Project descriptions
Contact person: Sune Dueholm Müller
Status: Project
This project proposes a paradigm shift in the deployment of AI and machine learning models for complex, mission-critical settings, moving from a static “waterfall” approach to an Action-Oriented Modeling framework. While the scientific community has made substantial progress in developing statistically accurate predictive models, a critical “implementation gap” routinely prevents these tools from supporting real-world decision-making. Current conceptions of veridical AI focus on the internal validity of models, i.e., predictability, computability, and stability, but largely neglect the sociotechnical conditions required for veridical outcomes in operational deployment. By embedding model development within a continuous cycle of co-creation with end users and decision-makers, this project will develop and validate a framework that extends veridical AI to encompass external validity: how models remain accurate, interpretable, and trustworthy when transferred to new institutional and data contexts at scale. The framework will be grounded and validated through an initial use case in climate-sensitive health modeling across Low- and Middle-Income Countries, and is designed to serve as transferable methodological infrastructure for TRUST’s broader portfolio of projects deploying AI in complex institutional settings.
Status: Project
Contact person: Ole Jakob Elle
Contact person: Francesco di Ruscio
Status: Project
Understanding and predicting changes in human behaviour is a complex challenge. Yet representing these changes is essential for mathematical models of real-world systems, such as traffic and mobility systems or infectious-disease transmission models. Human behaviour and societal activity leave signals across routinely collected data streams, including public transport usage, financial transactions, mobility data and electricity consumption. These data differ substantially in temporal and spatial resolution, structure and uncertainty, ranging from conventional time series to graph-based data.
This project will investigate how AI and machine-learning methods can combine such heterogeneous data sources to detect, characterise and forecast changes in behaviour and infer their impact on contact structures in near real time. The central methodological challenge lies in integrating signals with different structures, resolutions and uncertainties, and connecting them to mechanistic models through hybrid AI-mechanistic approaches. The ambition is to turn continuous societal data streams into dynamic representations of human behaviour that can update mathematical models as the systems they represent evolve. Rather than replacing mechanistic models with black-box predictions, the aim is to make them more adaptive while preserving their interpretability and explicit representation of underlying processes.
Infectious-disease modelling will provide a key application and methodological test case. Transmission is strongly influenced by changes in contact patterns, mobility and behaviour, which traditional mechanistic models often do not incorporate dynamically. Building on our agent-based modelling framework for infectious disease transmission, we will investigate how AI-derived estimates of changing contact structures can continuously inform the model, creating a digital twin of epidemic dynamics. This will allow us to test whether early signals of behavioural transitions can improve situational awareness, forecasting and the exploration of emerging transmission scenarios.
The findings will support public-health authorities by improving situational awareness and scenario analysis. More broadly, the methods could be applied to other complex societal systems modelled using heterogeneous and continuously changing data. User partners could contribute data, real-world use cases and domain expertise, and help validate both the identified behavioural signals and the practical usefulness of the resulting models.
Participating institutions: Norwegian Institute of Public Health (NIPH); Norwegian Computing Center (NR); Arctic University of Norway (UiT)
Key project members: Francesco Di Ruscio (NIPH); Birgitte F. de Blasio (NIPH); Jørgen E. Midtbø (NIPH); Filippo Maria Bianchi (UiT); Ingrid Dæhlen (NR); Geir Storvik (NR, UiO)
Contact person: Francesco Di Ruscio (francesco.diruscio@fhi.no)
Relevant Action Clusters: AC1 - AI in healthcare and medicine; AC10 - AI for time-series analysis and anomaly detection; AC11 - AI-guided decision making; AC12 - AI for security and preparedness; AC15 - AI for sciences.
Contact person: Anders Strand
Status: Project
AI tools are being adopted rapidly across professional fields from medicine to research to education. Typically, this happens faster than we gain understanding of their effects. This project asks: how does reliance on AI change what responsible, trustworthy professional conduct looks like? And what's the best way to study this?
