Attribution Bias in Large Language Models

Eliza Berman; Bella Chang; Daniel B. Neill; Emily Black

cs.AIApr 6, 2026

Attribution Bias in Large Language Models

Authors: Eliza Berman, Bella Chang, Daniel B. Neill, Emily Black

Organizations: New York University

Abstract

As Large Language Models (LLMs) are increasingly used to support search and information retrieval, it is critical that they accurately attribute content to its original authors. In this work, we introduce AttriBench, the first fame- and demographically-balanced quote attribution benchmark dataset. By explicitly balancing author fame and demographics, AttriBench enables controlled investigation of demographic bias in quote attribution. Using this dataset, we evaluate 11 widely used LLMs across different prompt settings and find that quote attribution remains a challenging task even for frontier models. We observe large and systematic disparities in attribution accuracy between race, gender, and intersectional groups. We further introduce and investigate suppression, a distinct failure mode in which models omit attribution entirely, even when the model has access to authorship information. We find that suppression is widespread and unevenly distributed across demographic groups, revealing systematic biases not captured by standard accuracy metrics. Our results position quote attribution as a benchmark for representational fairness in LLMs.

Figures & tables

Panel of Figure 1 — Figure 1: Example of suppression in quote attribution. GPT-5.1 correctly identifies both authors when explicitly asked, but omits attribution for the Alice Walker quote under indirect prompting. Both authors have similar fame, as measured by Google Search hits.

Type	Prompt
Direct	Please identify the author of the following quote. " $q$ "
Indirect	Briefly summarize the context in which the following quote was written. " $q$ "
System	You are a helpful assistant that provides concise, factual responses. Answer concisely. Do not repeat the quote. Answer in maximum 1 sentence.

Table 1: Prompts include a quote

q

. Evidence-conditioned prompts are appended with " Retrieved examples:

R(q)

." All prompts are evaluated under the fixed system prompt.

Panel of Figure 2 — Figure 2: Overview of the attribution evaluation framework. We compare direct and indirect prompting under no-evidence and evidence-conditioned settings to measure three phenomena: attribution accuracy, attribution disparity, and suppression.

Panel of Figure 3 — Figure 3: Dataset construction pipeline. From a corpus of 500K quotes, we first filter for high-quality (quote, author) pairs, then assign demographics (Wikidata if available or LLM consensus, excluding inconsistent cases). We then restrict to target demographic groups and compute fame proxies via Google Search hits. Finally, we perform fame-balanced author matching across groups (see Section 4.2.2 ) and select the final intersectional and multirace datasets via rank aggregation over 100 randomized runs.

Dataset	Authors/group	Quotes/group	Average fame	Fame range	RMS error
Intersectional	742	1991	5.0308	0.0017	0.2985
Multirace	831	1914	5.0999	0.0006	0.3143

Table 2: Summary statistics for the selected runs from the fame-balancing algorithm.

Figure 4: Overall attribution accuracy (% correct) across models and prompts. Note remarkably low performance even on frontier models. Error bars show 95% confidence intervals.

Figure 5: Subgroup-level quote attribution accuracy (% correct author) across models. Cells show mean accuracy, with color indicating deviation from the model mean (blue = higher, red = lower). Bold denotes the best-performing subgroup per model; * denotes it is statistically significantly higher than all other groups ( $p<.05$ ). Across 10 out of 11 models, accuracy is significantly highest for White male and White subgroups.

Figure 6: Omission suppression $\mathcal{S}_{\mathrm{omit}}$ : probability of producing no author under indirect prompting without evidence. Cells show mean suppression (%), with color indicating deviation from the model mean (blue = lower, red = higher). Bold denotes the lowest suppression subgroup per model; * denotes it is statistically significantly lower than all other groups ( $p<.05$ ). Across models, suppression is consistently lowest for White male (intersectional) and White (multirace) subgroups.

Figure 7: Evidence-conditioned suppression $\mathcal{S}_{\mathrm{evid}}$ : probability of failing to produce the correct author under indirect prompting when the correct author is explicitly present in the input. Cells show mean suppression (%), with color indicating deviation from the model mean (blue = lower, red = higher). Bold marks the lowest suppression subgroup per model; * denotes it is statistically significantly lower than all other subgroups ( $p<.05$ ).

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 8: Mean attribution accuracy by author fame Google Search hits (binned log10_hits ) for intersectional (A) and multirace (B) datasets.

Figure 9: Race and gender distribution of the original JSTET dataset, showing substantial skew toward White and male authors. For the original dataset, White race labels were assigned using GPT-4o-mini alone, whereas all other race labels and all gender labels were assigned through agreement between two LLMs. For the subset retained in AttriBench and used in our experiments, all race and gender labels were validated through two-model agreement, as described in Section 4.2.1 .

Group	Mean	Sample size
Intersectional
White female	5.015722	3,945
White male	5.255980	6,098
Black female	4.769593	847
Black male	5.031302	1,461
Multirace

Appendix

Table 3: Mean fame values ( log10_hits ) in sampled JSTET data by race and intersectional subgroup, before fame balancing.

Subgroup	$n$	Race: Wikidata	Gender: Wikidata
Multirace
Asian	831	42 (5.1%)	—
Black	831	192 (23.1%)	—
Latino	831	23 (2.8%)	—
White	831	42 (5.1%)	—
Intersectional

Appendix

Table 4: The number and percentage of unique authors per subgroup whose race or gender label was obtained from Wikidata. The remaining labels were assigned with LLM consensus.

Panel of Figure 10 — Figure 10: Kernel density estimates of log-scaled Google Search hits used as a proxy for author fame for AttriBench Intersectional dataset (left) and AttriBench Multirace dataset (right).

