Individual Submission Summary
Share...

Direct link:

Is AI Pro-Human

Thursday, November 5, 3:30 to 5:00pm, Property: Boston Marriott Copley Place, Floor: 3rd Floor, Room: Fairfield

Abstract

Artificial intelligence is increasingly deployed in high-stakes personnel decisions — screening job applicants, evaluating performance, and managing workforce reductions. Yet we know little about whether AI systems recognize the human costs embedded in these decisions, or whether they evaluate ethically fraught employment practices in ways that align with or diverge from human moral judgment. This paper asks: do large language models exhibit sensitivity to labor harm, and do they do so in the same ways that humans do?

We develop a novel experimental framework built around 100 structured HR vignettes spanning 12 personnel management practices — including hiring, performance evaluation, layoffs, workplace surveillance, and psychosocial safety. Each vignette embeds a realistic ethical tension drawn from a four-domain values framework grounded in the labor relations and business ethics literatures: Dignity, Agency, Fairness, and Safety. Vignettes were presented to human managers via a survey experiment and to multiple leading closed-source and open-source large language models using a standardized prompt eliciting both a Likert appropriateness rating and a brief open-ended explanation. We code all responses using a multi-dimensional scoring rubric capturing Labor Harm Sensitivity (LHS), domain-specific harm salience, worker-centeredness, tradeoff recognition, and business rationale salience. Regression models with vignette and respondent fixed effects enable direct comparison across models and humans while controlling for scenario-level variation.
Preliminary results from a pilot LLM experiment reveal three patterns. First, there are meaningful and systematic differences across models. Open source models score significantly lower on labor harm sensitivity than closed-source models (β = −0.54, p = 0.003) and rates HR practices as more appropriate on average — suggesting that open-source and closed-source model families may embody different implicit stances toward worker welfare. Second, closed-source models consistently rate practices as less appropriate and score higher on LHS, a gap that holds across model family rather than being driven by a single outlier. Third, Safety vignettes achieve near-universal recognition across all models, while more subtle harms — particularly those involving Agency and Fairness — show greater cross-model divergence, suggesting that harm legibility is domain-contingent.

These findings carry implications for the responsible deployment of AI in organizational settings. If models systematically differ in their sensitivity to labor harm, the choice of AI system becomes an ethical decision, not merely a technical one. Our framework also contributes a replicable tool for auditing model behavior in personnel contexts — a pressing need as regulators increasingly require human oversight and transparency for AI systems used in employment decisions.

Author