WebSci Companion '26: Companion Publication of the 2026 18th ACM Web Science Conference

Generative AI enables synthetic visual media at a scale that exceeds human verification capacity. This widens an asymmetry on the Web: users encounter images of uncertain provenance with few practical means of assessment. We present Origin Lens, a privacy-first mobile framework that performs cryptographic image provenance verification and AI detection entirely on-device. Built on a Rust/Flutter hybrid architecture, the system combines multiple verification signals—C2PA content credentials, generative model fingerprints, and optional retrieval-augmented context—to provide users with graded confidence indicators at the point of consumption. Unlike centralized platform moderation, this approach allows individuals to verify visual content independently. We discuss how client-side verification infrastructure relates to EU regulatory frameworks (AI Act, DSA) and can complement platform governance for information integrity on the Web.

This text advances the hypothesis that the meaning of the Web as an object of study has diluted as a clear research domain. One example of this phenomenon is the identity crisis of the Web Conference and the International Semantic Web Conference. At its root is the Web’s evolution from a focused technological object into a universal digital environment, a transition whose very success has fragmented its academic community and obscured its core identity. We chart this trajectory from a well-defined object of study to a fragmented backdrop, identifying key pressures such as the “academic tragedy of the commons” and the disruptive force of AI. We conclude that a fundamental community discussion is needed to define what it means to study the Web now that it has become the universal infrastructure for global digital activity.

Manual topic coding of election advertising images is highly time-consuming, yet increasingly required in survey-based research that collects photos and screenshots of campaign materials. We evaluate privacy-compliant, locally deployable vision–language models for automated topic classification of election advertising collected via a smartphone-based high-frequency panel survey during the 2025 German federal election. We compare two approaches: (1) direct image-to-topic classification using Llama 4 (109B) and Qwen2.5-VL-7B, and (2) a modular two-step pipeline in which these models first generate structured image descriptions that are subsequently classified by topic using Llama 4 (109B), Qwen2.5-VL-7B, and the gpt-oss-20b text-only model. Model outputs are evaluated against a human-coded reference of 500 images. Results show that the two-step pipeline substantially improves topic classification when combined with a strong text-based classifier, increasing Macro-F1 from 0.43 (best direct model) to 0.54 (best two-step model). The study provides methodological guidance for designing transparent and privacy-aware pipelines for automated analysis of heterogeneous political visuals in survey research.

Computational methods have opened up new avenues for research in many disciplines, especially in those concerned with large volumes of data like web science. However, many researchers have difficulties to find or apply novel methods or reproduce results because the method documentation is often not useful—especially for less technically oriented scientists—or incorrect. This paper introduces the Methods Hub, a new service dedicated to bring advances from computer and data science to other research communities, with a focus on social scientists. The Methods Hub assists researchers in finding, sharing, and applying computational methods. The service provides computational methods and tutorials with structured and carefully edited descriptions and instructions. To assist scientists in using the methods, the Methods Hub integrates interactive environments, in which methods can be executed and tested online. To increase the reproducibility of research, the Methods Hub allows for exact references to each version of a method via DOIs. Furthermore, the Methods Hub provides method guidelines and templates to simplify the implementation of best practices. The Methods Hub thus serves as community-based toolbox to find, share, and apply methods to social science questions, like those asked in web science.

In the digital world today, online social networks have signif- icantly changed how people communicate, form connections, and share information. While social media platforms provide ample opportunities for connections, they can also lead to divisiveness among the users, such as the formation of echo chambers. An echo chamber is a virtual en- vironment where a person is only exposed to information and opinions that confirm their existing beliefs. In this initial study, we propose a simulation study that considers different aspects of user profile and user behavior and the presence of structural holes to identify the susceptibility and prevention of echo chambers.

Commercial search engines are shifting from being intermediaries to generative answer engines, significantly altering Health Information-Seeking Behavior (HISB). Auditing these “black-box” systems is challenging due to dynamic client-side rendering and restricted data access. To address this, we use a custom-built browser extension to collect search engine results data. In a study of 900 health-related queries in Google US, we found an AI prevalence of 87.7% and a substantially low source overlap of 46.2% between organic results and AI citations. We observe a distinct platform shift: while organic rankings favor clinical authorities, generative AI elevates user-generated multimedia, making YouTube the second-most-cited source. Given high user trust in AI responses, this shift poses risks to public health, highlighting the need for independent auditing toolkits.