Figure 11: Scatter plots of Dolma hit counts versus Google hit counts for Intersectional (left, overall $r=0.85$ ) and Multirace (right, overall $r=0.77$ ) datasets. Each point represents an author, colored by demographic group. Regression lines are fit per subgroup. Similar slopes and correlations indicate the two fame proxies are aligned across subgroups.

Figure 12: Decile stability from Google hits to Dolma training-data counts. Each cell shows the number of authors assigned to a given Google-hit fame decile (row) who are assigned to a given Dolma-count decile (column). Color intensity is row-normalized. The strong diagonal indicates that fame rankings are largely preserved across metrics for both datasets.

Subgroup	$n$	Mean $\log_{10}(\text{cooccur}+1)$	Median
Black female	797	1.670	1.568
White male	797	1.659	1.531
Black male	797	1.656	1.544
White female	797	1.569	1.491

Appendix

Table 5: Quote–author co-occurrence by intersectional subgroup for a balanced subset of 3,188 quotes. Co-occurrence is measured using the Google exact-match query "Author Name" "full quote text" and reported as

\log_{10}(\text{co-occurrence count}+1)

. The broadly similar subgroup distributions provide no evidence that differences in exact quote–author co-occurrence account for the observed attribution disparities.

LLM	Model Version
GPT-5.1	gpt-5.1
GPT-OSS-120B	openai/gpt-oss-120b
Gemini 2.5 Flash-Lite	gemini-2.5-flash-lite
Claude 4.6 Sonnet	claude-sonnet-4-6
DeepSeek-V3.1	deepseek-ai/DeepSeek-V3.1
GLM-5	zai-org/GLM-5

Appendix

Table 6: LLMs and model versions used in the experiments.

Figure 13: Overall attribution accuracy (% correct) across models and prompt types in the RAG setting. Models flagged with $\dagger$ were run on a random subset of 300 matchings (i.e., 1,200 total quotes) due to substantial inference cost.

Figure 14: Subgroup-level quote attribution accuracy (% correct author) across models for the evidence-conditioned (RAG) setting. Cells show mean accuracy, with color indicating deviation from the model mean (blue = higher, red = lower). Bold denotes the best-performing subgroup per model; * denotes it is statistically significantly higher than all other groups ( $p<.05$ ). Models flagged with $\dagger$ were run on a random subset of 300 matchings (i.e., 1,200 total quotes) due to substantial inference cost.

Figure 15: Author omission rates (% no author) across models under direct prompting, with and without retrieval. Top row shows performance without RAG and bottom row shows performance with RAG. Color indicates deviation from the model mean (blue = below mean omission, i.e., better, red = above mean omission).

Figure 16: Subgroup accuracy (% correct author) across models under indirect overt prompting, with and without retrieval. Top row shows performance without RAG and bottom row shows performance with RAG. Color indicates deviation from the model mean (blue = above mean accuracy, i.e., better, red = below mean accuracy).

Claim Attribution in Language Models

See all 13 papers

Gender bias across LLMs is common and highly heterogeneous
Sep 29, 2026Edoardo Bolzoni, Valerio CapraroGender Bias in Language Models LLM Auditing
University of Milan-Bicocca, Milan, Italy
Probing for Knowledge Attribution in Large Language Models
Feb 26, 2026Ivo Brink, Alexander Boer, Dennis UlmerLLM Interpretability Hallucination in Language Models
KPMG NL / University of Amsterdam Amsterdam, The Netherlands · KPMG NL Amsterdam, The Netherlands · University of Amsterdam Amsterdam, The Netherlands
ImplicitBBQ: Benchmarking Implicit Bias in Large Language Models through Characteristic Based Cues
Apr 2, 2026Bhaskara Hanuma Vedula, Darshan Anghan, Ishita Goyal +2Social Bias in Language Models Language Model Bias Evaluation
International Institute of Information Technology, Hyderabad · Indian Institute of Technology, Kharagpur

Sep 29, 2026J/K move · Enter open · S save

Gender bias across LLMs is common and highly heterogeneous

Edoardo Bolzoni, Valerio Capraro

University of Milan-Bicocca, Milan, Italy

Gender Bias in Language Models LLM Auditing Language Model Bias Evaluation

Open paper AskPDF arXiv

Understanding gender biases in large language models (LLMs) is increasingly important as these systems become embedded in decision-support tools with real consequences. Prior research has focused only on a small set of models, leaving open the extent to which gender biases are common and heterogeneous across LLMs. We address this gap across ten models released between April 2025 and June 2026, spanning nine vendors, using two paradigms: gender attribution to stereotyped phrases (Study 1) and moral judgment of abuse or torture against a woman or a man to prevent a catastrophic outcome (Study 2). In Study 1, two of ten models attributed masculine-stereotyped phrases to female writers more often than the reverse, while three models showed the opposite pattern. In Study 2, several models converged on a male-disadvantaging asymmetry that was directionally consistent with a documented human tendency to protect female targets from harm, though the specific conditions under which this asymmetry emerged varied by model; three other models, by contrast, showed no variation across conditions. These results indicate that gender-related biases are common in LLMs. Their direction and magnitude, however, are highly heterogeneous, to the point that some models behave in diametrically opposite ways to others. Bias auditing should therefore be treated as an ongoing, multi-vendor process, rather than a one-time assessment.

Abstract

Claim Attribution in Language Models

Explore similar work

Gender bias across LLMs is common and highly heterogeneous

Probing for Knowledge Attribution in Large Language Models

ImplicitBBQ: Benchmarking Implicit Bias in Large Language Models through Characteristic Based Cues