arXiv subjects

Lorenzo Cima

Publications and source records attributed to Lorenzo Cima.

13 recordsLinked to original sources

arXiv2026

Contextualized Counterspeech Can Be More Persuasive Than Generic Counterspeech

AI-generated counterspeech offers a scalable and effective strategy to mitigate online toxicity by promoting more constructive dialogue. Yet, existing approaches adopt a generic, one-size-fits-all paradigm, overlooking the conversational context and characteristics of the targeted users. Here, we propose and evaluate multiple strategies for generating contextualized counterspeech that is adapted to the moderation setting and personalized to the moderated user. In detail, we explore a range of configurations that integrate different forms of contextual information and fine-tuning techniques. We conduct a comprehensive evaluation combining quantitative indicators with a pre-registered, mixed-design crowdsourcing experiment. To ensure robustness, we implement algorithmic measures of counterspeech quality based on ROUGE, BLEU, and BERTScore, observing overall consistent results across metrics. Furthermore, we analyze which characteristics of both the generated counterspeech and the moderated toxic message most strongly influence perceived persuasiveness, yielding insights into how contextualized interventions can be made more effective. Our findings show that personalization can be effective, but not uniformly so. Lightweight strategies combining conversational context and user history improve perceived adequacy and persuasiveness, whereas several other contextualization strategies degrade human-perceived counterspeech quality. Taken together, these results provide actionable directions for developing more personalized, effective, and responsible counterspeech systems, ultimately advancing human-AI collaboration in online content moderation.

cs.HC↗

arXiv2026

A Geometric Analysis of Small-sized Language Model Hallucinations

Hallucinations -- plausible but factually incorrect responses -- pose a major challenge to the reliability of Large Language Models (LLMs), especially in multi-step or agentic settings. Existing work largely frames hallucinations as a consequence of missing knowledge; we show instead that, even when the relevant factual knowledge is present, models still produce hallucinated answers, pointing to retrieval instability rather than knowledge gaps. Building on this observation, we introduce APORIA (Aggregate Prompt-wise Observation Retrieving Instability via Asymmetry -- the state of puzzlement-in-contradiction that hallucinations embody), a geometric framework that studies repeated responses to the same prompt in sentence-embedding space. Our central hypothesis is that genuine responses cluster more tightly than hallucinated ones; we empirically validate this and show that, after Fisher projection, the two response classes become consistently separable. We leverage this asymmetry in geometry via APORIA-LP, an efficient label-propagation method that classifies large collections of responses from as few as 30--50 annotations, achieving F1 scores above 90% across ten small-sized LLMs. To support further research, we release SOCRATES-300K, a fully labelled dataset of 300,000 responses, together with the code for both dataset generation and result reproduction. Our key finding -- framing hallucinations from a geometric perspective in the embedding space -- complements traditional knowledge-centric and single-response evaluation paradigms, paving the way for further research.

cs.CL↗

arXiv2026

Assessing the Reliability of Persona-Conditioned LLMs as Synthetic Survey Respondents

Using persona-conditioned LLMs as synthetic survey respondents has become a common practice in computational social science and agent-based simulations. Yet, it remains unclear whether multi-attribute persona prompting improves LLM reliability or instead introduces distortions. Here we contribute to this assessment by leveraging a large dataset of U.S. microdata from the World Values Survey. Concretely, we evaluate two open-weight chat models and a random-guesser baseline across more than 70K respondent-item instances. We find that persona prompting does not yield a clear aggregate improvement in survey alignment and, in many cases, significantly degrades performance. Persona effects are highly heterogeneous as most items exhibit minimal change, while a small subset of questions and underrepresented subgroups experience disproportionate distortions. Our findings highlight a key adverse impact of current persona-based simulation practices: demographic conditioning can redistribute error in ways that undermine subgroup fidelity and risk misleading downstream analyses.

cs.CY↗

arXiv2026

The Big Ban Theory: A Pre- and Post-Intervention Dataset of Online Content Moderation Actions

