At this year's workshop of the Special Interest Group ABIS of the German Informatics Society (Gesellschaft für Informatik), held as part of WebSci 2026 in Braunschweig, Germany, seven accepted papers were presented and discussed. The workshop brought together researchers from the Web Science community working on personalization and adaptive systems across a wide range of application domains, including social media, explainable AI, delivery support, digital surveillance, and health education.

As personalization increasingly becomes part of our everyday lives, understanding its broader societal implications is becoming just as important as advancing the underlying technology. Therefore, each paper presentation was followed by an interactive discussion focusing on one central question: How might the proposed personalization approach affect society—today and in the future? The reflections summarized below combine the key ideas presented in the papers with the discussions that emerged during the workshop.

Evaluating LLM-Based Bias Detection on Wikipedia

The paper Evaluating LLM-Based Bias Detection on Wikipedia: A Prompt Strategy Analysis Using NPOV Guidelines (Ashik Ahamed, Max Wang, and Jeanna Matthews) investigated how different prompting strategies influence the ability of LLMs to detect biased Wikipedia content. The authors demonstrated that model performance strongly depends on prompt engineering and varies considerably across domains and discourse communities.

Regarding the societal implications of using LLMs as large-scale moderation tools, we discussed several aspects. While adaptive prompting enables more effective content moderation, incorrect classifications—particularly in medical, scientific, or technical domains—could amplify misinformation or reduce access to reliable information. Since Wikipedia influences search engines, recommendation systems, and AI training data, these effects may extend far beyond the platform itself. The paper illustrates how adaptive AI systems require careful domain-specific customization and continuous evaluation rather than one-size-fits-all solutions.

Literacy Level based Nutrition Education in Virtual Reality

The paperLiteracy Level based Nutrition Education in Virtual Reality (Dennis Wüppelmann, Sarah Claudia Krings, Carina Micheel, and Corinna Ostwinkel) presented a gamified virtual reality environment that adapts educational content and feedback to a user's food literacy level.

Concerning the societal impact, we discussed the considerable potential of such personalized educational systems to promote healthier lifestyles in an engaging way. At the same time, questions regarding accessibility and social equity arise. Personalized health interventions should consider differences in family background, financial resources, and social circumstances to avoid unintentionally reinforcing existing inequalities. Looking ahead, similar concepts could also be transferred to real-world shopping scenarios using augmented reality, smart glasses, or mobile applications.

Exploring the Limits of Predicting User Watching Behavior with Short-Form Videos on TikTok

The paper Exploring the Limits of Predicting User Watching Behavior with Short-Form Videos on TikTok (Carolina Coimbra Vieira, Sepehr Mousavi, Oshrat Ayalon, Abhisek Dash, Krishna Gummadi, and Savvas Zannettou) examined how users' watch duration can be predicted using content characteristics, user behavior, and demographic information. Interestingly, demographic features contributed only marginally to prediction accuracy.

We discussed that this finding suggests that detailed user profiling may not always be necessary for effective personalization and that, in some domains, content metadata alone may be sufficient. From a societal perspective, two contrasting developments are worth considering: On the one hand, increasingly accurate prediction models may further optimize platforms for engagement and time spent, potentially reinforcing behaviors such as doomscrolling. On the other hand, the same knowledge could support interventions that encourage healthier and more conscious social media use by helping users better regulate their online behavior. From both perspectives, an important question is how such research findings should ultimately be used and who benefits from them.

LLM-Mediated XAI Explanations

The paper LLM-Mediated XAI Explanations: A Decision Co-Pilot for Fast and Calibrated Judgments on Potential Misinformation (Valentin Grimm, Eelco Herder, Jessica Rubart, and Carsten Röcker) proposed personalized AI-generated explanations that help users assess the trustworthiness of online content more efficiently.

Concerning societal effects, the discussion centered on balancing usability and critical thinking. While adaptive explanations can improve decision-making and reduce cognitive effort, we also emphasized the importance of preventing overreliance on AI recommendations. Future personalized explanation systems should therefore aim not only to improve efficiency but also to support users in developing and maintaining their own critical judgment.

Structuring the Last Mile with ReActV

The paper Structuring the Last Mile with ReActV: Tool-Augmented Delivery Planning with Verification (Pushkal Agarwal, Joyjit Chatterjee, Akash Marar, and David Gamtenadze) introduced an adaptive delivery planning system that personalizes routing instructions while incorporating human-in-the-loop verification through stepwise confirmation and escalation.

During the reflection session, participants discussed how adaptive technologies can also strengthen rather than diminish human agency. In contrast to fully automated decision-making, the proposed approach demonstrates how personalization can support workers while preserving opportunities for human intervention and control in algorithmically-managed work environments.

Detecting Persona-Generated AI Text

The paper Detecting Persona-Generated AI Text with Interpretable Linguistic Features (Elena Slavova, Angelina Parfenova, and Juergen Pfeffer) addressed the challenge of distinguishing AI-generated from human-written text using interpretable linguistic features.

Concerning its effects on society, both the opportunities and limitations of such approaches were discussed. Reliable detection could improve transparency and support content moderation, but false positives may negatively affect users who legitimately use AI as a writing assistant. Furthermore, as AI-assisted writing becomes increasingly common, human language itself may evolve over time, raising questions about the long-term validity of datasets and evaluation methods used for AI text detection.

Tracking Childhoods

The paper Tracking Childhoods: Inter-generational Perspectives on Parent-Child Relationships and the Ethics of Digital Monitoring (Eelco Herder and Ana Cîrnu) explored how different generations perceive digital monitoring in parent-child relationships.

The discussion extended beyond the presented study to broader societal questions concerning privacy, trust, and autonomy in the digital era. Participants reflected on whether increasing reliance on digital monitoring technologies might gradually replace interpersonal trust within families as such practices become increasingly common. While monitoring technologies may provide safety benefits in critical situations, they also create a fundamental trade-off between protection and privacy that becomes increasingly relevant as digital monitoring becomes more pervasive.

Overall Reflections

Across all discussions, several recurring themes emerged. First, personalization is not limited to traditional recommender systems but spans a wide range of application domains, increasingly driven by AI-driven approaches for intelligent decision support, adaptive education, and workplace assistance. Second, the societal impact of personalization depends not only on technical performance but also on how these systems influence autonomy, trust, privacy, fairness, and digital well-being.

Overall, one of the most important insights from the workshop was that personalization rarely has purely positive or purely negative consequences. The same adaptive technology may empower some users while disadvantaging others, depending on the context in which it is deployed. Reflecting on these trade-offs is therefore essential for designing future adaptive systems that benefit both individuals and society.

See all papers in the companion proceedings of WebSci'26: https://dl.acm.org/doi/proceedings/10.1145/3795513#heading3