enough to obtain nondiscriminatory algorithms. Although such regulations can serve as a framework, they cannot cover all aspects and decisions involved in the development of algorithms. Naturally, there are still many degrees of freedom in the development and training of algorithmic systems, and it is impossible for anyone to foresee all eventualities. However, on a higher level, the goals of an algorithm are largely determined by humans—primarily the developers, entrepreneurs, and users who create, disseminate, and train these systems. Given this context, understanding the extent to which humans can influence algorithmic output is paramount. We argue that a stakeholder’s motivation will find its way into an algorithm and its output—even independent of the algorithm's primary goal. Motivation refers to the psychological driving force behind behavior [16]. In many contexts (e.g., in the context of environmental motivation), motivation is based on a person's values [24, 29]. In other words, motivation directs a person's inherent values into behavior (e.g., a decision). In this study, we focus on a spam-filter algorithm—a case where, in training, individuals may tend to make value-based decisions that are driven by their motivation. A spam filter has the primary objective to reduce the amount of spam (i.e., unsolicited messages sent to many users [38]) that lands in users’ inboxes. However, there will be many (spam) emails whose classification as spam is highly dependent on a person's individual preferences (e.g., values), which form a person's motivation and, hence, guide their behavior [23]. Such motivation is continuously and often unconsciously at work in individuals’ daily lives and work lives. Applied to the context of algorithms, it can thus be assumed that a person's motivation can influence their decisions while developing and training algorithms. An infamous example of motivation that made it into negatively motivated speech is Microsoft's AI chatbot "Tay," which rapidly descended into bigotry on Twitter after interacting with users. Conversely, it is also likely that stakeholders with a strong concern for the welfare of others influence the algorithmic decision process in such a way that its output is biased toward the common good. For instance, a company or a team of developers who are strongly environmentally motivated would perhaps set the default for a navigation system to choose the route with the lowest CO2 emissions, just as they would in “real life” [57].Algorithmic output is tied to the data on which the algorithm was trained, as the algorithm learns patterns from its training data and applies these patterns to make predictions in new data [
17]. Thus, if these data reflect biased decisions, the output will be biased as well. It is likely that individuals for whom social and ecological sustainability is a core value that is reflected in their altruistic motivation (i.e., the motivation to enhance the welfare of others [
4]) or environmental motivation (i.e., the motivation to benefit the natural environment, see, e.g., [
33]) also make decisions that may shape the algorithmic output toward sustainability while developing and training algorithms.
Therefore, the aim of this study was to examine how individual differences in stakeholders' environmental and altruistic motivation could—through the rating of training material—influence an algorithm's decisions and thus its outcome. More precisely, we examined whether environmental and altruistic motivation are related to decisions when training a spam-filter algorithm. The role of stakeholders’ motivation in algorithm training has been largely unexplored thus far, and understanding these influences can help to either mitigate biases arising from value-based decisions, or, by contrast, build on this effect to design algorithms that serve society as a whole. Our study contributes to research on algorithmic bias by addressing the following research questions:
Can stakeholders’ (i.e., those who develop or train an algorithm) motivation influence an algorithm's output?
a.
Does environmental motivation influence individuals’ decisions when training an algorithm?
b.
Does altruistic motivation influence individuals’ decisions when training an algorithm?
4 Results
The correlations of the independent variables with the spam ratios are presented in Table
2. Altruistic and environmental motivation were correlated to a small to medium extent (
r = .26***). In support of H1 and H3b, environmental motivation was further negatively correlated with categorizing emails from environmental organizations (
r = -.19***) as well as humanitarian organizations (
r = -.14***) as spam. That is, individuals with a more pronounced environmental motivation showed a tendency to categorize emails from environmental (H1) and humanitarian organizations (H3b) as spam less often than those with less environmental motivation. Contrary to our expectations in H1 and H3a, altruistic motivation was not significantly correlated with the categorization of emails from environmental or social organizations.
We further calculated correlations between altruistic and environmental motivation and the eight email ratings. Table
3 presents the results.
The correlations between environmental motivation and the ratings of emails from humanitarian and environmental organizations were consistent for both personalized and nonpersonalized emails. Moreover, a positive correlation was observed between environmental motivation and ratings of the genuine spam email as spam (r = .13). Altruistic motivation was not correlated with either of the ratings of emails from environmental or humanitarian organizations. However, it was slightly negatively correlated with the spam rating of the tax authority's email (not spam, r = -.10*) and the personalized email from a bank (r = -.11**). Thus, individuals with higher levels of altruistic motivation were less likely to classify these two emails as spam. Furthermore, predominantly significant positive correlations were found between the different email ratings.