We aim to build a transdisciplinary framework that brings together philosophy, ethics, law, and social science to assess the real-world impact of AI in professional settings (WP1). We combine diverse disciplinary methods with a bottom-up approach rooted in the perspectives of those who use these tools in practice. Real-world use often reveals risks that abstract or lab-based standards miss. For instance, studies have found that routine reliance on AI can gradually erode professionals' own skills and judgment, a risk that could extend to research practices such as data analysis or writing.
The project includes case studies. One is a PhD position (WP2) examining how use of generative AI affects either research and research practices (case 1) or health services (case 2). A third case (WP3) concerns university education, where we plan to study how AI affects teaching and critical thinking in ex.phil (a mandatory introductory philosophy course with a representative student population).
Together, the case studies will inform a practical, transdisciplinary and evidence-based framework for navigating and assessing the role of AI tools in professional conduct.
Anders Strand (SPS, OsloMet), Joey Pollock (IFIKK, UiO), Mike Barnes (IFIKK, UiO), Bjarte Askeland (BI), Hege Hermansen (SPS, OsloMet), Andreas Eriksen (SPS, OsloMet), Anna Smajdor (IFIKK, UiO), Kim Pedersen Phillips (IFIKK, UiO).
Relevant action clusters include: AC1: AI in healthcare and medicine, AC3: AI for case processing and management in the public and private sector, AC14: Lifelong learning and AI, and AC15: AI for science.
Contact person: Tuukka Mäkitie
Status: Project
AI-ST investigates how AI use can be governed to contribute to decarbonization of energy and transport sectors. AI-ST develops a first-of-its-kind transformative innovation policy framework for governing AI applications in sustainability transitions (i.e., fundamental changes in production and consumption patterns). The development of this framework requires theorization, methodological toolkits and exploration of empirical cases. To achieve this the project builds on literature and policy reviews, interviews, empirical case studies, regulatory analysis and transdisciplinary research. AI-ST’s aim is to equip policymakers, researchers and other stakeholders with knowledge and tools to nudge AI use in the direction of solving societal problems like climate change, rather than letting AI undermine the efforts to solve such challenges.
Contact person: Crina Damsa
Status: Pre-project
SimAI examines how AI-supported technologies offer new ways to understand feedback, communication, coordination and stress during teamwork among students and professionals in healthcare and police training. High-stakes professions rely on simulation training to build teamwork and decision-making skills under pressure. SimAI develops an AI-enhanced solution that transforms critical team performance data into feedback moments and investigates whether and how this automated feedback fosters learning of team skills.
What makes this novel is combining a reflection methodology, multimodal data and generative AI to give participants a near-real-time, evidence-based picture of their own performance — including automated identification of team communication and actions, tagging of key events, and potentially AI-generated "best practice" video examples.
Our dream outcome is new knowledge on evidence-based learning of collaborative behaviour, and a validated, AI-enhanced feedback system that makes debriefing more objective, timely and pedagogically effective.
In practice, SimAI offers a methodology and digital platform for nursing and police education, with potential to redefine evidence-based, scalable learning across high-risk professional domains that use simulation training. Training programmes and other high-risk environments — including emergency services and defence — could potentially use SimAI to give students and professionals clearer, data-backed feedback on teamwork, improving both training quality and preparedness for real emergencies.
Contact person: Thomas Espeseth
Status: Project
Our research asks how information is prioritised in minds, brains and artificial networks, and what happens when information moves between such systems. Does priority in one system affect what is selected for transmission, and what becomes prioritised or preserved in the next?
We use memorability as a measurable expression of what persists in memory. Some images and texts are reliably more memorable than others, and these differences can be measured, predicted and experimentally varied.
In WP1, we study how AI systems select information. We use attribution and intervention methods to investigate what information models rely on, and compare this with what they ultimately retrieve, summarise or present. One possibility is that AI systems preferentially surface information that humans also find memorable, even though they were not designed to do so. We also test whether properties of the models' own representations better explain what gets selected. Where these measures diverge may be as informative as where they converge.
