Program

Keynote: Giada Pistilli, PhD - Principal Ethicist at Hugging Face

Bio: Read more about Giada here

Title: AI Ethics Across Borders: Values, Data, and Cultural Context

Abstract: Ethics in AI isn't just a checkbox: it's about making deliberate choices that shape how technology impacts our lives and societies. As Large Language Models (LLMs) become integrated into our daily activities, we face important questions about their development and deployment — those conversational AI systems are rapidly transforming how we work and communicate. But whose values do they reflect, and who controls the data they're built on? Each language carries unique concepts, values, and worldviews that reflect distinct cultural value systems and ways of understanding reality. When AI training prioritizes dominant languages and perspectives, we risk amplifying existing inequalities and losing the rich diversity that multiple linguistic traditions offer to our collective understanding of reality. Drawing from recent research initiatives like CIVICS, which examines how AI systems respond to culturally sensitive topics, and data governance frameworks developed through the multilingual BigScience workshop, we'll explore how values and cultural contexts shape these technologies. By connecting ethical principles to practical implementation strategies, we can build AI systems that genuinely respect human diversity while delivering meaningful benefits across communities and contexts.

Contributed talk: Prof. Lilja Øvrelid – Leader of the Language Technology Group, UiO

Bio: Read more about Lilja here

Title: How good is your LLM? Systematic evaluation of Norwegian skills and knowledge for Large Language Models

Abstract: In recent years, a number of Large Language Models (LLMs) have been trained specifically for Norwegian. The motivation behind these efforts have been to make available models that generate higher quality Norwegian text and that are better attuned to Norwegian language and values. A critical question moving forward is therefore how we can assess the quality of LLM-generated Norwegian text. In this talk, I will briefly outline the current LLM landscape of Norway and Scandinavia. I will then present NorEval: a new and comprehensive evaluation suite for large-scale standardized benchmarking of Norwegian generative language models. NorEval covers a broad spectrum of task categories targeting Norwegian language understanding and generation, establishes human baselines, and focuses on both of the official written standards of the Norwegian language: Bokmål and Nynorsk. I will present this new resource and the results of benchmarking 19 open-source pretrained and instruction-tuned LLMs for Norwegian. While NorEval constitutes an important first step toward a more informed use of LLMs in Norway, I will conclude with thoughts on the missing pieces and important next steps for LLM research in Norway.

Contributed talk: Anita Schjøll Abildgaard - CEO Iris.ai

Bio: Read more about Anita here

Title: Iris.ai - reflections on a 10-year journey of AI for Research

Abstract: Iris.ai is Norway's first AI startup, and we've been on the forefront of AI for Deep Knowledge and Research for almost a decade. In this talk, we will walk you through our journey - both from a startup, technology and research perspective. The ups and downs, the past and current challenges and the opportunities that lie ahead. The moments where we were about to give up, and the moments where everything fell into place. Of course we will dig into the research and development of our tech stack, presenting some things old and some brand new - like our newly published context sensitivity metric and our critical analysis of reliability in LLM evaluation metrics. And finally we will cover how we navigate our product offerings in a highly shifting AI market.

Contributed talk: Assoc. Prof. Erlend Aune - CTO HANCE

Bio: Read more about Erlend here

Title: ML for real-time Audio Enhancement

Abstract: In AI there is an increasing number of papers achieving state-of-the-art performance without
considering the computational cost or how to achieve real-time inference. These models are -
for the most part - unusable for real-time AI inference on the edge.
In this talk I will present how HANCE works with smaller quasi-causal models, capable of
running efficiently on mobile devices in real-time. I will touch upon the design choices we meet
when making new models, and how we address training of these models.
Throughout the presentation, I will give examples of models, how they work, and which
application they are intended for, while connecting these applications to the design choices we
face when making the models.

Contributed talk: Svein Arne Brygfjeld – Head of the AI Lab, National Library of Norway

Bio: Read more about Svein Arne here

Title: Norwegian language models for Norwegian conditions – dream or reality?

Abstract: The race among tech giants to develop the best language models is intense. At the same time, we see that these same giants are adapting their services to the political frameworks within which they operate, and clear political biases can be observed. Increasingly, people are using these services in more and more areas, and perhaps Norway as a nation is being influenced linguistically and culturally.

As a measure to ensure access to language models that better reflect the Norwegian language and Norwegian society, the Norwegian Parliament (Stortinget) has allocated 40 million Norwegian kroner in the 2025 national budget for the National Library of Norway (NB) to deliver language models for free use by all parts of society. The models will support variants of Nynorsk and Bokmål, as well as Sami languages. They will be built on and reflect the daily reality of Norwegian life. Computational resources will be provided by Sigma2.

