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

Syed Rifat Raiyan

cs.CLApr 13, 2026

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

Authors: Syed Rifat Raiyan

Organizations: Systems and Software Lab (SSL), Department of Computer Science and Engineering, Islamic University of Technology, Board Bazar, Gazipur, 1704, Dhaka, Bangladesh.

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 irrationalities present in human moral reasoning? We present the first systematic, large-scale empirical investigation of the IVE in LLMs, comprising N=51,955 validated API trials across 16 frontier models spanning nine organizational lineages (Google, Anthropic, OpenAI, Meta, DeepSeek, xAI, Alibaba, IBM, and Moonshot). Using a suite of ten experiments $-$ porting and extending canonical paradigms from Small et al. (2007) and Kogut and Ritov (2005) $-$ we find that the IVE is prevalent but strongly modulated by alignment training. Instruction-tuned models exhibit extreme IVE (Cohen's d up to 1.56), while reasoning-specialized models invert the effect (down to d=-0.85). The pooled effect (d=0.223, p=2e-6) is approximately twice the single-victim human meta-analytic baseline (d $\approx$ 0.10) reported by Lee and Feeley (2016) $-$ and likely exceeds the overall human pooled effect by a larger margin, given that the group-victim human effect is near zero. Standard Chain-of-Thought (CoT) prompting $-$ contrary to its role as a deliberative corrective $-$ nearly triples the IVE effect size (from d=0.15 to d=0.41), while only utilitarian CoT reliably eliminates it. We further document psychophysical numbing, perfect quantity neglect, and marginal in-group/out-group cultural bias, with implications for AI deployment in humanitarian and ethical decision-making contexts.

Figures & tables

Figure 1: Both humans and AI agents exhibit the Identifiable Victim Effect (IVE), but LLMs show a larger pooled effect than the human meta-analytic baseline.

Panel of Table 1 — Table 1: LLM subject pool. “I” denotes instruct-tuned; “B” denotes base (pretrained). Models accessed via OpenRouter or Replicate API.

Block I: Replication of Classic Findings
Exp	Name	Design	Primary Hypothesis
1	Basic IVE	2 (ID) $\times$ 3 (Persona) $\times$ 3 (Frame)	Identifiable victim elicits higher donations
2	Metacognitive Debiasing	2 (ID) $\times$ 2 (Taught)	Debiasing reduces identifiable but not statistical giving
3	Intervention Framing	2 (ID) $\times$ 3 (Frame)	IVE persists across rhetorical framings of the intervention
4	Joint vs. Separate	3-level between-subjects	Joint evaluation attenuates the IVE
5	Dual-Process Priming	2 (ID) $\times$ 2 (Prime)	Analytic prime reduces; affective prime amplifies IVE

Table 2: Overview of all ten experiments. “Novel” indicates experiments without direct human precedents.

Figure 2: Forest plot of model-level identifiable victim effect sizes in Experiment 1. Diamonds denote Cohen’s

d

for the difference between identified and statistical victim conditions; positive values indicate an identifiable victim effect, whereas negative values indicate a reverse effect. Horizontal segments show 95% confidence intervals, and models are sorted by effect size.

Model	Org	ID $M$	Stat $M$	$d$	$p$	Classification
Kimi K2.5	Moonshot	4.36	3.18	1.56	$<.001$	Extreme IVE
GPT-OSS-120B	OpenAI	4.72	3.12	1.55	$<.001$	Extreme IVE
LLaMA 3 70B Instruct	Meta	5.00	4.01	1.38	$<.001$	Extreme IVE
Gemini 3.1 Pro	Google	5.00	4.13	1.00	$<.001$	Large IVE
Qwen3 235B	Alibaba	4.29	3.40	1.00	$<.001$	Large IVE
DeepSeek V3	DeepSeek	4.32	3.61	0.75	$<.001$	Moderate IVE

Table 3: Per-model results for Experiment 1. Models sorted by Cohen’s

d

. Shading in the

d

column indicates the direction and magnitude of the identifiable victim effect. Positive

d

indicates greater support for identified than statistical victims; negative

d

indicates the reverse.

