Weizenbaum Library (Weizenbaum Institute)
Not a member yet
864 research outputs found
Sort by
Tik-to-Tok: Translating Language Models One Token at a Time: An Embedding Initialization Strategy for Efficient Language Adaptation
Get PDFTraining monolingual language models for low and mid-resource languages is made challenging by limited and often inadequate pretraining data. In this study, we propose a novel model conversion strategy to address this issue, adapting high-resources monolingual language models to a new target language. By generalizing over a word translation dictionary encompassing both the source and target languages, we map tokens from the target tokenizer to semantically similar tokens from the source language tokenizer. This one-to-many token mapping improves tremendously the initialization of the embedding table for the target language. We conduct experiments to convert high-resource models to mid- and low-resource languages, namely Dutch and Frisian. These converted models achieve a new state-of-the-art performance on these languages across all sorts of downstream tasks. By reducing significantly the amount of data and time required for training state-of-the-art models, our novel model conversion strategy has the potential to benefit many languages worldwide
Shaping digital transformation for a sustainable society
No full textThe second ‹Bits & Bäume› conference took place in Berlin in 2022. Once again, it provided a space for critical tech and sustainability communities to share ideas and collaborate towards the common goal of shaping digitalisation to foster sustainability. This companion book compiles the insights, work, research and opinions of more than 65 authors with a ‹Bits & Bäume› background, including practitioners, researchers and activists. The articles included in this journal demonstrate the progress made in merging ‹Bits› and ‹Bäume› (Trees) topics since our first publication in 2019 by addressing different sub-areas of the intersections between digitalisation and sustainability. Encompassing a wide range of topics, the articles delve into pressing challenges such as the resource consumption, power implications and democratic governance of digital infrastructures, AI, blockchains, mobile apps, and other software applications, as well as the need to address the unsustainable practices and paradigms of e.g., the platform economy. Offering not only transparency but also solutions, the journal presents practical approaches and concepts related to the necessary transformation, such as the Computer Science for Future programme. It also contains articles commenting on current political developments, such as the EU legislation on sustainability and freedom-related aspects of ICT devices. Further articles highlight the power of and need for an active civil society, aiming to inspire activism. This journal caters for everyone: Are you just getting into the topics around Bits & Bäume? Have you been involved in this field for many years, or are you an expert in one of the areas touched on here? In this journal you will find both introductory topics, such as illustrations on the challenges of today’s digitalised society, and also advanced topics, such as conceptual and regulatory discussions. Whatever your background, we think you’ll enjoy the read, learn something new on the way, and get inspired. Ultimately, we are all united by the overarching goal of shaping digitalisation as part of a necessary socio-ecological change; one which contributes to a sustainable and just society.The Bits & Bäume Conference 2022 and this subsequent publication were funded by the Deutsche Bundesstiftung Umwelt (DBU) and the Federal Ministry for the Environment, Nature Conservation, Nuclear Safety and Consumer Protection (BMUV). The work of Germanwatch and the Forum InformatikerInnen für Frieden und gesellschaftliche Verantwortung (FIfF) in the context of the conference was funded by Stiftung Mercator
Einsatz von KI in der Medizin: Haftung und Versicherung
Get PDFDer zunehmende Einsatz von künstlicher Intelligenz (KI) in der Medizin gibt Anlass, das Verhältnis von Arzthaftung (Behandelndenhaftung) und KI-Haftung näher zu beleuchten. Dabei sind drei grundsätzliche Arten der Risikozuweisung zu unterscheiden: Die Risikotragung durch die Patienten bzw. Sozialversicherungen als Ausgangspunkt, die Risikotragung durch die Behandelnden (Behandelndenhaftung), sowie die Risikotragung durch die Hersteller bzw. Betreiber der KI (KI-Haftung). Um die rechtliche Risikozurechnung beim Einsatz von KI in der Medizin darzustellen, wird zunächst kurz das allgemeine Risikopotenzial beim Einsatz von KI in der Medizin dargestellt (I.), um sodann die drei wesentlichen Zurechnungspole zu beleuchten, nämlich die Behandlungsrisiken, KI-Risiken und Patientenrisiken (II.), wobei die wesentlichen Risikoquellen sowie die Risikoverteilung