WP2 moves to the human receiver. We vary the memorability of AI-generated information and examine retention, understanding, inference, transfer, confidence and calibration. We also ask whether gist-level information can influence later cognition even when detail is lost.
WP3 asks whether these effects extend beyond a single encounter, influencing what people later seek, repeat or use, and what adaptive AI systems subsequently present.
We welcome collaboration involving AI selection, retrieval, explanation or summarisation, and human-AI, learning or decision settings where information exposure can be linked to later cognition or behaviour.
Contact person: Malcolm Langford
Status: Project
CASPAI (Case Processing with AI) will develop and test an interdisciplinary meta-framework for trustworthy AI in public decision-making, using asylum adjudication as a deliberate "hard case". Asylum falls within four of the EU AI Act's high-risk categories and concerns highly vulnerable people, so solutions that work there are designed to generalise across public administration and the courts, including TRUST user partners. The project builds on a research infrastructure of almost one million Nordic-Baltic asylum decisions, and runs alongside the NordForsk-funded NordAId project (Oslo, Copenhagen, Uppsala, Vilnius). Three questions guide it: when can AI-based public decision-making be trustworthy, how can trustworthiness be enhanced, and do these innovations produce calibrated trust – neither overtrust nor undertrust? Five work packages cover veridical and privacy-secure AI; anticipatory governance under the AI Act; a spectrum of decision-support models, from legal source recommenders to generative drafting and agentic AI, plus a preparation tool for asylum seekers; explainable AI and uncertainty quantification; and responsible evidence-based innovation through two-phase experiments with stakeholders in four countries. Participatory design and shared value metrics run throughout.
Contact person: Guendalina Righetti
Status: Project
AI systems increasingly rely on complex internal representations that can be difficult for humans to understand. This project asks: can cognitively grounded formal frameworks anchor statistical and neural models, and can cognitive models of concepts help us understand and constrain their latent representations?
The project investigates whether formal models of human conceptual knowledge can guide how AI represent features and internal representations, with the aim of making AI more interpretable and capable of robust generalisation. A central goal is to clarify what constitutes a meaningful feature and how features relate to concepts, bridging Cognitive Science, Philosophy, Knowledge Representation, and Machine Learning.
The expected outcomes are: (1) a principled, AI-operational account of features and their relation to concepts; (2)computational formalisations of cognitive models that can guide AI learning; (3) benchmarks for evaluating whether AI representations capture meaningful conceptual features, with a focus on interpretability, generalisation, and cognitively plausible failure; and (4) contributions to hybrid AI combining machine learning, logical reasoning, and cognitive models of concepts.
In practice, the results could support AI systems with more understandable representations and more robust, predictable behaviour, particularly in domains where trustworthy AI is important, such as law.
The project is intrinsically interdisciplinary and would benefit from collaboration with Machine Learning and statistical modelling researchers to integrate and test the formal models. User partners could contribute use cases, data, and validation environments to identify and evaluate relevant application scenarios.
Contact person: Martin Jullum
Status: Project
What parts of an LLM’s training data are responsible for the answer it generates to a given prompt? This project develops scalable, fine-grained methods for tracing the outputs and behaviour of large language models back to influential training examples and sources. Combining explainable AI, machine learning, natural language processing and statistical methods, we aim to move beyond simple influence scores and distinguish different contributions, such as factual content and linguistic style. The methods will be implemented and tested on modern LLMs and across training settings, including fine-tuning and learning from examples. The results could help model developers, auditors and other users understand and verify model outputs, diagnose errors and bias, improve training-data curation, support targeted unlearning, and build fairer, safer and more trustworthy language models.
