mlsanityData to 7 Oct 2026

Two Wrongs, No Right: Auditing Social-Desirability Bias in LLM Annotators for Computational Social Science

arXiv 2606.12426 ↗First posted 12 May 2026cs.CY cs.CL cs.LGcited by 0

Not on the map. Why?

Abstract (arXiv)

LLM annotators are increasingly used in computational social science (CSS), but it is unclear whether their alignment-shaped errors preserve the empirical conclusions a researcher would report. We audit three open-source 7B instruction-tuned models (Zephyr, Mistral-Instruct, Qwen2.5-Instruct) across six TweetEval tasks under four prompt conditions (72 cells) and find that social-desirability failures do not run in a single direction. Zephyr exhibits leniency bias, systematically under-applying harmful labels (offensive language: false benign rate 0.729, false alarm rate 0.031). Mistral and Qwen exhibit overcorrection, over-applying the same labels (Mistral hate-speech FAR = 0.604). All three models exhibit neutrality bias on abortion stance, underestimating opposition prevalence by 24 to 40 percentage points and inflating the neutral label. None of the four prompting interventions we test (neutral, safety framing, depersonalized, chain-of-thought) corrects these failures across models; safety framing can worsen stance distortion. Strikingly, Zephyr's hate-speech prevalence estimate matches the gold rate exactly while its class-conditional errors are large in both directions, an accidental cancellation that misleads aggregate validation. We translate these patterns into a three-part taxonomy with diagnostic FBR/FAR signatures and a lightweight gold-sample validation protocol. The headline for trustworthy CSS: a model that looks calibrated on aggregate metrics can still flip the substantive empirical conclusion a researcher would report.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Builds on 24 of its 26 references that are in the map. First posted on arXiv 12 May 2026.

Builds on

24 of 26 references.

References · 24

  1. 2022Chain-of-Thought Prompting Elicits Reasoning in Large Language Modelscited by 12,302
  2. 2024Qwen2.5 Technical Reportcited by 4,713
  3. 2023Mistral 7Bcited by 3,065
  4. 2024A Survey on LLM-as-a-Judgecited by 1,384
  5. 2023ChatGPT Outperforms Crowd-Workers for Text-Annotation Taskscited by 849
  6. 2023Whose Opinions Do Language Models Reflect?cited by 759
  7. 2022Out of One, Many: Using Language Models to Simulate Human Samplescited by 737
  8. 2023Zephyr: Direct Distillation of LM Alignmentcited by 498
  9. 2020TweetEval: Unified Benchmark and Comparative Evaluation for Tweet Classificationcited by 414
  10. 2021Dealing with Disagreements: Looking Beyond the Majority Vote in Subjective Annotationscited by 306
  11. 2019Measurement and Fairnesscited by 299
  12. 2021Annotators with Attitudes: How Annotator Beliefs And Identities Bias Toxic Language Detectioncited by 258
12 more
  1. 2023Can Large Language Models Transform Computational Social Science?cited by 245
  2. 2023Do LLMs exhibit human-like response biases? A case study in survey designcited by 176
  3. 2023AnnoLLM: Making Large Language Models to Be Better Crowdsourced Annotatorscited by 171
  4. 2022The 'Problem' of Human Label Variation: On Ground Truth in Data, Modeling and Evaluationcited by 133
  5. 2024Measuring Political Bias in Large Language Models: What Is Said and How It Is Saidcited by 89
  6. 2023Automated Annotation with Generative AI Requires Validationcited by 75
  7. 2025Specializing Large Language Models to Simulate Survey Response Distributions for Global Populationscited by 50
  8. 2024Human and LLM Biases in Hate Speech Annotations: A Socio-Demographic Analysis of Annotators and Targetscited by 31
  9. 2024Investigating Annotator Bias in Large Language Models for Hate Speech Detectioncited by 6
  10. 2025On the Inevitability of Left-Leaning Political Bias in Aligned Language Modelscited by 3
  11. 2026PromptPort: A Reliability Layer for Cross-Model Structured Extractioncited by 3
  12. 2026Not All Queries Need Rewriting: When Prompt-Only LLM Refinement Helps and Hurts Dense Retrievalcited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026BiasTrace: Linking Reasoning Behaviours to Biased Outputs in LLMs
  2. 2025Do LLMs Align Human Values Regarding Social Biases? Judging and Explaining Social Biases with LLMs
  3. 2024With a Grain of SALT: Are LLMs Fair Across Social Dimensions?
  4. 2024A Multi-LLM Debiasing Framework
  5. 2026Edu-MMBias: A Three-Tier Multimodal Benchmark for Auditing Social Bias in Vision-Language Models under Educational Contexts
  6. 2025Confident, Calibrated, or Complicit: Safety Alignment and Ideological Bias in LLM Hate Speech Detection
  7. 2024BiasAlert: A Plug-and-play Tool for Social Bias Detection in LLMs
  8. 2026Overstating Attitudes, Ignoring Networks: LLM Biases in Simulating Misinformation Susceptibility