Encyclopedic knowledge platforms are key gateways for online information exploration. The recent release of Grokipedia, a fully AI-generated encyclopedia, introduces an alternative to established systems like Wikipedia, raising new questions about how search mechanisms guide users’ exploratory paths. While Wikipedia has been widely studied, no prior work has compared its search outputs with those of newer AI-driven platforms. We address this gap with the first comparative analysis of the search engines of Wikipedia and Grokipedia. Using nearly 10, 000 neutral English words and their substrings, we collect over 70, 000 search results and analyze their semantic alignment, overlap, and topical structure. We find that both platforms often return weakly related or unexpected content from innocuous queries, yet they frequently diverge in the recommendations shown for the same input. Topical annotation and trajectory analysis further reveal systematic differences in category exposure and how results evolve across multi-stage exploration.

Can humans tell whether a news article was written by a person or a large language model (LLM)? We investigate this question using JudgeGPT, a study platform that independently measures source attribution (human vs. machine) and authenticity judgment (legitimate vs. fake) on continuous scales. From 2,318 judgments collected from 1,054 participants across content generated by six LLMs, we report five findings: (1) participants cannot reliably distinguish machine-generated from human-written text (p >.05, Welch’s t-test); (2) this inability holds across all tested models, including open-weight models with as few as 7B parameters; (3) self-reported domain expertise predicts judgment accuracy (r =.35, p <.001) whereas political orientation does not (r = −.10, n.s.); (4) clustering reveals distinct response strategies (“Skeptics” vs. “Believers”); and (5) accuracy degrades after approximately 30 sequential evaluations due to cognitive fatigue. The answer, in short, is no: humans cannot reliably tell. These results indicate that user-side detection is not a viable defense and motivate system-level countermeasures such as cryptographic content provenance.

The systematic curation of the Web remains a central challenge for national libraries and memory institutions that aim to preserve culturally and regionally relevant content. Existing directory-based approaches such as Curlie implement a predominantly topic-centric, one-dimensional hierarchy, where geographic aspects are intertwined with topical and linguistic categories. To address this limitation, we present WebKurator.de, a collaborative platform for combined regional and topical web curation, initially focused on the German web. WebKurator introduces a two-dimensional curation model that explicitly separates topical categorization and geographic annotation. The system integrates LLM–based topic classification and imprint-based address extraction with geocoding, and supports user suggestions together with moderated review. The platform is bootstrapped from the German Imprints Dataset, a large-scale collection of 5.54 million websites. Among them, 3.14 million contain imprint pages, for which we successfully extracted and geocoded postal addresses. Of these, 2.58 million (85.17%) are located in Germany and also have an assigned topic label. These websites form the initial foundation of WebKurator.de and can be continuously extended through user suggestions.

Large language models (LLMs) generate content such as news summaries, product descriptions, and chatbot responses that may reflect and propagate harmful stereotypes. We investigate whether LLMs’ internal representations of stereotypes are encoded linearly and whether we can steer text generated by LLMs when describing social groups. Using mass-mean probing on attention head outputs, we find representations of all 16 dimensions of the Agency-Beliefs-Communion (ABC) model are linearly encoded in Mistral-7B-Instruct and Llama-2-7B-chat. We further demonstrate that inference-time intervention along these directions produces controlled shifts in LLM-generated descriptions of social groups without modifying model weights, offering a transparent and efficient mechanism for mitigating stereotypes in AI-generated content.

The fragmentation of public data in Brazil reflects broader sociotechnical challenges in digital governance, institutional coordination, and knowledge organization. Heterogeneous standards and limited interoperability restrict collective knowledge production, weaken evidence-based policymaking, and reinforce data asymmetries. We introduce Brazil Data Commons, a semantic infrastructure designed to integrate heterogeneous Brazilian public datasets within a shared knowledge framework. By adopting globally recognized ontologies and interoperable data standards, Brazil Data Commons contributes to the development of digital public infrastructure and connects Brazilian data to the broader Data Commons ecosystem. Beyond technical integration, the platform supports new forms of cross-domain analysis, public transparency, and data accessibility. Through intuitive interfaces and flexible access mechanisms, Brazil Data Commons lowers barriers to data use and strengthens the capacity of researchers, policymakers, and civic actors to generate data-driven insights. We hope our system may help researchers to correlate ‘offline’ social-economic indicators with online digital traces. This opens new avenues for digital demography, allowing researchers to integrate traditional data with social media and web activity to better understand Brazil’s complex population dynamics.

