Depending on how an AI is trained, it will either care way too much about one person or be cold-heartedly obsessed with the many.
April 16, 2026
Original Paper
Narrative over Numbers: The Identifiable Victim Effect and its Amplification Under Alignment and Reasoning in Large Language Models
arXiv · 2604.12076
The Takeaway
Humans have a bias called the 'Identifiable Victim Effect,' where we’d rather save one person with a name and a face than 100 anonymous people. This study found that AI models don't just mimic this—they magnify it to the extreme. Basic models are intensely biased toward the individual, but 'reasoning' models can actually flip the switch and become hyper-utilitarian, ignoring the individual entirely. This means the 'ethics' of an AI isn't a fixed thing; it’s a dial that can be turned to either extreme. It highlights how dangerous it is to assume an AI will make 'human' decisions when faced with a tragedy.
From the abstract
The Identifiable Victim Effect (IVE) $-$ the tendency to allocate greater resources to a specific, narratively described victim than to a statistically characterized group facing equivalent hardship $-$ is one of the most robust findings in moral psychology and behavioural economics. As large language models (LLMs) assume consequential roles in humanitarian triage, automated grant evaluation, and content moderation, a critical question arises: do these systems inherit the affective irrationaliti