The National Library's digital collection, combined with data from other sources, is tailor-made for training language models. However, the copyright restrictions that apply to the collection otherwise remain in place. Therefore, the government has asked the National Library to negotiate with rights holders for permission to use copyrighted textual content in the training of language models.

If successful in both areas, both the public sector and the business community will gain access to Norwegian language models that can be used in service development. The licenses for the models will provide good opportunities for widespread use, including for commercial product development.

Contributed talk: Hans Eide - Special Adviser for User Services, Sigma2

Bio: Read more about Sigma2 here

Title: From Research and Development – to Society and Industry, in the Public Interest

Abstract: To stay competitive globally, to safeguard the environment and resources, and to deliver the best services to customers and citizens, it is essential to adopt the latest advancements in AI. This requires access to vast computational resources and expertise. Equally important is ensuring a seamless and efficient transfer of knowledge and development between research, society, and industry.

AI computational resources demand continuous large-scale investments, deep expertise, and careful management of energy and environmental challenges. Additionally, computational capacity must be utilized as efficiently as possible, striking a balance between national control and governance and relying solely on commercial providers.

For over a decade, Sigma2 and NRIS have provided national high-performance computing (HPC) and data services, primarily for research. Moving forward, there will be a stronger focus on integrating these efforts with industry and public administration. In addition to existing AI resources, Sigma2 is making significant investments to meet growing demands and support this increasingly important collaboration.

In this presentation, we will provide insights into Norway’s computational resources—both existing and upcoming—as well as the expertise and support services available, including the forthcoming LUMI AI Factory.

Contributed talk: Michael Riegler - Head of AI at Simula

Bio: Read more about Michael here

Title: Beyond Text: Alignment Challenges in Multimodal AI Systems

Abstract: As AI systems evolve to process and generate content across multiple modalities - including text, images, audio, video, and physical interactions - the challenge of alignment becomes increasingly complex. This talk examines how the multimodal nature of modern AI systems introduces novel considerations for alignment research. We explore how different modalities might interact in ways that complicate traditional alignment approaches, from potential divergences in concept grounding across modalities to emergent behaviors that arise from modal interactions. Special attention will be given to three key challenges: ensuring consistent ethical behavior across modalities, developing effective oversight mechanisms for multimodal systems, and addressing potential failure modes unique to cross-modal learning. Drawing from recent research in multimodal AI and alignment theory, we propose a framework for thinking about alignment that explicitly accounts for the unique challenges and opportunities presented by multimodal systems.

Keynote: Prof. Robert Jenssen - Director of Visual Intelligence, UiT

Bio: Read more about Robert here

Title: Aspects of XAI in neural multimodal learning (MML)

Abstract: The first part of this talk will outline some basic principles for neural multimodal learning (MML). Focus then shifts to aspects of eXplainable AI (XAI) in the context of MML. Leveraging the so-called CLIP model, the second part of the talk presents an approach to have vision networks “explain” how they function in terms of capturing concepts (such as colors, patterns, objects, etc) with particular focus on “transformer” networks. This is important for the general understanding of such AI systems and to shed light on pretraining vs finetuning. The third part of the talk presents ideas for developing self-explaining MML by learned masks for classification purposes. The aim is to identify the most informative features within each modality that influence predictions.  This is potentially important for users, for reliability and trustworthiness.

Panel discussion: Norwegian foundation models?

Panelists: Lilja Øvrelid, Svein Arne Bygfjeld, Hans Eide, Michael Riegler, Robert Jenssen.

Moderator: Klas Pettersen

Abstract: As powerful AI systems continue to advance, governments around the world are grappling with fundamental questions about technological sovereignty, national security, and responsible deployment. With the rise of large foundation models and agentic AI systems, the stakes are rapidly increasing. Many now argue that AI models and systems are becoming critical national infrastructure—alongside the data centers, compute resources, and electricity that power them.

In this panel, we turn the spotlight on Norway. What level of control should the nation retain—from pretraining to fine-tuning to deploying full AI agents? And how should we address the growing safety and security risks posed by increasingly autonomous AI systems?

Join leading Norwegian experts as we explore strategic options for shaping our national AI future.

Conference dinner at Fredriksten Hotell and NORA Annual Awards Ceremony

Schedule for the evening:

Transportation (Please be on time, the bus will leave):
Arrival:
19:10: Pick-up at Clarion Collection Hotel Park
19:20: Pick-up at Thon Hotel
19:30: Arrival at Fredriksten Hotell (Generalveien 25, 1769 Halden)

Dinner: 
19:30 - 20:00: Mingling/ arrivals:
20:00 - 21:30: Dinner and NORA Awards
21:30 - 23:30: Music 

Return: 
22:30: Pick-up at Fredriksten Hotell (Generalveien 25, 1769 Halden)
22:40: Drop-off at Thon Hotel
22:50: Drop-off at Clarion Collection Hotel Park

Host Keynote: Prof. Irina Engeness - Dept. of Languages, Literature and Culture, HiØ

Bio: Read more about Irina here

Title: AI and the Unsettling Shift in Education: Are We Approaching a Tipping Point or a Transformational Change?

