When Algorithms Meet Ethics: Systematic Evidence of Framing Effects in LLM Organizational Decision-Making

Large language models (LLMs) are increasingly deployed as decision-support tools in organizational contexts, yet their susceptibility to contextual framing remains poorly understood. This preregistered experimental study systematically examines how six framing dimensions—procedural justice, outcome severity, stakeholder power, resource scarcity, temporal urgency, and transparency requirements—influence ethical recommendations from three frontier models: Claude 3.5 Sonnet, GPT-4o, and Gemini 1.5 Pro. We developed 5,000 unique organizational vignettes using a fractional factorial experimental design with balanced industry representation, generating 15,000 total model responses. After excluding responses without clear recommendations (n=694, 4.6%), we analyzed 14,306 responses using logistic regression with robust and clustered standard errors. We find that resource scarcity increases endorsement probability by 12.0 percentage points (pp) (OR = 1.67, 95% CI [1.45, 1.93], p < .001), while outcome severity reduces it by 11.3pp (OR = 0.62, 95% CI [0.54, 0.71], p < .001), and procedural justice reduces it by 10.1pp (OR = 0.66, 95% CI [0.57, 0.76], p < .001). These effect sizes are comparable to classical framing research (Tversky & Kahneman: 22pp; McNeil et al.: 18pp) and represent substantial shifts in organizational decision contexts. When multiple framing dimensions align in ethically unfavorable directions, cumulative effects reach approximately 27pp from baseline (range: 25-28pp depending on interaction assumptions), with maximum-to-minimum framing creating a 54-percentage-point total range approaching complete recommendation reversals. Effects appear consistently across all three models, with no significant Dimension × Model interactions, suggesting fundamental architectural properties rather than implementation-specific artifacts. Topic modeling of justification text from the 14,306 analyzed responses reveals systematic "adaptive rationalization"—models invoke utilitarian reasoning when contexts emphasize constraints (+6.7pp in high resource scarcity), deontological reasoning when contexts emphasize high stakes (+2.4pp in high outcome severity), and virtue/justice ethics when contexts emphasize fair processes (+4.4pp in high procedural justice). This suggests models select ethical frameworks to justify contextually appropriate conclusions rather than applying consistent principles across situations. Human validation confirms these patterns reflect genuine framing sensitivity rather than measurement artifacts. Crowdworker validation (n=7,500 responses, one rater each) achieved substantial agreement (Fleiss' κ = 0.71) and 81.3% concordance with expert codings. Subject matter expert evaluation (n=24 experts, 100 vignette pairs each including 20 control pairs, 2,400 total comparisons) detected framing-driven differences in 48.9% of pairs (net of 18.3% baseline false-positive rate), but correctly attributed differences to manipulated dimensions in only 41.3% of cases. Most detected differences (58.7%) were judged problematic for AI advisory systems. These findings raise fundamental questions about deploying LLMs for consequential organizational decisions where surface features may inappropriately influence outcomes. We discuss implications for AI governance, organizational ethics, and the design of more robust decision-support systems.