Geotargeting is a common affordance of search engines that enables the tailoring of content based on a user’s geolocation. This capability is important during political elections, especially for queries that facilitate participation, such as “where to vote” or ”voting near me”. While U.S. federal law mandates that certain local election administrators provide this information in multiple languages, such as Spanish, our bilingual audit of Google Search results across 223 U.S. localities reveals significant discrepancies in information quality. We find that geotargeting is often inaccurate at both the county and city levels, with a higher mistargeting rate for Spanish-language results. These findings highlight the need for increased efforts to make important electoral information more discoverable by search engines and accessible through alternative means.

Artificial Intelligence (AI) chatbots are increasingly used for emotional, creative, and social support, leading to sustained and routine user interaction with these systems. As these applications evolve through frequent version updates, changes in functionality or behavior may influence how users evaluate them. However, work on how publicly expressed user feedback varies across app versions in real-world deployment contexts is limited. This study analyzes 210,840 Google Play reviews of the chatbot application Character AI, linking each review to the app version active at the time of posting. We specifically examine negative reviews to study how version-level rating trends, and linguistic patterns reflect user experiences. Our results show that user ratings fluctuate across successive versions, with certain releases associated with stronger negative evaluations. Thematic analysis indicates that dissatisfaction is concentrated around recurring issues related to technical malfunctions and errors. A subset of reviews additionally frames these concerns in terms of potential psychological or addiction-related effects. The findings highlight how aggregate user evaluations and expressed concerns vary across software iterations and provide empirical insight into how update cycles relate to user feedback patterns and underscore the importance of stability and transparent communication in evolving AI systems.

We investigate platform-native citation farming on ResearchGate by analyzing almost 3000 papers uploaded by five suspected boosting-service provider accounts. From the uploaded papers and associated metadata, we construct both paper-level and author-level citation networks. We introduce an interpretable structural signal for coordinated boosting, equal references groups: clusters of papers with equal reference lists. We find that many papers from our collection exhibit this motif, that is, they disproportionately cite a small set of authors, consistent with coordinated or automated boosting rather than independent scholarly practice. Finally, we show that for some authors in our dataset a substantial share of their citations can be attributed to these suspicious groups. A different citation network was used to validate the rareness of such motifs in legitimate scientific work.

The trust in the electoral discourse on social media has been eroding in the last years due to a number of factors, including the detection of foreign influence operations, the use of AI-generated content and impersonation, and a shifting geo-political context. Without rigorous analysis of political-related conversations on social media, it is difficult to identify malicious actors and enforce relevant legislation. However, data access restrictions, platform governance, and cross-platform heterogeneity make data collection and analysis a costly process. Using the German 2025 federal election as a case study, we identify key bottlenecks for a systematic analysis of multi-platform social media content related to electoral campaigns, focusing on data collection under constrained access, multi-platform monitoring, and content analysis requirements. We argue that platform observability should be treated as a research infrastructure: a reusable foundation of tools, workflows, and documentation that enables more reliable comparisons across platforms, time, and political contexts.

Vision-language models (VLMs) are increasingly used for image retrieval in cultural heritage collections. Yet these models are trained and evaluated on modern web-scraped data which might cause unobserved biases when applied to historical imagery. While prior work has documented biases related to gender and skin tone in VLMs using contemporary datasets, no evaluation datasets exist for historical archives. In this paper, we address this gap by introducing a new dataset for evaluating demographic bias in VLMs on photograph collections from colonial contexts. Our dataset comprises annotated photographs for demographic queries spanning gender and race. Using this dataset, we examine bias in multiple VLMs through retrieval performance across queries, measuring disparities using established information retrieval metrics. Experimental results show that gender biases documented in contemporary datasets persist on historical imagery, with all tested models retrieving certain demographic groups substantially better than others. Notably, models specifically trained for debiasing fail to mitigate bias on historical images and can introduce new disparities across skin tones. These findings underscore the need for domain-specific debiasing approaches and caution against deploying off-the-shelf VLMs in archival contexts without proper evaluation.