To examine the extent to which environmental and altruistic motivation jointly predicted spam categorization, we ran two separate multiple linear regression analyses after determining that the assumptions (homoscedasticity, normally distributed residuals, and no multicollinearity) were met. These analyses aimed to explore the relationships between these types of motivation and the two spam ratios. The results of these regression analyses are presented in Tables
4 and
5, respectively.
The model in Table
4 significantly predicted the spam ratio (environment),
F(2, 689) = 16.3,
p < .001,
R2 = .05, with environmental motivation being the strongest predictor, β = -0.22,
p < .001. When the two predictors were considered jointly, altruistic motivation was positively related to categorizing emails from environmental organizations as spam, but the effect was very small (β = 0.10,
p = .011).
The model in Table
5 significantly predicted the spam ratio (humanitarian),
F(2,704) = 7.11,
p < .001,
R2 = .02. Again, environmental motivation was the strongest predictor, β = -0.15,
p < .001, whereas altruistic motivation did not significantly predict the spam ratio for emails from humanitarian organizations.
The aim of this study was to investigate the influence of individuals’ motivation, specifically environmental and altruistic motivation, on the classification of training material for an algorithm. Our findings revealed that individuals with a more pronounced environmental motivation were less likely to classify emails from environmental or humanitarian organizations as spam. By contrast, although correlated with environmental motivation, altruism was not related to the classification of emails from either environmental or humanitarian organizations as spam. In fact, when examined jointly with environmental motivation in the regression analysis, altruistic motivation even exhibited a small positive association with the categorization of environmental emails as spam.
Our results show that individuals’ environmental motivation influenced their decisions during the rating task, thus implying that, depending on their motivation, individuals would have trained an algorithm to permit more emails from environmental and humanitarian organizations to pass through. This finding is in line with Lammert's [
31] expectation that software engineers’ values affect the sustainability outcome of a software product. Furthermore, our results show that also in the training of algorithms, individuals’ motivation influences individuals' behavior and, more specifically, their decisions, a finding that is consistent with the general expectation that motivation and attitudes influence individual behavior [
23]. Thereby, our study provides empirical evidence for the emergence of the
user-to-data bias [
37]. Furthermore, the relationship observed between environmental motivation and decisions during the algorithm training task is aligned with existing research on involvement [
16]. Individuals' involvement (i.e., their value-based motivation), particularly in areas of personal relevance (e.g., environmental issues), can significantly sway their decision-making process toward fostering value-based decisions during algorithm training. Our findings further corroborate research on social media content sharing [
12], where personal values have been found to guide sharing decisions, thereby creating a bias in the transmission of information [
52]. The present study extends this understanding to the domain of algorithm training, suggesting that similar principles of value-based decision-making apply. Just as individuals are more likely to share content on social media when it is related to topics that are personally relevant to them, they also tend to train a spam filter algorithm in ways that correspond to their motivation.
Contrary to our expectations, altruistic motivation did not have an impact on spam classification, a finding that might be explained by several factors. The task of classifying emails might not have been perceived as an opportunity to benefit other humans. As the correlation table shows, participants with a higher altruistic motivation were less likely to classify genuine personal emails (e.g., those from the financial administration or a bank) as spam. This could indicate a focus on the welfare of the recipients, ensuring that legitimate personal emails do not end up in the spam folder. Furthermore, perceptions of humanitarian organizations might have influenced the results. Participants may have held a more negative view of humanitarian organizations than of environmental organizations (visible in an overall higher rate of classifying emails from humanitarian organizations as spam versus emails from environmental organizations). Due to negative press on some organizations regarding the mismanagement of donations, e.g., [
9], participants with a higher altruistic motivation might have aimed to prioritize protecting email recipients over promoting humanitarian organizations. This tendency is supported by the finding that altruistic motivation was not related to classifying emails from environmental organizations as spam either. In addition, when examined jointly with environmental motivation, altruistic motivation was even positively related to categorizing environmental emails as spam. Thus, the aim of helping other individuals might have been channeled differently in this categorization task.
The finding that individuals with a more pronounced environmental motivation were also less likely to classify emails from humanitarian organizations as spam is consistent with research on the relationship between prosocial and pro-environmental behavior [
40,
44]. It is likely that environmentalists have a more inclusive self-concept, that is, they more easily identify with human and nonhuman others [
21]; therefore, they want not only to protect the environment but also to benefit other humans. Thus, this view could extend to a more favorable view of social organizations and a lower likelihood of classifying their emails as spam.
