How Tech Workers Contend with Hazards of Humanlikeness in Generative AI

Paper Title

How Tech Workers Contend with Hazards of Humanlikeness in Generative AI

Publication Info

• Topic area: Human-computer interaction and responsible AI, focusing on the risks of humanlike generative AI.
• Keywords: Generative AI, humanlikeness, anthropomorphism, sociotechnical harms, technology workers, responsible AI, human-AI interaction, epistemic challenges, risk mitigation, conceptual mapping.

Background and Problem

• Problem / challenge: The rapid adoption of generative AI with humanlike qualities introduces risks such as overtrust, manipulation, and sociotechnical harms. Current industry standards lack clear guidance for identifying and mitigating these hazards.
• Significance: Understanding and addressing these risks is critical for responsible AI development, ensuring safety, trust, and effective integration into workflows and society.
• Motivation and related work: Prior research has explored anthropomorphism and humanlike design features in AI but has not sufficiently examined how tech workers perceive and navigate these risks. This paper aims to fill that gap by focusing on workers' experiences and conceptualizations of humanlikeness.

Solution

• Proposed approach: A qualitative investigation into tech workers' perceptions of humanlike generative AI and its associated hazards, culminating in a conceptual map to articulate these risks.
• Novelty:
  1. Differentiates between "humanlikeness" and "anthropomorphism," addressing their conflation in prior work.
  2. Identifies six specific hazards of humanlike generative AI, grounded in tech workers' experiences.
  3. Proposes a conceptual map linking humanlikeness features to potential hazards and harms.
  4. Highlights the need for role-specific guidance and cross-functional support in mitigating risks.
• Procedure and key techniques:
  • Conducted focus groups with 30 U.S.-based tech workers across six job roles (e.g., ML engineers, product managers, UX researchers).
  • Used thematic analysis to identify workers' conceptualizations of humanlikeness and perceived hazards.
  • Synthesized findings into a conceptual map distinguishing humanlikeness features, hazards, and anthropomorphism.

Results

• Concrete findings:
  • Workers identified six key hazards of humanlike generative AI: additional labor, miscalibrated trust, believable content, worsened user experience, disparate AI literacy, and worker replacement.
  • Workers conceptualized humanlikeness through interaction dynamics, conversational input, humanlike outputs, task sophistication, and anthropomorphizing tendencies.
• Advantage over baselines: Provides a nuanced understanding of humanlikeness hazards, moving beyond simplistic views of anthropomorphism and offering actionable insights for responsible AI development.
• Experiments / evaluation:
  • Focus groups with 30 participants representing diverse roles and organization sizes.
  • Thematic analysis of transcripts to identify patterns in workers' reasoning and hazard perception.
• Limitations and future work:
  • Limited generalizability due to purposive sampling and focus group dynamics.
  • Findings may not capture the full range of perspectives across all tech roles or industries.
  • Future work should empirically validate the conceptual map and explore hazard mitigation strategies across different contexts.

Summary

This paper investigates how tech workers perceive and navigate the risks of humanlike generative AI, identifying six key hazards and varied conceptualizations of humanlikeness. Through focus groups and thematic analysis, the authors provide a conceptual map linking humanlike features to potential harms, offering a foundation for responsible AI development. The study highlights the need for clearer guidance, cross-functional support, and role-specific strategies to mitigate risks, emphasizing the importance of addressing humanlikeness as a multifaceted construct.