One important challenge in the fight against misinformation is evaluating the reliability of news sources. Although professional organisations manually profile outlets in terms of factors such as political bias and the quality of factual reporting, this process is costly and difficult to scale up. In this preliminary study on News Source Profiling, we aim to predict a publisher’s political bias and level of factual reporting. We construct two datasets, covering 1,571 sources for bias and 1,363 for factual reporting, using Media Bias Fact Check ratings as supervision. We then fine-tune the RoBERTa Large model under a strict, source-disjoint evaluation protocol to prevent data leakage, in order to assess whether stylistic signals alone provide useful cues at the publisher level. Results show that political bias leaves detectable stylistic patterns, while factual reporting proves substantially more challenging, especially at fine-grained levels. We release datasets, model checkpoints, and evaluation metrics to support reproducibility and future research. Our findings provide a baseline and highlight both the potential and the limitations of single-article source profiling.

An AI-generated podcast companion article to augment the PhD Symposium Session of the WebSci Companion '26

NOTE: AI content can contain errors and should be double-checked before use.

The proliferation of Generative AI has precipitated an “Authenticity Crisis,” where synthetic media effectively breaches the “Indistinguishability Threshold” of human perception. This dissertation presents an integrated Design Science framework to quantify this systemic vulnerability across nine chapters. We introduce three open-source artifacts: RogueGPT, a parametric stimulus engine for generating controlled synthetic news; JudgeGPT, a dual-axis perception platform measuring both authenticity and credibility judgments; and Origin Lens, a mobile application implementing C2PA cryptographic provenance verification. Our empirical pipeline reveals four core findings: (1) human detection accuracy for AI-generated news converges to random chance (∼ 50%), independent of age or digital literacy—challenging the “digital native” assumption; (2) a significant “Fatigue Effect” causes detection capacity to degrade under volume, supporting the “Firehose of Falsehood” strategy; (3) a measurable “Liar’s Dividend” where authentic content is dismissed as AI-generated; and (4) expert practitioners express skepticism toward detection tools and preference for provenance-based standards [25]. Building on the CSET RICHDATA and DISARM frameworks, we validate an adapted “Disinformation Kill Chain” using empirical perception data to identify stage-specific countermeasures. Our findings establish that mitigation must shift from “post-delivery” fact-checking to “pre-delivery” provenance and “post-exposure” cognitive inoculation. This work connects technical AI capabilities with human factors research, contributing evidence-based strategies for platform governance in the Web Science domain.

Generative Artificial Intelligence (GenAI) and the creative industries are embroiled in an existential controversy regarding the role of copyrighted creative works. While ongoing research examines the effects of GenAI on creative work in society and the governance of GenAI as a technology, there is limited research that examines this as a single network of associations through the lens of UK public policy. This paper applies Actor-Network Theory (ANT) as an interdisciplinary framework to conceptualise a network and discuss its associations. This paper argues that a) this network is a continuation of an existing paradigm that emerged with platform capitalism, b) that neoliberal characteristics are the key values driving the coalescence of this network, c) the concentration of corporations that represent several parts of the network create an oligopolistic environment where participation is coercive, and d) the long-term stability of this network depends on the extent and nature to which copyright, intellectual property, competition, and other related laws and norms are interpreted and applied.

The regulation of Artificial Intelligence (AI) faces the challenge of balancing innovation with safety and societal well-being. This work challenges the narrative that frames public trust as a means to achieve strategic goals such as attracting investment and maintaining global AI leadership, treating public trust as a strategic instrument. The Society-in-the-Loop (SITL) concept is proposed as an approach to collective and participatory governance. The study focuses on Brazil, an innovation hub in the Global South characterised by a regulatory gap and low levels of political trust. The main contribution is a three-phase research design aimed at operationalising SITL and investigating, from the perspective of Brazilian AI specialists, how public engagement may mediate the relationship between effective regulation and legitimate public trust.

