Posts

Blog: Tackling Out-of-Context Misinformation with Self-Supervised Learning

Misinformation refers to false or misleading information that spreads without the intent to deceive. It differs from disinformation, which is deliberately created to mislead. Misinformation can take many forms—completely fabricated stories, manipulated images, or misrepresented facts. In today’s digital age, social media plays a significant role in the rapid spread of such content, often amplifying misleading narratives before they can be fact-checked. The consequences of misinformation range from minor misunderstandings to large-scale public confusion, political unrest, and harm to communities. One of the most challenging types of misinformation to detect is out-of-context misinformation—a tactic that misuses authentic images, videos, or quotes by presenting them in a misleading way. Unlike fabricated content, out-of-context misinformation is harder to identify because the media itself is real—the misleading part lies in how it is framed. For example, an image from a past natural disa...
Read more

Blog: Discussion on AI Assistance in Academic Writing

During this week's group meeting, I2SC group members shared specific ways they have used LLMs to improve their writing and discussed best practices for incorporating LLMs into a writing workflow. As a teaser to the discussion, Jianlong prepared a "Journalist or AI" game in which group members were invited to attribute authorship to pairs of news-like paragraphs. We noticed that: LLM-generated text tends to have a perfect sentence structure but is often verbose.  Human-written text may break secondary-school conventions but is often effective in succinctly communicating ideas and data.  Ingmar is very good at this game. Consensus reached by participants included: Never using LLMs to create an initial draft of a paper, because *you* should be the one to decide the overall structure and direction of the writing. LLM-generated feedback for your writing can be helpful, but it tends to be overly positive--even if you specify that you are aiming for a competitive ...
Read more

Upcoming Block Seminar: AI in the Global South

Are you curious about how AI is shaping the world? Do you want to make a difference in communities with fewer resources? Are you a student at UdS looking to explore new ideas and technologies?   If your answer to any of these is yes , then our seminar on AI in the Global South is for you!   AI in the Global South: Artificial Intelligence (AI) and Large Language Models (LLMs) are rapidly transforming sectors such as education, healthcare, and agriculture across the world. While these technologies hold immense potential, their application in the Global South —which includes regions such as India, Brazil, and Nigeria, as well as marginalized communities within developed nations—faces unique challenges. These include digital inequality, limited language representation, and pressing ethical concerns that demand more inclusive and context-aware AI solutions. This seminar is offered by Dr. Vikram Kamath Cannanure , a postdoctoral fellow at I2SC. He specializes in Human Computer Int...
Read more

Blog: Definitions of Forest in Remote Sensing and Climate Change

In the forestry, ecology, climate change, conservation, and remote sensing fields, accurately defining what constitutes a forest is essential for effective monitoring, dynamics assessment, and policy formulation in areas focused on carbon offsetting and tracking. A clear, general, and comprehensive understanding of forest characteristics is, therefore, necessary for establishing baselines for reforestation, tracking deforestation, mapping ecological dynamics, and accurately quantifying the achievements of carbon offset goals. In different fields, the forest is defined based on structural, ecological, and functional criteria, and some definitions also consider the socioeconomic benefits that forests provide. Most definitions of forests focus on tree cover, canopy density, area covered, and tree heights [7]. For instance, the Food and Agriculture Organization (FAO) defines forests as areas with a tree cover of more than 10%, a minimum height of 5 meters, and an area of at least 0.5 hecta...
Read more

Blog: Large Language Models for Curriculum Mapping with a Focus on Employable Skills

In the era of rapid technological evolution, where required employable skills are expected to change rapidly, higher education institutions (HEIs) need to prepare students for an uncertain and dynamic job market. Large Language Models (LLMs) - tools like GPT-4 - are emerging as a promising way to tackle this challenge by assisting with curriculum mapping and design.   Curriculum mapping is the process of aligning course descriptions, learning activities, and assessments with the skills needed for employment. Traditionally, experts carry out this process manually. However, this manual approach requires considerable resources and faces several difficulties. For instance, experts may interpret employable skills differently, and human error or bias can occur. These issues make it harder to update curriculum maps frequently, even though the job market is rapidly changing.  Recent studies have investigated whether large language models (LLMs) can automate this process [1]. Their fin...
Read more

Bevorzugte Twitter die Tweets rechter Politiker? Internationale Studie dazu endet unfreiwillig 2023

Read more

Blog: Why should we be careful when using LLMs as human participants?

Large Language Models (LLMs) have the ability to act as a specific personality based on a given description. Researchers have used this ability in simulating LLMs as replacement for human participants in user studies, annotation tasks, opinion surveys, and Computational Social Science. When researchers are using LLMs in such a way, they are assuming that LLMs are able to represent the perspectives of different demographic identities because of the vast training data they were trained on. However, the paper by Wang et. al. [1], ' Large language models that replace human participants can harmfully misportray and flatten identity groups ' show that there is an inherent flaw in the current approaches of training these LLMs that make them misportray and flatten representation of identity groups. The authors thus urge caution in use cases where LLMs are intended to replace human participants.  There are mainly three limitations in using LLMs as human participants' replacement: Mi...
Read more