Contact person: Nina Sunde
Status: Pre-project
The Deepfakes and Crime project examines how deepfakes and other forms of synthetic media may affect criminal investigations, digital evidence, and decision-making throughout the justice system. The project comprises three interrelated sub-projects: (1) a systematic review of research on the implications of deepfakes for criminal investigations, evidential integrity, media authentication, and judicial assessment; (2) a mapping of existing technologies and methods for deepfake detection and media authentication, including their capabilities and limitations; and (3) interviews with police investigators, digital forensic practitioners, judges, and relevant public- and private-sector stakeholders to explore practical challenges, needs, and perspectives concerning synthetic media across the justice system. Bringing together perspectives from law, social science, and technology, the project will produce an evidence-and-gap map of the current research landscape, a technology map of relevant detection and authentication approaches, and a stakeholder-informed research agenda for a larger follow-up project. These outputs will identify knowledge gaps and research priorities that can inform the future development of investigative practices, evidential guidance, training, and independent evaluation of deepfake-detection and authentication technologies. The work is informed by use cases developed by police user partners and input from partners associated with Action Cluster AC6, Combatting Criminality with AI.
Contact person: Pierre Lison
Status: Project
Can we tell whether a text was written by a human or generated by AI? Several techniques exist, based on AI-detectors, advanced statistical analysis, or even hidden watermarks embedded in the model responses. Yet all of these approaches remain imperfect, often struggling with new models and deliberate attempts to evade detection. The project objectives are to:
- advance the state of the art in AI detection methods
- connect those to legal/regulatory aspects, notably transparency requirements
- and explore how AI detection methods may affect the use of AI tools in various settings, notably in education.
Contact person: Nandor Knust
Status: Pre-project
ECO-WAR-NET will ask what happens to the legal protection of the environment in armed conflict when the facts on which that protection depends are increasingly produced by machines. International humanitarian law protects the natural environment through the prohibition on "widespread, long-term and severe" damage and the requirement that environmental harm enter the proportionality assessment, yet in four decades these rules have gone almost entirely unapplied — not, the project will argue, because states disagree about the norms, but because the law demands factual determinations about the extent, severity and expected consequences of harm that it has no method for making. Artificial intelligence is now transforming this evidentiary landscape on both sides of the strike: after the event, machine-learning change detection, satellite remote sensing and open-source investigation document conflict damage at scale; before the event, digital twins — AI-driven computational replicas of oceans, battlespaces and infrastructures already entering military planning — simulate how an environment would respond to intervention, making credible prediction of environmental harm possible for the first time. Taking the Arctic marine environment as its object — the ocean region where the law is least developed, where a fragile, slow-recovering ecosystem makes any harm most likely to be long-term and severe, where military activity and undersea-infrastructure vulnerability are intensifying fastest, and where Norway's own security and ocean-governance stakes are highest — the project will ask whether, when and with what authority algorithmically generated and simulated evidence can ground — or be demanded for — a legal judgement, and how such evidence reshapes the legitimacy of targeting decisions, drawing on the Black Sea, Red Sea and Baltic as the contemporary practice against which the Arctic framework will be tested. To be anchored at the Norwegian Centre for the Law of the Sea (NCLOS) at UiT The Arctic University of Norway — Norway's national centre on UNCLOS, marine environmental law and Arctic ocean governance — the project would unite NCLOS's IHL/ICL and law-of-the-sea scholarship with NUPI's expertise on the data, modelling and machine-learning systems through which militaries render the world targetable, green criminology, and INTERPRT's forensic-visual practice at NTNU, within TRUST — The Norwegian Centre for Trustworthy AI, whose Action Cluster on AI in Armed Conflict the PI co-leads. ECO-WAR-NET would build a machine-readable ontology of Arctic marine conflict harm, theorise the conditions under which AI-derived evidence can carry legal weight, validate the reconstructed doctrine experimentally with legal practitioners, and prototype EcoProp, an AI-supported methodology for anticipatory environmental proportionality assessment — testing, for the first time, how far legal judgement itself is computable, and where human judgement must remain, as simulation and algorithmic evidence become the infrastructure of decision-making in war.
Contact person: Arnoldo Frigessi
Status: Pre-project
When deploying and aligning modern Large Language Models, two controls can be adjusted when you prompt the LLM and ask to answer a question or solve a problem: sampling temperature and alignment steering.
