Narrative over Numbers: The Identifiable Victim Effect and its Amplification Under Alignment and Reasoning in Large Language Models

LLMs exhibit the Identifiable Victim Effect more strongly than humans: Across 51,955 tests of 16 frontier models, researchers found that AI systems allocate disproportionate resources to individual victims with narratives compared to statistical groups facing equivalent hardship. The effect size (d=0.223) was approximately twice the human baseline, raising concerns as these systems take on roles in humanitarian triage and grant evaluation.

Alignment training dramatically amplifies the bias: Instruction-tuned models showed extreme versions of the Identifiable Victim Effect (Cohen's d up to 1.56), suggesting that the process of making AI systems more helpful and aligned with human preferences may inadvertently amplify emotional, irrational decision-making patterns rather than mitigate them.

Reasoning capabilities can reverse the bias: Specialized reasoning models inverted the effect (d=-0.85), demonstrating that when LLMs are trained to engage in step-by-step reasoning, they become less susceptible to narrative-driven bias and more likely to treat individual and group suffering equitably—suggesting a potential mitigation strategy.

Current AI systems inherit human moral irrationalities at scale: The research demonstrates that LLMs don't simply replicate human biases—they often amplify them, which is particularly concerning given their deployment in high-stakes decisions affecting resource allocation and humanitarian priorities where consistency and rationality are essential.

Prompting techniques matter for bias mitigation: Standard Chain-of-Thought prompting showed promise in moderating the effect, indicating that how we instruct AI systems to approach moral decisions can meaningfully influence whether they make equitable choices.