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

arXiv: https://arxiv.org/abs/2606.12426 · HTML: https://mlsanity.com/paper/2606.12426/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 24 of its 26 references that are in the map.
- First posted on arXiv 12 May 2026.

## Builds on (most cited first, 24 of 24)

- [Chain-of-Thought Prompting Elicits Reasoning in Large Language Models](https://mlsanity.com/paper/2201.11903/) (2022)
- [Qwen2.5 Technical Report](https://mlsanity.com/paper/2412.15115/) (2024)
- [Mistral 7B](https://mlsanity.com/paper/2310.06825/) (2023)
- [A Survey on LLM-as-a-Judge](https://mlsanity.com/paper/2411.15594/) (2024)
- [ChatGPT Outperforms Crowd-Workers for Text-Annotation Tasks](https://mlsanity.com/paper/2303.15056/) (2023)
- [Whose Opinions Do Language Models Reflect?](https://mlsanity.com/paper/2303.17548/) (2023)
- [Out of One, Many: Using Language Models to Simulate Human Samples](https://mlsanity.com/paper/2209.06899/) (2022)
- [Zephyr: Direct Distillation of LM Alignment](https://mlsanity.com/paper/2310.16944/) (2023)
- [TweetEval: Unified Benchmark and Comparative Evaluation for Tweet Classification](https://mlsanity.com/paper/2010.12421/) (2020)
- [Dealing with Disagreements: Looking Beyond the Majority Vote in Subjective Annotations](https://mlsanity.com/paper/2110.05719/) (2021)
- [Measurement and Fairness](https://mlsanity.com/paper/1912.05511/) (2019)
- [Annotators with Attitudes: How Annotator Beliefs And Identities Bias Toxic Language Detection](https://mlsanity.com/paper/2111.07997/) (2021)
- [Can Large Language Models Transform Computational Social Science?](https://mlsanity.com/paper/2305.03514/) (2023)
- [Do LLMs exhibit human-like response biases? A case study in survey design](https://mlsanity.com/paper/2311.04076/) (2023)
- [AnnoLLM: Making Large Language Models to Be Better Crowdsourced Annotators](https://mlsanity.com/paper/2303.16854/) (2023)
- [The 'Problem' of Human Label Variation: On Ground Truth in Data, Modeling and Evaluation](https://mlsanity.com/paper/2211.02570/) (2022)
- [Measuring Political Bias in Large Language Models: What Is Said and How It Is Said](https://mlsanity.com/paper/2403.18932/) (2024)
- [Automated Annotation with Generative AI Requires Validation](https://mlsanity.com/paper/2306.00176/) (2023)
- [Specializing Large Language Models to Simulate Survey Response Distributions for Global Populations](https://mlsanity.com/paper/2502.07068/) (2025)
- [Human and LLM Biases in Hate Speech Annotations: A Socio-Demographic Analysis of Annotators and Targets](https://mlsanity.com/paper/2410.07991/) (2024)
- [Investigating Annotator Bias in Large Language Models for Hate Speech Detection](https://mlsanity.com/paper/2406.11109/) (2024)
- [On the Inevitability of Left-Leaning Political Bias in Aligned Language Models](https://mlsanity.com/paper/2507.15328/) (2025)
- [PromptPort: A Reliability Layer for Cross-Model Structured Extraction](https://mlsanity.com/paper/2601.06151/) (2026)
- [Not All Queries Need Rewriting: When Prompt-Only LLM Refinement Helps and Hurts Dense Retrieval](https://mlsanity.com/paper/2603.13301/) (2026)

## 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.