im Allgemeinem überblicksartig dargestellt werden. Darauf aufbauend werden drei grundlegende Mechanismen der rechtlichen Risikozurechnung dargestellt: Zunächst wird das Arzthaftungsrecht als klassischer Risikoverteilungsmechanismus bei medizinischen Behandlungen aufgezeigt, das über eine Risikotragung durch den Behandelnden oder durch den Patienten (bzw. dessen Versicherung) entscheidet (III.). Sodann wird das bei KI-Sachverhalten hinzutretende KI-Haftungsrecht dargestellt, das zwischen einer Risikotragung durch die Hersteller, durch die Betreiber oder durch die Betroffenen unterscheidet (IV.). Zuletzt wird im Rahmen der besonderen Patientenrisiken gezeigt, welche Möglichkeiten einer Vergemeinschaftung von Risiken sowohl bei Behandlungsrisiken als auch bei KI-Risiken diskutiert werden (V.). Es wird deutlich, dass die Haftung de lege lata bereits zu einer angemessenen Risikoverteilung führt, es durch die Einführung neuer KI-Haftungsregelungen de lege ferenda aber zu einer Verschiebung der Risikotragung hin zu den Herstellern von KI-Systemen kommen könnte
Exploring the German-Language Twittersphere: Network Analysis of Discussions on the Syrian and Ukrainian Refugee Crises
Get PDFThis study conducts a comparative analysis of Twitter communication networks relating to the Syrian and Ukrainian refugee crises. Employing a network analysis approach, the study uses approximately 660,000 tweets to gain insights into the online discussion communities surrounding these crises. Tweets specifically discussing Syrian refugees were collected between 2015 and 2023, while those about Ukrainians were harvested from 2022 to 2023, utilizing the full-archive search endpoint of the Twitter API. By transforming retweets into communication networks between users, the study investigates the community structure within these networks. The findings reveal that the online anti-refugee community is smaller in size, more active, highly interconnected, and transcends national boundaries, in contrast to the opposing communities. These results underscore the need for increased social media engagement of pro-refugee voices and improved moderation practices to foster a more inclusive virtual public sphere.This work has been funded by the Federal Ministry of Education and Research of Germany (BMBF) (grant no.: 16DII121, 16DII122, 16DII123, 16DII124, 16DII125, 16DII126, 16DII127, 16DII128 – “Deutsches Internet-Institut”)
Defending Informational Sovereignty by Detecting Deepfakes? Opportunities and Risks of an AI-Based Detector for Deepfakes-Based Disinformation and Illegal
No full textThis paper investigates possible contributions that an AI-based detector for deepfakes could make to the challenge of responding to new forms of cyberthreats, including fraud and disinformation as a threat to democracy. The paper investigates the implications of such a tool for the emerging European discourse on digital sovereignty in a global environment. While cybersecurity and disinformation are certainly not new topics, recent technological developments relating to AI-generated deepfakes have increased the manipulative potential of video and audio-based content spread online, making it a specific but important challenge in the global and interconnected information context.The work in this paper is partly funded by the German Federal Ministry of Education and Research (BMBF) under the FAKE-ID project (grant numbers OVGU FKZ: 13N15736 and HWR/FPÖS FKZ: 13N15737)
From Joseph Weizenbaum to ChatGPT: Critical Encounters with Dazzling AI Technology
Get PDFThe paper considers AI systems from a use perspective. It focuses on conversational chatbots, starting from Weizenbaum’s ELIZA and sketching the major scientific advances leading up to ChatGPT. The main discussion builds upon several experiment-reflection cycles conducted by the author to explore ChatGPT as a knowledge resource. The analysis considers ChatGPT responses in terms of accuracy, structure, context, perspective, and bias. The critical evaluation begins with the observation that ChatGPT produces a mixture of clear and precise results and arbitrary misinformation without ever clarifying its own scope. This leads to the identification of the system’s key problem, namely, how it contends with truth, which involves replacing the idea of truth with a probabilistic surrogate based on textual correlation. In responsible use, a system like ChatGPT must be embedded in a human learning culture. A framework for this process should include an insistence on truthfulness, an impulse towards enhancing human competence, and strengthened responsibility structures within communities.This publication has been funded by the Federal Ministry of Education and Research of Germany (BMBF) (grant no.: 16DII121, 16DII122, 16DII123, 16DII124, 16DII125, 16DII126, 16DII127, 16DII128 – “Deutsches Internet-Institut”)