Temperature controls diversity of the answer or solution that you get. A low temperature will make the answer deterministic, and you would observe a lack of creativity and it will soon get boring. A too high temperature, then words come at random, and you get just garbage out. Mathematically, the temperature is a rescaling the probability distribution of the answer.
Alignment steering can be used to steer the answer or solution in a wished direction, so that it aligns to what you want: have a polite tone, align with a law, behave safely. It works by adding a vector to your prompt, that directs the model's internal state to shift its behaviour, all without retraining. However if you steer too hard, again, the model sticks to your alignment vector too much and does not solve your problem well. A too low alignment steering will not help to align.
Both controls share a major flaw: they have a failure threshold past which the generated text completely collapses. Until now, the only way to find these usable limits was trial and error.
In this pre-project we work on this problem. We try to control such negative emerging behaviors a priori, by deciding the temperature and the strength of the alignment steering before we use the LLM. We do this by using the lens of statistical mechanics: the probabilities of the tokens needed to write the output act as energy levels and the steering vector acts as an external field.
What does this mean in practice? It means we can compute the exact usable window in advance of temperature and alignment steering. So, our results should give developers zero-cost diagnostic tools computed straight from the weights, eliminating brute-force parameter optimisation and allowing deployed LLM systems to operate safely within their true limits. Please take contact with us if you wish to test our tools in your LLMs.
Contact person: Filippo Bianchi
Status: Project
TruSTFM will develop a reusable foundation model for Earth-observation and climate data. Its core idea is to learn jointly from heterogeneous spatio-temporal sources—including satellite imagery and weather fields on regular grids, irregular station networks, and event records—across different spatial resolutions and uneven observation times. The model will combine a common representation for grids and graphs with multi-scale and continuous-time modelling.
Current models typically specialise in one data geometry, spatial resolution, sensor, or task. They therefore transfer poorly to new regions and struggle with missing observations or irregular sampling. TruSTFM is novel in addressing these limitations within one model and coupling this with a standardised, uncertainty-aware evaluation framework. This will make claims about transfer, robustness, and “foundation” capabilities more transparent and comparable.
Expected outcomes include a trained model core with input modules for multiple data types; open software for data processing, training, and evaluation; reusable benchmarks and transfer protocols; documentation; scientific publications; and two demonstrators. The first will combine coarse weather forecasts, geographic information, and sparse station observations for regional climate prediction and downscaling. The second will use SAR/InSAR time series and contextual data for surface-change and hazard monitoring. These results could support weather and climate services, environmental monitoring, renewable-energy analytics, natural-hazard management, Earth-observation services, and research communities working with sparse or heterogeneous sensor data.
Collaboration with user partners is central to ensuring practical relevance. Partners can contribute operational datasets, representative use cases, domain requirements, expert-labelled events, and validation under realistic conditions such as sensor outages, missing modalities, changing station networks, and deployment in new regions. They could also help integrate and test project outputs within existing forecasting or monitoring workflows.
Contact person: Tobias Mahler
Status: Project
Agentic AI is evolving quickly, and definitions remain contested. A workable characterisation is that such systems break high-level goals into sub-tasks, plan, retain memory across tasks, invoke external tools and services, and coordinate with other agents. This marks a shift from static, single-task models towards autonomous actors that intervene in the world, potentially with minimal human oversight. The shift promises significant gains in productivity, but it implies risks and strains existing governance frameworks. GALTS brings together law, computer science, philosophy, social science and risk governance in four prongs: (1) Design for safety, accountability and compliance: how agentic AI systems can be technically designed to ensure safety and to support traceability and legal compliance. (2) Systemic risk management: how organisations can anticipate, assess and manage systemic risks from agentic AI, which often stem from feedback loops between multiple agents, their environments, and wider socio-technical systems. (3) Regulatory challenges: whether current frameworks, including the EU AI Act and the GDPR, are adequate to the risks that agentic AI poses. (4) Human and AI agency: goal formation in human-AI interaction, empirically and normatively, including how agentic systems shape immediate preferences and longer-term interests, whose objectives they pursue, and what standing machine-generated objectives can claim in collective decisions.
