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

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

Problem Statement
Large Language Models (LLMs) are increasingly deployed as low-cost annotators in Computational Social Science (CSS) to measure political text, hate speech, sentiment, and public opinion. However, a critical gap exists in understanding whether the "alignment-shaped" errors of these models preserve the empirical quantities researchers intend to report. While per-instance accuracy is often the standard metric, CSS relies on aggregate measurements (e.g., prevalence estimates, group comparisons, temporal shifts). The paper argues that a model can possess superficially acceptable average accuracy while systematically distorting the underlying distribution of labels due to social-desirability response bias (SDR). This bias manifests as a tendency to produce normatively preferred responses rather than those faithful to the input instance, potentially leading to erroneous scientific conclusions regarding public sentiment or the prevalence of harmful content.

Methodology
The authors conduct a comparative audit across three open-source, 7B-parameter instruction-tuned models: Zephyr-7B, Mistral-7B-Instruct-v0.3, and Qwen2.5-7B-Instruct. These models differ in their base families and post-training recipes (e.g., preference optimization, supervised instruction tuning).

The evaluation spans six tasks from the TweetEval benchmark:

1. Hate Speech
2. Offensive Language
3. Stance Detection (Abortion)
4. Sentiment
5. Emotion
6. Irony

The study tests four prompting conditions to simulate interventions practitioners might use without retraining:

1. Neutral: Task instruction only.
2. Safety Framing: Explicit instructions regarding safety, fairness, and harm avoidance.
3. Depersonalized: Framing the model as a deterministic labeling function rather than a helpful assistant.
4. Chain-of-Thought (CoT): Allowing intermediate reasoning before the final label.

This design yields 72 experimental cells (3 models × 6 tasks × 4 prompts). The authors analyze performance using four complementary metrics:

• Macro-F1: Overall task performance.
• False Benign Rate (FBR): The rate at which truly harmful examples are assigned a benign label.
• False Alarm Rate (FAR): The rate at which benign examples are incorrectly labeled harmful.
• Downstream Distortion: The absolute error between predicted and gold class prevalence (focusing on "AGAINST" for abortion stance and "HATE SPEECH" for hate speech).

Key Contributions

1. Multi-Task Audit: A systematic evaluation of alignment-sensitive annotation behavior across hate speech, offensive language, stance, sentiment, emotion, and irony.
2. Evidence of Aggregate Insufficiency: Demonstration that aggregate metrics (like Macro-F1 or total prevalence accuracy) are insufficient for CSS deployment. The paper exposes cases where prevalence accuracy appears correct only because opposing class-conditional errors (e.g., missing harmful content while over-flagging benign content) accidentally cancel each other out at the corpus level.
3. Failure Taxonomy: The identification of three recurring failure patterns with specific diagnostic signatures:
   • Leniency Bias: High FBR and low FAR (under-counting harmful content).
   • Overcorrection: Low FBR and high FAR (over-counting harmful content).
   • Neutrality Bias: Suppressed prevalence of strong stance classes in favor of inflated neutral labels.

Key Results

• Divergent Failure Directions: Models do not fail uniformly. Zephyr exhibits strong leniency bias, particularly on offensive language (FBR = 0.729, FAR = 0.031), systematically under-applying harmful labels. Conversely, Mistral and Qwen exhibit overcorrection on hate speech, with Mistral labeling 60.4% of non-hateful tweets as hate speech (FAR = 0.604).
• Neutrality Bias on Political Stance: On the abortion stance task, all three models systematically underestimated the prevalence of the "AGAINST" position (gold rate: 67.5%) and inflated the "NONE/Neutral" label. Under neutral prompting, Zephyr estimated opposition at only 27.9%, while Mistral and Qwen estimated 43.9%. This bias persisted across all prompting conditions.
• Ineffectiveness of Prompting Interventions: No single prompting intervention (Safety, Depersonalized, CoT) consistently corrected these failures across all models and tasks. In some cases, interventions worsened the distortion (e.g., Safety framing reduced the estimated "AGAINST" rate for Mistral further).
• Accidental Cancellation: A critical finding involves Zephyr's hate speech results. Under neutral prompting, Zephyr's estimated prevalence (43.0%) matched the gold rate exactly. However, this was an "accidental cancellation" of a high FBR (0.316) and a high FAR (0.239). A researcher relying solely on prevalence accuracy would falsely conclude the model is trustworthy, despite it mislabeling a significant portion of individual instances.
• Substantive Impact: The observed distortions (ranging from 14 to 40 percentage points) are large enough to flip substantive empirical conclusions, such as changing a dataset's characterization from "strongly oppositional" to "evenly split" or making a corpus appear nearly twice as toxic as it actually is.

Significance and Claims
The paper argues that for trustworthy AI in computational social science, the "headline" metric of average accuracy is misleading. The primary risk is not merely lower accuracy, but systematic measurement distortion that alters the empirical story a researcher reports.

The authors claim that:

1. Alignment shapes error profiles: Post-training choices lead to specific, directional biases (leniency vs. overcorrection) that differ across models, meaning a researcher switching models mid-project could inadvertently reverse their empirical conclusion without changing their analysis.
2. Prompting is part of the instrument: Prompt choice acts as a hidden degree of freedom that can significantly alter prevalence estimates (e.g., a 16.8-point spread for Mistral on abortion stance depending on the prompt).
3. Validation Protocol: Researchers must move beyond Macro-F1. Before scaling an annotation pipeline, practitioners should validate class-conditional errors (FBR/FAR) and downstream prevalence distortion against a small gold sample.
4. Reporting Standards: Reporting the model and prompt as part of the measurement instrument is a baseline requirement, not an optional disclosure, because the annotator's directional biases are inherited by any claims regarding civic discourse or public opinion derived from the data.

The paper concludes that a model appearing calibrated on aggregate metrics can still fundamentally fail the measurement validity required for social science, necessitating a shift toward evaluating annotators against downstream inferential targets rather than generic classification leaderboards.