Search query suggestions shape early exposure to political information by steering user attention before search results are retrieved. While biases and harmful representations in search systems are well documented, less is known about how potentially harmful or non-neutral query suggestions shift across geographic locales and language settings, or how their perceived severity can be assessed at scale. This paper presents a large-scale audit of query suggestions for 352 U.S. politicians, comparing Google and Bing across U.S. and EU regional contexts and across different language settings. To quantify the perceived severity of potentially harmful or biased phrasing, we use a scalable framework that combines Large Language Model (LLM) listwise ranking with a Bradley-Terry rating model. The resulting scores are aggregated with an exposure-aware metric to compare search engines, locales, and politician groups. Our analysis reveals clear cross-platform differences in how suggestion spaces adapt to local contexts, with Bing exhibiting stronger linguistic and topical shifts between the U.S. and Europe, whereas Google suggestions remain more static. For both search engines, exposure to potentially biased and harmful query suggestions is highest when the interface language matches the local language setting. We also observe that this exposure correlates with politician popularity and differs across selected demographic groups in some subsets. In a limited human evaluation, LLM-based listwise ranking aligns more closely with aggregated human judgments than direct scalar assessment. Overall, this study provides a reproducible workflow for auditing potentially harmful political query suggestions.

Artificial intelligence tools, particularly large language models, are becoming deeply embedded in everyday writing and editing practices. As a result, Wikipedia now receives AI generated and AI assisted content in many forms, yet the platform lacks clear mechanisms for identifying when and how these tools were used. Existing MediaWiki tags offer only limited metadata about the interface or action that produced an edit and do not identify the type or extent of AI involvement. At the same time, fully prohibiting AI generated text is impractical because many editors already rely on tools such as Grammarly, ChatGPT, Claude, and Google Gemini for grammar correction, summarization, translation, and content drafting. Instead, a principled and well-designed approach for labeling AI involvement is needed that 1) reflects the full range of assistance and 2) supports transparency for both editors and researchers. Building on prior work analyzing inconsistencies in Wikipedia’s Special Tags system, this paper argues that current tagging practices do not match the realities of modern editing workflows. Tool usage is frequently underreported, tag adoption varies widely across languages and topic areas, and editors often have incentives to hide tool involvement to avoid being held responsible for errors introduced by automated systems. Recent work, such as the development of LLM-based image captioning tools for Wikimedia Commons, illustrates that AI participation is already widespread and expanding. Instead of attempting to restrict AI use entirely, this work proposes a more practical strategy. This paper outlines a set of new tags organized into four major categories: Content Creation Tags for textual editing, Assistance and Verification Tags for evaluation and support functions, Metadata Suggestion Tags for organizational elements, and Media-Specific Tags for images, audio, and video. These tags document how, where, and to what extent AI systems contributed to Wikipedia content, providing a foundation for greater transparency, accountability, and informed research on human-AI collaboration.

An AI-generated podcast companion article to augment the Workshop - ABIS 2026: The Effects of the Adaptive Web on Society Session of the WebSci Companion '26

NOTE: AI content can contain errors and should be double-checked before use.

ABIS is an international workshop organized by the SIG on Adaptivity and User Modeling in Interactive Software Systems of the German Gesellschaft für Informatik. For more than 30 years, the ABIS workshop has served as a highly interactive forum for discussing the state of the art in personalization, user modeling, and related areas. The 2026 edition will focus on personalization and recommendations, with particular attention to adaptation on the Web and its societal effects. To explore and discuss the effects of the adaptive Web on society, the workshop aims to bring together researchers and practitioners to share insights and discuss emerging findings and future developments. Our goal is to identify current trends, newly observed effects, and future research directions, with the overarching aim of fostering the development of this discipline and encouraging new collaborations.

This paper introduces an LLM-mediated AI Advisor that contextualizes and synthesizes heterogeneous explainable AI (XAI) outputs to support fast and calibrated misinformation judgments in time-sensitive social media settings. We define LLM-mediated XAI as a process in which a large language model aggregates, prioritizes, and translates heterogeneous XAI outputs into a context-sensitive explanation tailored to the user’s decision situation. Semantic features, XAI modules and LLM-based summarization and synthesis enable the generation of explanations that are adapted in three ways: compressed for time-efficient decisions, translated into non-technical language, and progressively expandable for deeper inspection. Through a mixed-methods user study, including a quantitative study and a qualitative study, we analyze how users interpret, challenge and strategically rely on LLM-mediated explanations during real-world misinformation assessment tasks. The findings indicate that the approach reduces time-to-decision and supports critical inspection without inducing over-reliance. Progressive disclosure and different techniques to present information favored different user needs while conversational functionality was rarely used due to unclear benefits and fear of confusion.