Abstract: Artificial Intelligence (AI) is affecting education, bringing both opportunities and challenges. As AI-driven technologies become more integrated into learning environments, fundamental questions arise: How will AI shape the roles of teachers and students? Will it enhance learning, teaching, and assessment, or does it risk undermining traditional educational practices? Can AI foster creativity and critical thinking, or will it lead to intellectual stagnation and passive knowledge consumption?

This presentation will explore these critical issues, examining the benefits and dilemmas of AI in education. It will consider AI’s role in assessment and feedback, its impact on student engagement, and the growing reliance on generative AI for content creation. Additionally, it will address concerns about the devaluation of human expertise, the risks of bias in AI-driven decision-making, and the broader implications for academic integrity and independent learning.

This session aims to provide a nuanced perspective on the future of AI in education: Are we approaching a tipping point of another wave of technological disruption, or will AI cause fundamental transformational changes in the educational landscape?

Contributed talk: Prof. Anna Smajdor - Practical Philosophy Group, UiO

Bio: Read more about Anna here

Title: Datafication and the Midas touch: ethics in AI-enabled healthcare and beyond

Abstract: The rapid deployment of AI in medicine has made datafication an increasingly integral feature of our health services. Patient data is systematically collected and analyzed; complex human experiences become mere numbers. Datafication transforms everything into digital artifacts, just as Midas turned all he touched to gold. Indeed, these digital artifacts are immensely valuable. The data we can yield is hungrily sought by those who develop AI systems. The ethical response to this has so far focussed on the need to bolster privacy, anonymity and confidentiality, stripping sensitive data of its connections to the individuals from whom it originates. However, this detachment makes it easier for data to be commodified and exploited in ways that may not necessarily align with the original donors’ interests. As an alternative, I propose a new approach to privacy and confidentiality. Instead of obliterating the links between people and their data, the significance of these connections should be recognized and preserved. In turn, this opens the possibility of offering individuals greater involvement and control over the AI technologies developed from their data. Ultimately, anonymisation is not ethically neutral: it feeds into a false perception of data as unclaimed property up for grabs

Keynote: Prof. Samuel Kaski - Leader of FCAI and ELLIS Institute Finland

Bio: Read more about Samuel here

Title: Fundamentals of Applied AI

Abstract: Research is our most promising tool for solving the multitude of important real problems we are facing. By improving the very engine of R&D (research and development) with machine learning, we can contribute to solving a large number of problems with domain experts of the respective fields. This requires contributions from a strong international network of collaborators in academia and industry, and solutions to fundamental machine learning problems that I call the Fundamentals of Applied AI. By these fundamentals I mean solutions that make AI more widely and reliably applicable for decision making in R&D, sometimes called AI4Science but could also be called AI4Research or AI4RDI. I will discuss in particular collaborative AI for human-AI teamwork and out-of-distribution transfer, and how this work is being carried out as collaboration of FCAI, ELLIS Institute Finland, Manchester Centre of AI Fundamentals, and their international partner network.

Keynote: Prof. Amy Loutfi - Program Director of WASP

Bio: Read more about Amy here

Title: Embodied AI: Learning Through Curiosity and Play

Abstract: In this keynote, Professor Amy Loutfi explores how embodied AI—intelligent systems that learn through physical interaction—can be enhanced by integrating human-inspired traits like curiosity and play. Drawing from her research at Örebro University and Linköping University, she discusses how these traits enable robots to adapt to real-world environments, improving perception, reasoning, and human-robot collaboration. As a Wallenberg Scholar, Loutfi's work advances AI's capacity for autonomous learning and interaction.

Panel discussion: From Paris to Halden – The AI Action Summit and AI in the Nordics

We’re moving the global conversation on responsible AI from Paris to Halden!
This panel brings together leading academics from the Nordics and France to explore how AI development can align with democratic values in an increasingly fragmented geopolitical landscape. Amid growing concerns about digital sovereignty, data control, and technological dependency, the discussion will focus on how the Nordic model—rooted in trust, openness, and strong public institutions—can help shape a human-centered AI future. The panel will also reflect on insights from France and the broader European context.

Panelists: Samuel Kaski, Amy Loutfi, Klas Pettersen, Giada Pistilli.

Moderator: Michael Solvang (NORA.ai)

Day 2 closing: Best Poster announcement and location of the NORA Conference 2026