The Ghost Annotator: a Framework to Explore Human Label Variation in Content Moderation through Conformal Prediction
Authors: Mirko Lai, Alessandra Urbinati, Simona Frenda, Fabiana Vernero, Marco Antonio Stranisci
Organizations: Heriot-Watt University, Edinburgh, Scotland · aequa-tech, Torino, Italy · Laboratory for the Modeling of Biological and Socio-technical Systems, Northeastern University, Boston, MA, USA · Università del Piemonte Orientale, Vercelli, Italy · Università degli Studi di Torino, Torino, Italy
Current research primarily focuses on model performance, while comparatively less attention has been devoted to uncertainty estimation, particularly in settings where LLMs are increasingly used to generate annotated data. We introduce a framework combining conformal prediction with Collaborative Filtering-style annotators' representation to model LLM behavior in relation to human annotators and to analyze patterns of agreement and disagreement. Using Non-Conformity Scores, we introduce the Ghost Prediction metric and the Ghost Annotator representation to quantify cases in which model predictions diverge from all available human annotations. We compute cosine similarity measures to explore differences in model behavior across sociodemographic axes. We evaluated four LLMs of different size and families across four content moderation datasets. Our finding shows that while we find that all models uncertainty increases with annotator disagreement, larger models tend to be more confident in the classification of texts that are not aligned with any human annotation. Finally, the Ghost Annotator framework reveals a consistent and robust pattern of demographic misalignment, suggesting a structural bias likely rooted in pretraining corpora.
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.
Hate speech detection is inherently subjective: people from different demographic groups perceive the same content very differently. Collecting enough annotations from multiple demographic groups is costly and difficult to scale. Persona-conditioned Large Language Models (models prompted to adopt a specific demographic identity) have been proposed as a way to simulate diverse perspectives at scale. But do they actually reflect how different groups disagree? We evaluate three aspects of human social judgement: (i) whether personas from different groups disagree in human-like ways (inter-group disagreement), (ii) whether they become more sensitive when content targets their own identity (in-group sensitivity), and (iii) whether they can accurately predict how another group would react (vicarious prediction). Our results show that no model consistently captures all three dimensions, and performance is highly model-dependent and does not emerge reliably from minimal identity prompts alone. However, vicarious prompting with Llama 3.1 yields the highest cross-group agreement in most demographic axes and provides the closest overall approximation to human disagreement patterns, indicating that this configuration may provide a more reliable setting for automatic annotation aligned with human judgements.
Hate speech annotation is costly, subjective, and prone to annotator disagreement, making large-scale dataset construction challenging. We systematically analyze how well large language models (LLMs) align with human judgments across ten theoretically grounded subjective attributes, such as dehumanization, violence, and sentiment, evaluating both small and large variants of Llama 3.1 and Qwen 2.5. Our analysis reveals a consistent split across all models: behaviorally explicit dimensions (insult, humiliate, attack-defend) correlate strongly with human annotations, while evaluative dimensions (respect, sentiment, hate speech) are systematically inverted. Demographic persona conditioning reduces model confidence without improving alignment. Building on these insights, we propose combining attribute-level LLM predictions via a confidence-weighted Ridge regression to reconstruct continuous hate speech scores from the Measuring Hate Speech corpus, achieving R2 of up to 0.71 and outperforming direct prompting baselines, demonstrating that structured attribute decomposition recovers a richer and more human-aligned signal than end-to-end label prediction alone.
Mohammad Amine Jradi, Faeze Ghorbanpour, Alexander Fraser