SustAInable: How Values in the Form of Individual Motivation Shape Algorithms’ Outcomes. An Example Promoting Ecological and Social Sustainability

Literature Title

SustAInable: How Values in the Form of Individual Motivation Shape Algorithms’ Outcomes. An Example Promoting Ecological and Social Sustainability

Literature Information

• Field of Study: Human-Computer Interaction (HCI), Algorithmic Bias and Ethics, Ecological and Social Sustainability
• Keywords: Humanism, Sustainability, Algorithmic Bias, Quantitative Methods, Human-Computer Interaction

Research Background and Problem

• Identified Problem or Challenge:
  Algorithm outputs are not solely driven by data and technology but are also influenced by the personal motivations and values of developers and users. This influence can lead to algorithmic bias, unintentionally reflecting tendencies such as ecological or social values. However, understanding of these influences remains limited.
• Significance:
  Algorithmic systems are increasingly critical in many societal domains (e.g., finance, human resource management, cybersecurity). Potential biases in these systems could impact fairness and the realization of public interests, necessitating an in-depth exploration of the roots of such biases.
• Research Motivation and Related Work:
  Existing research has discussed types of biases and their causes (e.g., data bias, user bias) and has attempted to develop technical or legal methods to mitigate these biases. However, there is limited research on whether the personal motivations of developers and trainers also influence algorithm outputs. This study aims to fill this gap and explore its potential applications.

Solution

• Research Methodology:
  The authors designed an online survey to collect data from 766 participants, investigating how environmental and altruistic motivations influence participants’ scoring decisions when training a spam filter algorithm. Correlation analysis and regression models were used to study the impact of these motivations on spam classification.
• Innovative Aspects:
  This study is the first to systematically explore how the personal motivations of developers and users influence decision-making during algorithm training, thereby introducing bias into algorithm outputs. Furthermore, the research expands the concept of algorithmic bias to include potential ecological and social sustainability impacts, rather than limiting it to traditional negative "discriminatory biases."
• Implementation Steps and Techniques:
  1. Design the survey and collect data: Participants classified eight emails as spam, including emails from environmental organizations, charitable organizations, and other personal emails.
  2. Measure independent variables: Environmental motivation was measured using the General Ecological Behavior Scale, and altruistic motivation was measured using the Self-Report Altruism Scale.
  3. Conduct correlation analysis and multiple linear regression to evaluate the impact of motivations on spam classification rates.

Research Findings

• Specific Results:

  1. Environmental motivation was negatively correlated with the frequency of marking emails from environmental and humanitarian organizations as spam. Participants with stronger motivations tended to reduce spam classification for such emails.
  2. Although there was some correlation between environmental and altruistic motivations, altruistic motivation had no significant impact on email classification.
  3. Environmental motivation demonstrated a broad impact on social behavior, not limited to environmental actions.

• Comparison with Existing Solutions:
  Compared to current research on algorithmic bias, this study focuses on the motivational effects behind biases, which are not necessarily harmful, and considers their potential to favor "common good" outcomes. The study suggests that beyond technical solutions, the values of individuals influencing algorithms should be considered and addressed through education.

• Experimental or Evaluation Results:

  • Environmental motivation was significantly negatively correlated with spam rates (Environmental: r=-0.19; Humanitarian: r=-0.14).
  • Regression analysis showed that environmental motivation had a significant negative predictive effect on spam classification rates (e.g., predictive effect for environmental emails’ classification rate β=-0.22, p<0.001).
  • Motivation-driven biases can yield unexpected results, such as enhancing ecological and social sensitivity in spam filtering.

• Limitations and Future Directions:

  1. The effect of altruistic motivation was not significant, possibly due to measurement methods or material design. Future research should optimize scales and email samples.
  2. The study used a cross-sectional design, which cannot infer causality and only indirectly indicates relationships.
  3. This study focused on spam filters, a relatively narrow scenario, and future research should extend to more complex algorithmic systems.
  4. Motivation-driven biases may have more profound impacts in other real-world scenarios, warranting further exploration of how to apply this research in various decision-making environments.

Practical Implications

• Recommendations include fostering developers’ sustainability values, such as implementing relevant training or establishing a culture of social responsibility and environmental protection within tech companies.
• This study provides theoretical support for designing algorithms that promote social and ecological values (e.g., intelligent recommendations, credit scoring).
• It also reminds algorithm creators and regulators to monitor and educate about motivation-driven biases to prevent negative impacts on training data.

Conclusion

This study reveals how personal motivations influence algorithm outputs through the training process, opening the door to designing algorithms "biased" toward social and ecological sustainability. The authors emphasize raising awareness of motivation-driven biases while proposing the dual approach of leveraging motivations that promote public interest and preventing negative impacts from unintended biases. This has significant implications for designing algorithms that better align with human values.