Large language models increasingly generate persona-driven texts for personalized and adaptive web environments, including educational tools, recommender interfaces, creative assistants, and conversational systems. As such systems blur the boundary between human and machine-authored communication, transparent methods for identifying AI-generated persona content become important for user trust, governance, and responsible personalization. We present a feature-based system designed to detect AI-written persona texts. Our approach combines seven groups of writing patterns: multi-domain stylometry, word patterns, reading level, AI predictability, emotional tone, syntactic structure, and lexical diversity. We tested our system on balanced datasets of 500 documents each, comparing human writing with AI text generated from both simple and detailed instructions. Our models identified AI writing with nearly perfect accuracy, reaching F₁ scores of 0.99. We found that AI writing follows predictable regularities-such as using fewer unique words, repetitive punctuation, and consistent reading levels-regardless of how the AI was prompted. These results demonstrate that analyzing multiple writing features provides a reliable and explainable way to detect AI-authored content.

This research set out to examine how using digital tracking tools shape the relationship between parents and children. For this purpose, semi-structured interviews were conducted with participants from four generations, followed by open-ended questions. The findings indicate that the impact of tracking technologies is shaped not only by the tools themselves, but largely by the parental approach and by the quality of the communication surrounding tracking, limits, and boundaries. We conclude that progress does not lie in redefining the mechanisms of surveillance, but in adaptive and context-aware design that promotes trust and communication.

Food literacy in students is a strong lever for future influences. We propose to use virtual reality to foster food literacy by providing adaptive, level-differentiated support. For this, we present a system tailoring decision support for students in a grocery shopping scenario to their personal food literacy levels. The decision support ranges from a fully guided experience to a completely free choice, with only overall result feedback in the end. Our contribution will be tested in an intervention study with a diagnostic questionnaire in German schools.

Wikipedia’s Neutral Point of View (NPOV) policy is central to maintaining the integrity of one of the world’s most widely accessed information sources. As AI-powered moderation tools become increasingly prevalent in adaptive web environments, understanding their reliability and limitations is critical. This paper investigates the effectiveness of Large Language Models (LLMs) in detecting neutrality violations in Wikipedia text using the Wikimedia Neutrality Corpus (WNC). We evaluate three Claude models—claude-3-haiku, claude-haiku-4-5, and claude-sonnet-4-5—across four prompt strategies ranging from direct instruction to policy-guided few-shot prompting. Performance is further analyzed across topical domains, macro-topic categories, bias types (framing, demographic, and epistemological), and discourse communities. Our results show that claude-sonnet-4-5 with NPOV guidelines and detailed explanations achieves the highest accuracy of 69.0%, yet performance varies substantially across domains—medical discourse showing the highest disagreement rate at 64.3%. These findings reveal important risks and limitations of deploying LLMs for automated content moderation on the adaptive web, and offer practical guidance for prompt design in policy-grounded classification tasks.

Short-form video platforms such as TikTok rely on highly adaptive algorithms to curate personalized content streams. While these platforms are widely perceived as effective, one might expect that improvements in personalization would change user-watching behavior, for example, by increasing the proportion of videos watched until the end. However, prior work shows that the fraction of videos watched until the end rarely exceeds 60% and remains largely stable over time. In this paper, we investigate the limits of predicting user-watching behavior—operationalized as whether a video is watched until the end—and examine the extent to which it can be inferred from observable features. We conducted a controlled experiment in which participants interacted with a curated TikTok playlist, allowing us to isolate content-related effects from personalization, and compared these results with real-world data. Across both controlled and real-world settings, simple video metadata, particularly video duration, are the strongest predictors of whether a video will be watched until the end. When incorporating user demographic information, predictive performance improves only marginally, suggesting fundamental limits to modeling user-watching behavior in short-form video contexts. These findings challenge common assumptions about the effectiveness of fine-grained personalization and point to a potential disconnect between perceived vs. actual adaptivity and actual user-watching behavior.