Contact person: Sten Ludvigsen
Status: Pre-project
How do individual and collective cognition emerge when using AI; and in which ways create AI disruptions and expansions of social and cognitive/knowledge processes in multiple types of settings.
Premise 1: When powerful new tools and infrastructures are designed and used, we need to understand and explain how cognition emerges and changes, and how this influences the division of labor between humans and AI (including different types of automation and augmentation). As AI systems are used and scaled, they become cultural tools that affect both micro- and macro-level dynamics.
Premise 2: After these descriptions, we must develop appropriate causal explanations in order to predict how different AI systems can affect human cognition in the short and long term. This part will contribute to new conceptualizations across multiple fields of knowledge.
This project aims to analyze different conceptualizations of AI and why they are important for studying emerging cognition and knowledge in specific domains. Two different conceptualizations are often used: hybrid intelligence and complementary intelligence. Both bring different ideas and require further conceptual work. However, the term complementary intelligence offers a clearer stance from both cognitive AI and machine AI perspectives and seems the most promising.
The goal is to develop new scientific causal explanations of the interplay between humans and AI, not only within AI systems but also in broader sociotechnical contexts.
User partners in higher education will gain new insights into the design of learning in domain-specific areas and into what kinds of knowledge about generative AI are needed to learn and use AI in those domains. In work settings, we can establish local causal procedures for implementation and generate new knowledge about what it takes to create collective cognition/intelligence/knowledge in organizational practice.
Contact person: Crina Damsa
Status: Pre-project
TeamLearnAI explores how students and teams learn when generative AI becomes part of the process — not merely a tool, but a thinking partner. We examine how AI-powered, adaptive feedback shapes individual and group learning across disciplines such as law, medicine and education. The study's central focus is understanding what human and non-human agency — and their interplay — look like once AI-generated feedback is woven into human activity.
What makes this novel is our empirical, real-world focus: rather than speculating about AI's potential, we test how students actually engage with AI feedback, how their thinking and collaboration change, and who remains "in control" of learning when AI is involved.
Our dream outcome is a validated model, combining a reflection-based training method with an AI-enhanced dashboard. In practice, this means universities could adopt a ready-to-use platform that that gives students tailored, discipline-specific feedback and supports their critical thinking and study progression.
Contact person: Monika Nerland
Status: Project
AI-based tools have the character of epistemic technologies with capacities to generate knowledge, perform analytical processes, and act upon the knowledge generated. These capacities bring new opportunities, but also serious challenges to professional expertise. On the one hand, they offer potential for augmented human capabilities, on the other hand the risks of undermining human safeguarding of knowledge claims and the further development of expert communities. How these challenges are met, and with what implications, will need to be studied in specific professional and technological contexts.
This project will examine TRUST-developed AI technologies as they are tested or taken into use in socially critical work in the public sector. That is, work where judgement is critical, demands for reliability are high, and trust in decisions is at stake.
The project will a) develop novel, empirically grounded insights into different forms of human-AI interaction and their implications for expertise, which b) lay the ground for targeted learning and professional development initiatives in the work settings and c) inform the further development of trustworthy AI in relevant domains.
Contact person: Alex Lenkoski
Status: Project
The digital revolution has facilitated markets for a broad range of goods and services. As the volume and complexity of these markets grows, algorithmic pricing naturally takes over, with agent-based reinforcement learning (RL) algorithms becoming increasingly prevalent. When a small number of RL algorithms control the pricing for a large share of a market, there is increasing concern that they may begin to exhibit the same sort of anticompetitive behavior that competition law prohibits when perpetrated by human actors. This project will develop a class of RL algorithms designed to trade in digital markets with a particular focus on understanding, anticipating and mitigating the negative consequences of interaction between independent agents. Our intent is to (1) extend the theory of RL trading algorithms to markets driven by complex heterogeneous processes; (2) create a general and extendable testbed for new RL research, development and deployment; (3) develop a flexible class of "whistle blower" agents that are capable of monitoring market actors and assessing the chances of tacit or explicit algorithmic; and (4) clarify and update important legal concepts in competition law related to collusion, use of insider information and systemic risk when algorithms act and interact autonomously in digital markets.