To summarize, our findings provide the following main contributions to the literature on algorithmic biases: Our study emphasizes the need to consider the motivation (and on a higher level, the values) of stakeholders of an algorithmic system and their role in the emergence of bias. A growing body of research is dedicated to discovering and developing technical methods for minimizing biases in training data, [
37]. Our research shows that in addition to technical means, it is also important to consider the workforce by taking into account not only their diversity [
3] but also their motivation and underlying values. Furthermore, our findings introduce a new perspective: examining algorithmic biases as an opportunity to align the algorithms with other values if needed. Our study underscores the idea that algorithm development is not just about minimizing algorithmic biases but also about understanding and potentially leveraging the algorithms so that they are aligned with broader societal values. For instance, to assess creditworthiness, a credit scoring algorithm typically focuses on historical financial data. However, if the developers value financial inclusivity, they might incorporate additional (not typically used) data points (e.g., rental history) in the model. Such inclusivity would offer better opportunities to certain groups of people who are disadvantaged by other scoring algorithms.
4.1 Expanding the scope beyond spam filters
In our study, we chose spam filters as a specific case study due to their relatively narrow but rather controlled environment. The behavioral shifts observed in the spam filter training task may significantly impact the flow of information when scaled to larger systems. In these broader contexts, slight biases, akin to those in this study, can impact the visibility and dissemination of content, thereby influencing public access to information. While the specific task of categorizing emails may seem limited in scope, it mirrors the decision-making processes found in many more complex algorithmic systems. In other, more complex contexts, the basic mechanism will be quite similar if not the same because even though an amalgamate of several motives (and values) can be at work in complex contexts, motives as well as values are usually compensatory, and thus, they simply add up independently.
Just like spam filters, such systems often involve sorting and classifying information. For instance, large language models learn through curators’ ratings of whether the content is appropriate or not. In the realm of social media, content moderation algorithms function similarly to spam filters to determine the appropriateness of content. Similarly, news feed algorithms curate a user's news stream on the basis of perceived interests. In recommendation systems—whether for e-commerce, streaming services, or online advertising—the sorting and recommending of items to users is another domain where the decision-making process can be subtly influenced by the values and biases of the stakeholders.
Whereas the systems described above are algorithms that are similar to spam filters with respect to their training (i.e., information is sorted by humans), the fundamental assumption that values can influence algorithmic output through motivation is applicable to a variety of contexts. Our research focused on content curators; however, similar effects could emerge among developers, users, and other stakeholders. Developers’ motivations could influence their decisions in the development process in a similar manner. On the other hand, user feedback further shapes algorithmic decision-making. Thus, the supposedly small influences of the values and motivations of certain groups could have a large impact all together.
4.2 Practical implications
Our findings contribute to the understanding and management of biases in algorithmic systems. Recognizing these underlying values can help developers and policymakers create algorithms that are sustainable and socially responsible. For instance, it may be instrumental to integrate sustainability-focused training and interventions into the educations or workplaces of those who develop or curate algorithms. Fostering a culture of environmental stewardship and social responsibility in tech companies might provide a basis for these motivations to be reflected in the decisions made during the development and training of algorithms and, thus, algorithmic output. Whereas on the one hand, our results suggest that the influence of individual values on algorithmic output could be utilized to promote positive outcomes, see also [
61], our results also highlight that values may also have unintended consequences in the form of unintended biases that need to be mitigated. For instance, clients who want to apply an algorithm might not want sustainability motivation (or other motivations) to influence training decisions. The biased selection of training material, as observed in our study, echoes concerns similar to those found in research on social media, such as the selective sharing of posts. For instance, Shin and Thorson [
52] showed that partisans tend to spread content that favors their perspective while criticizing the opposition—a bias that could similarly affect algorithms responsible for content selection and dissemination. Thus, raising awareness about such value- or motivation-based biases is crucial for developing strategies to reduce them. For instance, it may be necessary to discuss (a) whether and how the influence of stakeholders’ values and motivations could or should be taken into account in regulations and guidelines pertaining to algorithm development and (b) potential ways to reduce such influences. In their recent work, Hardy et al. [
18] adapted a content selection algorithm to be used in a social network. The aim was for the algorithm to rearrange the content from individuals’ personal networks in such a way that it was representative of the perspectives of the population. This approach enables a more balanced flow of information and reduces echo chambers.
4.3 Limitations and future research
There are several limitations to consider when interpreting the results. First, the fact that we did not find an influence of altruistic motivation on the categorization of emails might be due to the measurement method and the material we used. The scale we used is widely applied to assess altruism; however, it might not have adequately captured the nuances of altruistic motivation that would affect spam ratings, or the ratings might be influenced by a negative image of humanitarian organizations. Future studies should assess this relationship using different measures and by using different and more diverse stimulus material. Second, we used a cross-sectional design; thus, the directions of the relationships can be derived only from theory. Third, our study focused on a specific use case, and the effect sizes were rather small. We chose spam filters as a specific case study, which, while narrow, provided a more controlled environment to observe the subtle influence of values on decision-making. Recognizing and understanding these effects lays the groundwork for further research that can investigate the implications of such value-based decisions in broader, real-world contexts where decisions are made at scale and where their impacts can be far-reaching.