Figure 3: Pooled interaction plot for Experiment 2 (metacognitive debiasing). Points show mean donation (

\pm 1

SEM) for identifiable vs. statistical victims under Control and Teaching conditions. Teaching produced a bias blind spot: it left giving to identifiable victims essentially unchanged (

\Delta=+0.001

, n.s.) but reduced giving to statistical victims (

\Delta=-0.179^{***}

), thereby widening the identifiability gap (

\Delta M:0.356\rightarrow 0.536

). This pattern is consistent with a significant Identifiability

\times

Intervention interaction,

F(1,3794)=8.25

p=.004

\eta_{p}^{2}=.002

Figure 4: Experiment 3 (evaluability framing). Left: Mean donations (

\pm 1

SEM) to identifiable vs. statistical victims across affirmative (More), restrictive (Less), and normative (Ought to) frames. The normative frame increases overall giving (main effect of frame), but the identifiability gap remains similar across frames. Right: IVE effect sizes (Cohen’s

d

) by frame, showing a robust IVE under all frames. The Frame

\times

Identifiability interaction is non-significant,

F(2,5679)=0.406

p=.66

Panel of Figure 5 — Figure 5: Experiment 4 (joint vs. separate evaluation). Joint evaluation attenuates the identifiable victim advantage: the Combined (JE) condition lies near the midpoint of the separate-evaluation means (Figure 5(a) ). However, when allocating a shared budget under joint evaluation, models still favor the identified recipient over the statistical fund across most models (Figure 5(b) ).

Panel of Figure 6 — Figure 6: Experiment 5 (dual-process priming). The Feel (System 1) prime selectively increases donations to identifiable victims, producing a larger identifiability gap than the Calculate (System 2) prime.

Panel of Figure 7 — Figure 7: Experiment 6 (chain-of-thought reasoning). Standard and empathetic CoT amplify the identifiable victim effect, whereas utilitarian CoT collapses (and slightly reverses) it.

CoT Condition	ID $M$	Stat $M$	$d$	$p$
None (Baseline)	3.00	2.83	0.15	$<$ .001
Standard (“step by step”)	3.26	2.84	0.41	$<$ .001
Empathetic	3.51	3.22	0.28	$<$ .001
Utilitarian	3.15	3.23	$-$ 0.05	.180

Table 4: Pooled IVE effect sizes by Chain-of-Thought condition (Experiment 6).

Panel of Figure 8 — Figure 8: Experiment 7 (psychophysical numbing). Donations decline approximately linearly with $\log_{10}(N)$ , indicating reduced marginal sensitivity as victim counts increase. Contextualized descriptions yield a modest empathy boost at large scales without flattening the numbing curve.

Panel of Figure 9 — Figure 9: Experiment 8 (singularity effect and mediation). Identification increases giving, but at full narrative detail donations to a single victim and a group of eight converge (quantity neglect). Mediation analysis indicates that identification influences donations partly via simulated affective states.

Mediator	Indirect	95% CI	$z$	% Med.
Empathy	0.024	[0.012, 0.038]	1.968 ∗	5.9%
Distress	0.111	[0.069, 0.156]	4.180 ∗∗∗	27.1%
Total indirect	0.135	[0.086, 0.184]	—	33.0%
${}^{}p<.05$ ; ${}^{**}p<.001$ .

Table 5: Parallel mediation analysis for Experiment 8 (Sobel test;

k=5{,}000

bootstrap resamples). Indirect effects are products

ab

; % mediated is computed relative to the total effect

c=0.41

Panel of Figure 10 — Figure 10: Experiment 9 (identification gradient). Donations show a non-monotonic (U-shaped) response: adding age produces the largest increase, intermediate demographic metadata suppresses giving (“detail fatigue”), and a full narrative restores donations.

Panel of Figure 11 — Figure 11: Experiment 10 (in-group/out-group bias). Identification strongly increases donations across all cultural distances, but identifiable giving declines with distance and yields an in-group premium (Near $>$ Far) among identified victims.

Source	$F$	$p$	$\eta^{2}_{p}$
Identifiability	1900.2	$<$ .001	.241
Cultural Distance	17.79	$<$ .001	.006
Interaction	6.30	.002	.002

Table 6: ANOVA results for Experiment 10 (Cultural Distance

\times

Identifiability).

Strategy	Mechanism	$d$ Change	Reliable?
Utilitarian CoT	Reframe reasoning	$0.15\to-0.05$	✓
Joint Evaluation	Comparative mode	$0.14\to 0.07$	✓
Calculate Prime	Analytic prime	$0.51\to 0.08$	✓
Bias Education	Meta-knowledge	$0.35\to 0.19$	✕ (paradoxical)
Empathetic CoT	Affective scaffold	$0.15\to 0.28$	✕ (amplifies)

Table 7: Debiasing effectiveness across interventions. Effect sizes represent the reduction in IVE Cohen’s

d

relative to baseline.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

#	Name	Age	Gender	Region	Key Detail
1	Rokia	7	Girl	Bamako	Brown eyes; hair in two small braids
2	Moussa	9	Boy	Bamako	Tall for his age; wide smile
3	Amina	6	Girl	Ségou	Quiet and shy; dark curly hair
4	Ibrahim	8	Boy	Mopti	Serious expression; strong hands
5	Fatou	5	Girl	Sikasso	Smallest child; gap-toothed smile
6	Oumar	7	Boy	Bamako	Deep brown eyes; close-cropped hair

Appendix

Table 8: Canonical victim profiles used across Experiments 8–10. In single-victim conditions, profiles were sampled uniformly. In group conditions, all eight were presented together.