Online platforms rely on moderation interventions to curb harmful behavior such as hate speech, toxicity, and the spread of mis- and disinformation. Yet research on the effects and possible biases of such interventions faces multiple limitations. For example, existing works frequently focus on single or a few interventions, due to the absence of comprehensive datasets. As a result, researchers must typically collect the necessary data for each new study, which limits opportunities for systematic comparisons. To overcome these challenges, we introduce The Big Ban Theory (TBBT) -- a large dataset of moderation interventions. TBBT covers 25 interventions of varying type, severity, and scope, comprising in total over 339K users and nearly 39M posted messages on Reddit and Voat. For each intervention, we provide standardized metadata and pseudonymized user activity collected three months before and after its enforcement, enabling consistent and comparable analyses of intervention effects. In addition, we provide a descriptive exploratory analysis of the dataset, along with several use cases of how it can support research on content moderation. With this dataset, we aim to support researchers studying the effects of moderation interventions and to promote more systematic, reproducible, and comparable research.

cs.SI↗

arXiv2025

Dark Personality Traits and Online Toxicity: Linking Self-Reports to Reddit Activity

Dark personality traits have long been associated with antisocial and toxic online behaviors, yet their relationship with observable online activity remains unclear. We investigate the association between validated dark personality measures, self-reported experiences of online incivility, and linguistic and behavioral features extracted from real-world user activity. To this end, we developed a Web application that securely links responses to validated psychological questionnaires collected via Amazon Mechanical Turk with participants' Reddit activity. This yielded a dataset of nearly 57K comments (2.2M tokens) from 114 users, represented through a broad set of linguistic and behavioral features. Our analyses reveal a clear distinction between self-reported and observed behavior. Dark personality traits show consistent associations with self-reported engagement in uncivil interactions. However, no validated dark personality dimension significantly predicts text-derived toxicity or linguistic features. In contrast, self-reported experiences of engaging in or being targeted by toxic behavior are robustly reflected in users' language, exhibiting consistent associations with measures of negativity, moral framing, and emotional intensity. Taken together, these findings highlight a gap between stable personality traits and their manifestation in surface-level linguistic signals. While computational features effectively capture behavioral engagement in online incivility, they do not provide reliable proxies for underlying personality constructs within the present framework. Our results underscore the importance of grounding computational approaches in validated psychological measures and point to the need for richer, context-aware representations to better understand the relationship between personality and online behavior.

cs.CY↗

arXiv2025

PRISM: Phase-enhanced Radial-based Image Signature Mapping framework for fingerprinting AI-generated images

A critical need has emerged for generative AI: attribution methods. That is, solutions that can identify the model originating AI-generated content. This feature, generally relevant in multimodal applications, is especially sensitive in commercial settings where users subscribe to paid proprietary services and expect guarantees about the source of the content they receive. To address these issues, we introduce PRISM, a scalable Phase-enhanced Radial-based Image Signature Mapping framework for fingerprinting AI-generated images. PRISM is based on a radial reduction of the discrete Fourier transform that leverages amplitude and phase information to capture model-specific signatures. The output of the above process is subsequently clustered via linear discriminant analysis to achieve reliable model attribution in diverse settings, even if the model's internal details are inaccessible. To support our work, we construct PRISM-36K, a novel dataset of 36,000 images generated by six text-to-image GAN- and diffusion-based models. On this dataset, PRISM achieves an attribution accuracy of 92.04%. We additionally evaluate our method on four benchmarks from the literature, reaching an average accuracy of 81.60%. Finally, we evaluate our methodology also in the binary task of detecting real vs fake images, achieving an average accuracy of 88.41%. We obtain our best result on GenImage with an accuracy of 95.06%, whereas the original benchmark achieved 82.20%. Our results demonstrate the effectiveness of frequency-domain fingerprinting for cross-architecture and cross-dataset model attribution, offering a viable solution for enforcing accountability and trust in generative AI systems.