Active Social Media Use and Its Impact on Well-being: An Experimental Study on the Effects of Posting Pictures on Instagram
Get PDFActive use of social networking sites (SNSs) has long been assumed to benefit users’ well-being. However, this established hypothesis is increasingly being challenged, with scholars criticizing its lack of empirical support and the imprecise conceptualization of active use. Nevertheless, with considerable heterogeneity among existing studies on the hypothesis and causal evidence still limited, a final verdict on its robustness is still pending. To contribute to this ongoing debate, we conducted a week-long randomized control trial with N = 381 adult Instagram users recruited via Prolific. Specifically, we tested how active SNS use, operationalized as picture postings on Instagram, affects different dimensions of well-being. The results depicted a positive effect on users’ positive affect but null findings for other well-being outcomes. The findings broadly align with the recent criticism against the active use hypothesis and support the call for a more nuanced view on the impact of SNSs.This work has been funded by the Federal Ministry of Education and Research of Germany (BMBF) under grant no. 16DII127 (“Deutsches Internet-Institut”) and grant no. 16DII131 (“Weizenbaum-Insitut e.V.”)
Same, same but different? Explaining issue agendas of right-wing parties’ Facebook campaigns to the 2019 EP election
No full textSocial media are important for right-wing parties to communicate with and mobilize potential voters in election campaigns. Our study focuses on the Facebook campaigns of right-wing parties in six European countries and aims to understand which issues were transnationally shared and which ones emphasize national perspectives on the agenda of the populist actors. We ask what context conditions on the party- and country-level determine the individual issue agendas. Using structural topic modelling, we analyze the communication of the Austrian FPÖ, the German AfD, the French RN, the Italian Lega, the Polish PiS, and the Swedish SD during the 2019 EP election campaign. To explain their issue agendas, we run logistic regression models testing the influence of country-specific and party-specific factors. Our analyses establish that while right-wing parties across Europe are similar in pushing a few populist issues like blaming elites and immigration, they still engage in campaigning on national politics
Digital Inclusion of Low-Literate Adults: Challenging the Sequential Underpinnings of the Digital Divide
Get PDFContemporary models of digital inclusion and the digital divide assume that developing the digital literacy that enables individuals to participate in society is a sequential and linear process that is more or less similar for all individuals in all contexts and requires basic linguistic skills. This paper challenges these understandings, arguing that such a technical, normative perspective excludes marginalized and disadvantaged publics, such as low-(digital) literate citizens. Based on a longitudinal ethnographic study of low-literate Dutch adults, we show that the often-described causal relation between (digital) literacies, (digital) participation, and (digital) inclusion is not as evident as it seems and neglects the important socio-cultural contexts through which (digital) literacies are often gained and enacted in everyday practice. Consequently, we argue that current conceptualizations of (digital) inclusion and (digital) participation need to be rethought in terms of the limitations, potential, and capabilities of low-literate people.This work has been funded by the Federal Ministry of Education and Research of Germany (BMBF) (grant no.: 16DII111, 16DII112, 16DII113, 16DII114, 16DII115, 16DII116, 16DII117 – „Deutsches Internet-Institut“
Weizenbaum Report 2023: Political Participation in Germany
Get PDFThe Weizenbaum Report is an annual publication on political participation in Germany. It presents findings from a survey representative of the population conducted by the Weizenbaum Panel research unit since 2019. The focus is on investigating various forms of political participation over time and their development against the backdrop of digitalization. In addition to the changing role of digital media for political action, the fourth wave from 2022 focused on authoritarian attitudes that act as drivers of anti-democratic participation. The effects of social inequality on political engagement were also investigated. The results reveal ambivalent attitudes towards the effects of AI and social media as well as patterns of political participation of authoritarian-minded people.This work has been funded by the Federal Ministry of Education and Research of Germany (BMBF) (grant no.: 16DII111, 16DII112, 16DII113, 16DII114, 16DII115, 16DII116, 16DII117 – „Deutsches Internet-Institut“