Last-mile deliveries form a critical part of logistics and supply chain management across both business-to-business (B2B) and business-to-customer (B2C) segments. Fulfilment of such deliveries is frequently hindered by incomplete delivery instructions, unreliable location signals, language barriers, and ambiguous addresses. We propose ReActV (Reason–Act–Verify), a prompt engineering methodology inspired by the classical ReAct and Chain-of-Verification (CoVe) paradigms, by incorporating an explicit verification stage and deterministic tools within the prompting loop for grounding courier actions with structured operational context for Large Language Models (LLMs). We present a multi-agent prototype implemented with the DSPy framework and applied to a real-world dataset, which transforms free-form text delivery instructions into structured and actionable insights for couriers. Our focus is on combining stepwise reasoning, action tools and verification tools for delivery planning. The paper provides a comparative evaluation of ReActV against Zero-Shot, Chain-of-Thought-only, and ReAct baselines – with ReActV demonstrating strong performance in verification quality and delivery issue coverage. The proposed approach can be viewed as a context-aware adaptive assistant. It can help address logistics challenges under low-resource settings, particularly for novice couriers in the present-day gig economy, by providing practical insights to enrich human–AI collaboration during last-mile deliveries while improving customer satisfaction and reducing operational costs.

With the growing importance of science-related discourse on the Web and the erosion of the classical knowledge order, it is important to identify different user groups, such as scientists, automatically. This work proposes an approach for identifying scientists and non-scientists on X/Twitter based on their user biographies and tweets. We show that we are able to classify accounts as ’scientists’ and ’non-scientists’ on two different datasets, reaching an F1-Macro score of up to 0.87 using Random Forests with linguistic features and up to 0.89 using a DeBERTa model in an ensemble setup. Furthermore, we provide two datasets with X users labeled as scientists or non-scientists and their respective tweets and user biographies.

Online spaces frequently expose women to sexualised and objectifying content, with documented harms including body dissatisfaction, anxiety, and depression. Automated moderation algorithms compound this through gendered bias by disproportionately classifying benign images of women as sexualised. Deepfake technologies have intensified the harms, with the victims being predominantly women. To counter these developments, we present BiasShield, a browser extension that identifies, audits, and enables users to manage exposure to misogynistic and deepfake content. We report on the design of a multimodal classifier and evaluate its capacity to detect misogynistic content while reducing gender-based false positives. By making algorithmic bias visible and actionable through exposure analytics and protective measures- including optional blurring of offensive content—BiasShield turns content moderation on the web into informed, user-based control.

Online abuse targeting women spreads rapidly in digital spaces, shaped by sociocultural norms that influence how such behavior is perceived, expressed, and addressed across communities. Current detection approaches are largely reactive, identifying harmful content after it occurs without modeling how abuse emerges, escalates, or propagates across communities. We propose a cross-cultural, ontology-guided causal approach that combines structured knowledge representation with causal reasoning to forecast abuse escalation, identify high-risk interactions, and evaluate interventions. By integrating a standardized abuse ontology with cultural context, the framework produces interpretable, explainable outputs that support culturally aware moderation and policy design. Preliminary results on synthetic data illustrate how cultural context shapes escalation dynamics and provides transparent explanations of underlying causal pathways. This paper is a position abstract proposing an ontology-guided causal framework to model and anticipate online abuse across cultural contexts.

Women politicians experience gendered harms facilitated by the use of web-based tools to exclude them from political life, with psychological, professional, and democratic ramifications [1, 2, 3]. Rooted in gender norms that have historically sought to confine women to the private sphere [4], resistance to their participation in politics is now enacted through digital means with unprecedented reach and scale. This presents a uniquely consequential site for examining emerging threats to women’s online safety: gendered digital harms faced by women politicians reinforce broader structures of misogyny and oppression while simultaneously undermining democratic representation, narrowing the range of voices shaping policy and legitimising the exclusion of women from the public sphere altogether [5].

These harms can manifest in several ways, including sexual harassment, threats of violence, disinformation, and image-based abuse [1, 6], and are gendered when their form, content, or targeting is shaped by the politician’s identity as a woman. They are not incidental to the digital environment, or simply a reproduction of offline misogyny. Rather, they are perpetrated using digital tools whose design, development, and deployment reflect and reproduce unequal gendered power dynamics positioning women’s participation in public life as transgressive [7, 8]. Social media platforms are some of the most prominent web-based tools used to perpetrate the aforementioned digital harms, amplifying misogynistic content through algorithmic design, granting anonymity through architectural features, and prioritising engagement over user safety [6, 9]. Generative artificial intelligence (AI) is poised to introduce a new layer into this landscape, especially as content made with generative AI tools is targeted and disseminated through the very same platforms already enabling gendered harms against women politicians [10, 11].