cs.CV↗

arXiv2025

Emoji Reactions on Telegram: Unreliable Indicators of Emotional Resonance

Emoji reactions are a frequently used feature of messaging platforms, yet their communicative role remains understudied. Prior work on emojis has focused predominantly on in-text usage, showing that emojis embedded in messages tend to amplify and mirror the author's affective tone. This evidence has often been extended to emoji reactions, treating them as indicators of emotional resonance or user sentiment. However, they may reflect broader social dynamics. Here, we investigate the communicative function of emoji reactions on Telegram. We analyze over 650k crypto-related messages that received at least one reaction, annotating each with sentiment, emotion, persuasion strategy, and speech act labels, and inferring the sentiment and emotion of emoji reactions using both lexicons and LLMs. We uncover a systematic mismatch between message and reaction sentiment, with positive reactions dominating even for neutral or negative content. This pattern persists across rhetorical strategies and emotional tones, indicating that emojis used as reactions do not reliably function as indicators of emotional mirroring or resonance of the content, in contrast to findings reported for in-text emojis. Finally, we identify the features that most predict emoji engagement. Overall, our findings caution against treating emoji reactions as sentiment labels, highlighting the need for more nuanced approaches in sentiment and engagement analysis.

cs.HC↗

arXiv2025

Hallucination Detection: A Probabilistic Framework Using Embeddings Distance Analysis

Hallucinations are one of the major issues affecting LLMs, hindering their wide adoption in production systems. While current research solutions for detecting hallucinations are mainly based on heuristics, in this paper we introduce a mathematically sound methodology to reason about hallucination, and leverage it to build a tool to detect hallucinations. To the best of our knowledge, we are the first to show that hallucinated content has structural differences with respect to correct content. To prove this result, we resort to the Minkowski distances in the embedding space. Our findings demonstrate statistically significant differences in the embedding distance distributions, that are also scale free -- they qualitatively hold regardless of the distance norm used and the number of keywords, questions, or responses. We leverage these structural differences to develop a tool to detect hallucinated responses, achieving an accuracy of 66\% for a specific configuration of system parameters -- comparable with the best results in the field. In conclusion, the suggested methodology is promising and novel, possibly paving the way for further research in the domain, also along the directions highlighted in our future work.

cs.CL↗

arXiv2024

Contextualized Counterspeech: Strategies for Adaptation, Personalization, and Evaluation

AI-generated counterspeech offers a promising and scalable strategy to curb online toxicity through direct replies that promote civil discourse. However, current counterspeech is one-size-fits-all, lacking adaptation to the moderation context and the users involved. We propose and evaluate multiple strategies for generating tailored counterspeech that is adapted to the moderation context and personalized for the moderated user. We instruct an LLaMA2-13B model to generate counterspeech, experimenting with various configurations based on different contextual information and fine-tuning strategies. We identify the configurations that generate persuasive counterspeech through a combination of quantitative indicators and human evaluations collected via a pre-registered mixed-design crowdsourcing experiment. Results show that contextualized counterspeech can significantly outperform state-of-the-art generic counterspeech in adequacy and persuasiveness, without compromising other characteristics. Our findings also reveal a poor correlation between quantitative indicators and human evaluations, suggesting that these methods assess different aspects and highlighting the need for nuanced evaluation methodologies. The effectiveness of contextualized AI-generated counterspeech and the divergence between human and algorithmic evaluations underscore the importance of increased human-AI collaboration in content moderation.

cs.HC↗

arXiv2024

Investigating the heterogenous effects of a massive content moderation intervention via Difference-in-Differences