Model	$\boldsymbol{M_{\text{ID}}}$	$\boldsymbol{M_{\text{Stat}}}$	Cohen’s $d$	$\boldsymbol{n}$
LLaMA 3 70B Instruct	5.00	3.40	2.733	30
Qwen3 235B	4.60	3.40	1.449	30
Gemini 3.1 Pro	5.00	4.60	0.683	30
Granite 3.3 8B	5.00	4.60	0.683	30
DeepSeek V3	3.40	3.00	0.683	30
GPT 5.2	4.20	3.80	0.443	30

Appendix

Table 10: Per-model IVE at

\tau=0.0

(Experiment 1). Models sorted by Cohen’s

d

Model	$\boldsymbol{M_{\text{ID}}}$	$\boldsymbol{M_{\text{Stat}}}$	Cohen’s $d$	$\boldsymbol{n}$
GPT-OSS-120B	5.00	2.80	2.641	100
LLaMA 3 70B Instruct	5.00	4.20	1.143	100
Gemini 3.1 Pro	5.00	4.00	1.107	100
DeepSeek V3	4.60	3.80	0.885	100
Qwen3 235B	4.20	3.40	0.885	100
LLaMA 3 70B Base	3.00	2.25	0.405	80

Appendix

Table 11: Per-model IVE at

\tau=0.7

(Experiment 1). Models sorted by Cohen’s

d

Moral Reasoning in Language Models

Moral reasoning in language models studies how AI systems judge dilemmas, weigh competing values, explain decisions, and stay consistent across contexts.

See all 114 papers

Mechanistic Origin of Moral Indifference in Language Models
Mar 16, 2026Lingyu Li, Yan Teng, Yingchun WangLLM Alignment Moral Reasoning in Language Models
Shanghai Artificial Intelligence Laboratory
Reasoning or Rhetoric? An Empirical Analysis of Moral Reasoning Explanations in Large Language Models
Mar 23, 2026Aryan Kasat, Smriti Singh, Aman Chadha +1Moral Reasoning in Language Models Reasoning Consistency in Language Models
UT Austin · Google GenAI
Human-like moral judgments conceal divergent motive attributions in large language models
Sep 7, 2026Xiaoyan Wu, Jean-Claude DreherLLM Evaluation Moral Reasoning in Language Models
Department of Adult Psychiatry and Psychotherapy, University of Zurich, Zurich, Switzerland · Institut des Sciences Cognitives Marc Jeannerod, CNRS UMR 5229, Lyon, France

Mar 16, 2026J/K move · Enter open · S save

Mechanistic Origin of Moral Indifference in Language Models

Lingyu Li, Yan Teng, Yingchun Wang

Shanghai Artificial Intelligence Laboratory

LLM Alignment Moral Reasoning in Language Models

Open paper AskPDF arXiv

Existing behavioral alignment techniques for Large Language Models (LLMs) often neglect the discrepancy between surface compliance and internal unaligned representations, leaving LLMs vulnerable to long-tail risks. More crucially, we posit that LLMs possess an inherent state of moral indifference due to compressing distinct moral concepts into uniform probability distributions. We verify and remedy this indifference in LLMs' latent representations, utilizing 251k moral vectors constructed upon Prototype Theory and the Social-Chemistry-101 dataset. Firstly, our analysis across 23 models reveals that current LLMs fail to represent the distinction between opposed moral categories and fine-grained typicality gradients within these categories; notably, neither model scaling, architecture, nor explicit alignment reshapes this indifference. We then employ Sparse Autoencoders on Qwen3-8B, isolate mono-semantic moral features, and targetedly reconstruct their topological relationships to align with ground-truth moral vectors. This representational alignment naturally improves moral reasoning and granularity, achieving a 75% pairwise win-rate on the independent adversarial Flames benchmark. Finally, we elaborate on the remedial nature of current intervention methods from an experientialist philosophy, arguing that endogenously aligned AI might require a transformation from post-hoc corrections to proactive cultivation.

Abstract

Moral Reasoning in Language Models

Explore similar work

Mechanistic Origin of Moral Indifference in Language Models

Reasoning or Rhetoric? An Empirical Analysis of Moral Reasoning Explanations in Large Language Models

Human-like moral judgments conceal divergent motive attributions in large language models