Contact person: Geir Kjetil Sandve
Status: Project
Imagine you are deploying AI in a low-resource area in Ethiopia to forecast climate-sensitive diseases like malaria. You have almost no local data, but a massive global model exists trained with lots of data from other places. Right now, you face a forced choice: either build a standalone local model that probably fails due to data scarcity, or accept a global model that forces generic assumptions on you. The real problem is that the global model doesn't care about your local context – for example it uses its own loss function. It doesn't know that for your specific task, you need another loss function.
Our project solves this fundamental tension between global learning and local specificity and autonomy. We propose a Hub-and-Spoke framework designed specifically for settings where spokes cannot simply take the hub model and use it. In our architecture, the Hub holds shared data, computational power, and common AI models. The Spokes represent local contexts, particular users or customers, even individual patients. The Spoke wants to solve the same task as the hub, but in its own way. Importantly, the spoke has an own private loss function that it would have liked the hub to use in training, and that it want the spoke-based model to follow. The core innovation of our work is Spoke Autonomy. Methodologically, this project links to world models, alignment steering, knowledge representation, agentic AI.
So this project bridges the gap between centralized AI power and local decision-making. It ensures that when AI is deployed in high-impact, data-sparse environments, stakeholders never have to sacrifice their local autonomy just to benefit from global tools.
So what real-world problem will we be able to solve in a new and very competitive way? If you have a situation where you have many users and you want to tell them that with your model, service or product, they really can personalise the model, adapt it to their needs and priorities. If you make autonomy and freedom of your clients/user your marketing strength. Talk to us, we need your problem as case!
Contact person: Erik Velldal
Status: Project
Multimodal text–image models have rapidly advanced in their ability to generate rich textual descriptions of visual content or, conversely, to generate images from a given textual description. However, neither their textual nor visual depictions are necessarily neutral. For example, when a model maps an image into language, this “translation” between modalities may also introduce bias through particular framing, emphasis, omission, or evaluative language. This project will develop methods and datasets for analyzing ideological bias in multimodal text–image models, addressing methodological challenges in multimodal benchmarking as well as broader questions about bias in how AI systems represent social and political realities.
Contact person: Arnoldo Frigessi
Status: Project
Contact person: Benjamin Ricaud
Status: Project
While Large Language Models (LLMs) are becoming increasingly sophisticated "thinkers," their "memory" and efficient access to large amounts of information remains fundamentally underdeveloped. This project proposes the development of graph-based, navigable memory modules that allow AI agents to reason, cross-reference, and explore information with the fluidity of a human expert. Any progress in this direction will be highly beneficial to applications with large corpora of documents, where important information is unstructured and disseminated among multiple places. This is the case, for example, for data in public administrations.
Contact person: Helene Gundhus
Status: Pre-project
PROMESS (Processing the Mess: Critical, Ethical & Legal Exploration of AI Tools in Policing) is a transdisciplinary project responding to today’s overload of messy information in police investigations and intelligence, where information management remains largely manual and time-consuming and computational methods are still limited. Using real police data, the project develops and evaluates machine learning tools to transform unstructured information into structured, searchable knowledge, including graph databases and tools to detect conflicting or missing information. In parallel, PROMESS traces how these technologies are domesticated and co-produced in everyday policing, examines how they affect professional judgement and critical thinking, and assesses their compliance with European legal frameworks, human rights standards and emerging AI regulations. Working closely with user partners such as Kripos, the project aims to bridge the gap between current investigative and intelligence practices and modern computational processing, delivering both practical applications for more efficient and trustworthy information handling, and clear guidelines for the responsible, transparent and accountable use of AI in policing.