"Look! It's a Computer Program! It's an Algorithm! It's AI!'': Does Terminology Affect Human Perceptions and Evaluations of Algorithmic Decision-Making Systems

Title of the Paper

“Look! It’s a Computer Program! It’s an Algorithm! It’s AI!”: Does Terminology Affect Human Perceptions and Evaluations of Algorithmic Decision-Making Systems?

Paper Information

• Domain: Human-Computer Interaction (HCI), Algorithmic Decision-Making (ADM) Systems, and Terminology Effects
• Keywords: ADM systems, Artificial Intelligence, Human-Computer Interaction, Terminology Usage, Research Methods, Trust, Fairness, Complexity
• Conference: CHI Conference on Human Factors in Computing Systems (CHI ’22)
• Publication Year: 2022
• Authors: Markus Langer, Tim Hunsicker, Tina Feldkamp, Cornelius J. König, Nina Grgić-Hlača

Research Background and Problem

• Issues and Challenges:
  • The terminology used to describe Algorithmic Decision-Making (ADM) systems in media, policymaking, and academic research varies widely, including terms like algorithm, artificial intelligence (AI), and computer program. This diversity may influence public perceptions and evaluations of ADM systems.
  • Inconsistent terminology can reduce the reproducibility of research findings and impact public trust and acceptance of ADM systems in different contexts.
• Significance:
  • As ADM systems are increasingly applied in fields such as healthcare, recruitment, and judicial decision-making, terminology choices may elicit different public reactions, influencing policymaking and research outcomes.
• Motivation and Related Work:
  • Existing literature suggests that terminology choices can affect perceptions of system attributes, such as complexity and trustworthiness. This effect is evident in various domains, including technology trust and AI interaction.
  • The authors aim to explore the specific impact of terminology differences on people’s perceptions and evaluations of ADM systems, emphasizing the importance of terminology use in research methodologies.

Solution

• Research Methods:
  • The authors conducted two experimental studies to systematically examine the effects of terminology usage.
    • Study 1: Investigated how terminology differences influence perceptions of ADM system attributes (e.g., complexity, familiarity) and their ability to perform various tasks.
    • Study 2: Specifically analyzed how terminology affects evaluations of fairness, trust, and procedural justice in workplace assessment and task allocation contexts.
• Key Techniques and Implementation Steps:
  • Experiment Design and Terminology Selection:
    • Using Google’s Universal Sentence Encoder to analyze semantic similarities between terms, the authors categorized terms (e.g., “AI,” “statistical model”) and selected representative terms for the experiments.
  • Study Variables:
    • Artificial Intelligence (AI) was used as the reference group, with other terms (e.g., “robot,” “computer program”) tested for their effects.
  • Multiple linear regression and linear mixed models were employed to analyze the impact of terminology on different attributes and task evaluations.

Research Findings

• Summary of Findings:
  • Terminology differences significantly influenced participants’ perceptions of ADM system attributes, such as:
    • Complexity: “Artificial intelligence” was perceived as more complex than “automated system.”
    • Familiarity: “Computer” was perceived as more familiar than “artificial intelligence.”
    • Trust and Fairness: In workplace assessment contexts, “statistical model” was trusted more than “artificial intelligence.”
  • For specific task evaluations, terminology differences did not significantly impact perceptions of ADM systems being superior to humans.
• Advantages:
  • The study highlights the significant influence of terminology choices on perceptions, trust, and fairness evaluations of ADM systems, providing valuable insights for experiment design, policymaking, and information dissemination.
• Limitations and Future Directions:
  • Limitations:
    • Data collection relied on self-reports, without directly measuring the impact of terminology on actual behavior.
    • The sample was drawn from an online platform (Prolific), limiting representativeness.
    • The study was conducted in an English-speaking context, and terminology effects may vary across languages.
  • Future Research Directions:
    • Investigate the direct impact of terminology differences on behavior, such as decision-making during actual ADM system interactions.
    • Extend research to different languages and cultural contexts to explore the universality and limitations of terminology effects.
    • Develop unified guidelines for terminology selection to minimize the potential impact of terminology differences on research outcomes.

Through this study, the authors emphasize the importance of terminology choices and their potential strategic implications (e.g., influencing public attitudes, increasing engagement), providing a reference framework for effectively utilizing terminology in policymaking, technology adoption, and research design.