In today's online environments, users encounter harm and abuse on a daily basis. Therefore, content moderation is crucial to ensure their safety and well-being. However, the effectiveness of many moderation interventions is still uncertain. Here, we apply a causal inference approach to shed light on the effectiveness of The Great Ban, a massive social media deplatforming intervention on Reddit. We analyze 53M comments shared by nearly 34K users, providing in-depth results on both the intended and unintended consequences of the ban. Our causal analyses reveal that 15.6% of the moderated users abandoned the platform while the remaining ones decreased their overall toxicity by 4.1%. Nonetheless, a small subset of users exhibited marked increases in both the intensity and volume of toxic behavior, particularly among those whose activity levels changed after the intervention. However, these reactions were not accompanied by greater activity or engagement, suggesting that even the most toxic users maintained a limited overall impact. Our findings bring to light new insights on the effectiveness of deplatforming moderation interventions. Furthermore, they also contribute to informing future content moderation strategies and regulations.

cs.CY↗

arXiv2024

Human and LLM Biases in Hate Speech Annotations: A Socio-Demographic Analysis of Annotators and Targets

The rise of online platforms exacerbated the spread of hate speech, demanding scalable and effective detection. However, the accuracy of hate speech detection systems heavily relies on human-labeled data, which is inherently susceptible to biases. While previous work has examined the issue, the interplay between the characteristics of the annotator and those of the target of the hate are still unexplored. We fill this gap by leveraging an extensive dataset with rich socio-demographic information of both annotators and targets, uncovering how human biases manifest in relation to the target's attributes. Our analysis surfaces the presence of widespread biases, which we quantitatively describe and characterize based on their intensity and prevalence, revealing marked differences. Furthermore, we compare human biases with those exhibited by persona-based LLMs. Our findings indicate that while persona-based LLMs do exhibit biases, these differ significantly from those of human annotators. Overall, our work offers new and nuanced results on human biases in hate speech annotations, as well as fresh insights into the design of AI-driven hate speech detection systems.

cs.CL↗

arXiv2024

Beyond Trial-and-Error: Predicting User Abandonment After a Moderation Intervention

Current content moderation follows a reactive, trial-and-error approach, where interventions are applied and their effects are only measured post-hoc. In contrast, we introduce a proactive, predictive approach that enables moderators to anticipate the impact of their actions before implementation. We propose and tackle the new task of predicting user abandonment following a moderation intervention. We study the reactions of 16,540 users to a massive ban of online communities on Reddit, training a set of binary classifiers to identify those users who would abandon the platform after the intervention -- a problem of great practical relevance. We leverage a dataset of 13.8 million posts to compute a large and diverse set of 142 features, which convey information about the activity, toxicity, relations, and writing style of the users. We obtain promising results, with the best-performing model achieving micro F1-score = 0.914. Our model shows robust generalizability when applied to users from previously unseen communities. Furthermore, we identify activity features as the most informative predictors, followed by relational and toxicity features, while writing style features exhibit limited utility. Theoretically, our results demonstrate the feasibility of adopting a predictive machine learning approach to estimate the effects of moderation interventions. Practically, this work marks a fundamental shift from reactive to predictive moderation, equipping platform administrators with intelligent tools to strategically plan interventions, minimize unintended consequences, and optimize user engagement.

cs.CY↗

arXiv2024

The Great Ban: Efficacy and Unintended Consequences of a Massive Deplatforming Operation on Reddit

In the current landscape of online abuses and harms, effective content moderation is necessary to cultivate safe and inclusive online spaces. Yet, the effectiveness of many moderation interventions is still unclear. Here, we assess the effectiveness of The Great Ban, a massive deplatforming operation that affected nearly 2,000 communities on Reddit. By analyzing 16M comments posted by 17K users during 14 months, we provide nuanced results on the effects, both desired and otherwise, of the ban. Among our main findings is that 15.6% of the affected users left Reddit and that those who remained reduced their toxicity by 6.6% on average. The ban also caused 5% users to increase their toxicity by more than 70% of their pre-ban level. Overall, our multifaceted results provide new insights into the efficacy of deplatforming. As such, our findings can inform the development of future moderation interventions and the policing of online platforms.

cs.SI↗