To better understand this layer, it is necessary to examine the capabilities of generative AI tools. Through large language models, image and video generators, and voice synthesis technologies, these tools are able to produce several types of synthetic content [12] that can undermine women politicians’ safety and credibility. Such content can be sexualised, like non-consensual synthetic intimate imagery [13] that predominantly targets women, or purposefully deceptive, like synthetic video or audio shared to misrepresent a politician’s character [14]. Gendered disinformation notably weaponises misogyny and prejudice against women in the dissemination of inaccurate and deceptive information, most frequently targeting women public figures like politicians [10]. While generative AI’s potential to exacerbate gendered disinformation has been recognised in grey literature [15], it has not yet been explored in-depth in academic scholarship. Generative AI may also facilitate additional forms of gendered harm, including coordinated automated attacks at scale, cyber-harassment templates, and the creation of synthetic histories misrepresenting individuals [16], all of which could be utilised against women politicians whose public visibility amplifies their exposure. Much like gendered disinformation, these forms remain prospective and largely unexamined in academic literature. Ultimately, this is an emerging problem at the intersection of technology and power, political representation, and digital governance, developing faster than any of these domains, alone or together, has so far been able to fully address.

This abstract draws on in-progress PhD research focused on the European Union (EU) as a context for investigating the impacts of generative AI on the existing landscape of gendered digital harms against women politicians. The research employs a two-phase qualitative design: first, focus groups with stakeholders across scholarship, practice, and policy domains to map the conceptual terrain, and second, interviews with women Members of European Parliament to centre the lived experiences and perspectives of those most directly affected, and those with the ability to inform EU-wide policy. The EU represents a unique context in this regard, hosting a supranational body with direct regulatory authority over digital technologies that has positioned itself as a global leader on both gender equality and digital governance. Through what has been termed the "Brussels Effect" [17], EU regulatory frameworks have the capacity to set de facto global standards for digital governance, meaning that how the EU responds could influence the global scope of protection available to women politicians. However, while the Digital Services Act and the Artificial Intelligence Act represent landmark interventions in digital policy, both have faced critiques for inadequately encompassing the gendered and political dimensions of the harms they aim to regulate [18, 19, 20].

This is troubling, as women politicians in the EU report experiencing gender-based harassment, sexualised abuse, and reputational attacks [21], experiences that research consistently indicates can deter women from political life and undermine inclusive representation [2, 3]. Evidence from Sweden [22] and Denmark [23] illustrates such chilling dynamics, with politicians reporting that online abuse impacts what they say, how they engage politically, their personal well-being, and their motivation to run for office. Furthermore, during the 2024 European Parliament elections, gendered abuse against women EU candidates surged, including sexualised and race-based attacks [24]. These trends now converge with generative AI tools becoming rapidly more accessible and realistic in their content generation capabilities, dramatically lowering the cost, skill, and time required to produce sophisticated content disproportionately harmful to women politicians. Yet, research investigating how generative AI may facilitate and exacerbate these trends remains limited, obscuring the erosion of conditions under which gender-equal representation becomes possible and constraining policy from adequately responding. This is even more troubling, as the EU risks falling short on both gender equality and digital governance fronts if left unaddressed.

Empirically, there is a distinct lack of research centring the experiences and perspectives of women politicians themselves in relation to gendered generative AI-facilitated harms, which must change to ensure the voices of those most vulnerable are foregrounded in developing adequate responses. Conceptually, existing frameworks for understanding digital harms facing women in politics require further development to capture the specific characteristics of those facilitated by the use of generative AI tools, most notably the realistic nature of content produced, lowered technical and financial barriers to creation, and the speed at which content can be made and disseminated. The role of social media platform architectures in increasing the reach and visibility of content must also be investigated as a core analytical and governance concern. Regarding regulation, the inadequacy of existing frameworks demands attention to several components, such as how harm is defined and measured and what gender-responsive enforcement mechanisms would require in practice. Importantly, women politicians from marginalised communities face compounding forms of targeting, and an intersectional lens should be applied to understand how race, gender expression, religion, and other dimensions of identity shape exposure and severity [25, 26]. This is critical for investigating contexts like the EU, which is made up of twenty-seven member states and a correspondingly diverse population of women politicians, and for developing responsive policies. Addressing generative AI-facilitated gendered harms facing women politicians therefore requires frameworks capable of capturing their entangled gendered and political dimensions, and the full spectrum of intersecting identities through which they are experienced. It is time to interrogate these gaps collectively, and to ask what the cost of failing to act will be for both women politicians and democracy at large.