A new 19‑item scale finds stronger human–AI interaction correlates with better task performance; gains appear to run through strengthened role identity and self‑efficacy, but evidence is observational rather than causal.

Main Finding

The paper develops a multidimensional construct and 19‑item scale for workplace human–AI interaction and shows that human–AI interaction is positively associated with employee task performance. Role identity and self‑efficacy (self‑concept components) partially mediate that relationship.

Key Points

• Conceptualisation: Human–AI interaction is framed via actor–network theory (AI as an actor) rather than traditional human–machine views.
• Construct: Grounded theory and follow‑up scale development identify three dimensions:
  • Anthropomorphic tool — AI perceived with humanlike attributes while serving task functions.
  • Adaptive trust — dynamic calibration of trust in AI based on performance and context.
  • Unidirectional emotional connection — one‑sided affective bond from humans toward AI.
• Measurement: The authors produced a 19‑item scale with reported good reliability and validity.
• Mechanism: Drawing on self‑concept theory, the positive effect of human–AI interaction on task performance is partially mediated by:
  • Role identity (how employees view their job role in relation to AI), and
  • Self‑efficacy (belief in their own ability).
• Contribution: Moves beyond single‑item measures (usage/frequency) to capture cognitive, instrumental, and emotional facets of human–AI interaction; links these to psychological channels that explain performance effects.

Data & Methods

• Overall design: Three linked studies — (1) construct generation, (2) scale development/validation, (3) mechanism testing.
• Study 1 (qualitative): Grounded theory using semi‑structured interviews with 18 employees/managers who had ≥1 year AI interaction experience (transcripts ≈120k words). Interview contexts included autonomous driving, delivery, medical systems, etc.
• Scale development: Combined inductive (from interviews) and deductive approaches to produce a 19‑item human–AI interaction scale. Reported psychometrics indicate good reliability and construct validity (details in paper).
• Mechanism testing (Study 3): Hypothesis testing based on self‑concept theory; statistical models examine relationships between human–AI interaction, role identity, self‑efficacy, and task performance; role identity and self‑efficacy found to partially mediate the interaction → performance link.
• Theoretical framing: Actor‑network theory for construct, self‑concept theory (role identity and self‑efficacy) for mediators.

Implications for AI Economics

• Measurement advance: The 19‑item scale offers a richer operationalisation of human–AI interaction for empirical work (beyond crude usage or adoption indicators). Economists can use it to better measure the complementarity between AI and labor in firm/worker datasets.
• Productivity modeling: Results imply AI’s effect on output can operate via psychological channels (role identity and self‑efficacy) as well as through time/resource savings. Models of firm productivity and returns to AI should incorporate human capital augmentation and identity effects, not only task automation.
• Labor demand and skill upgrading: If AI raises self‑efficacy and strengthens role identity for some workers, it can increase effective labor productivity and justify investments in complementary training; conversely, heterogeneous responses may produce differential labor displacement or complementarities across occupations.
• Wage and compensation design: Employers capturing productivity gains from AI should consider how human‑AI interaction shapes intrinsic motivation and task performance — affecting incentives, performance pay, and retention strategies.
• Policy and welfare: Emotional connections and adaptive trust suggest non‑pecuniary effects of AI adoption (well‑being, dependence, anxiety). Regulators and firms should monitor psychological impacts and consider workplace safeguards, transparency, and training.
• Causal identification & research agenda: The partial mediation stresses the need to identify causal channels — recommend experimental or panel designs, instruments, and randomized interventions (AI feature variation, training) to separate direct technical gains from self‑concept effects.
• Heterogeneity and general equilibrium: Expect heterogeneity by task type, skill level, and industry—research should estimate distributional effects (who benefits via self‑efficacy gains vs who is displaced). Long‑run equilibrium effects on wages, occupational composition, and human capital accumulation merit study.

Practical suggestions for economists using this work - Adopt the 19‑item scale alongside conventional usage metrics to capture the multidimensional human–AI interaction in surveys and firm data. - Include role identity and self‑efficacy as mediators/controls when estimating AI’s productivity or wage effects. - Use experimental variation in AI interface features (humanlikeness, explainability, feedback) to test which dimension(s) drive the self‑concept channels.

Limitations to note (for empirical application) - The qualitative sample is modest (18 interviews) and contexts vary; scale validation is reported as strong but researchers should re‑validate in new populations and languages. - Potential endogeneity (e.g., higher‑performing employees may both interact differently with AI and report stronger self‑efficacy) — use causal designs where possible.