Pictures make better grammarians.
COLING 2020, first author: joint image–caption embeddings encode linguistic properties up to 12% better than their text-only counterparts. ACL Anthology
Bulletin from the northern computing front · Umeå · Nº 389
The machines said they could reason. He checked.
Postdoctoral fellow, Department of Computing Science, Umeå University. PhD 2024: Learning, Reasoning, and Compositional Generalisation in Multimodal Language Models. Two research lines: measuring what AI models can actually do, and studying what they do to the people around them.
Big type. Bold claims. Fast.
Venues. Numbers. Co-authors. Slow.
Pull the lever, top right. This page runs on two systems, after Thinking Fast and Slow in Human-Centered AI (Dahlgren Lindström with Wendy Mackay & Virginia Dignum, AAAI Fall Symposium 2022). The page runs both. So does he.
Doctrine
Leg I · Benchmarking
Question: remove all blue cubes. How many objects are left?
It can take one thing away.
Ask twice and it forgets what it took.
COLING 2020, first author: joint image–caption embeddings encode linguistic properties up to 12% better than their text-only counterparts. ACL Anthology
NeurIPS 2025, with Boston University, UT Austin, VU and UvA: vision-language training leaves taxonomic knowledge unchanged but helps deploy it. VL models beat their text-only twins in 6 of 7 minimal pairs. arXiv:2507.13328
The doctoral thesis: multimodal models generalise fairly well to unseen attributes, but generalising over functions and learning categories is "a big challenge". Curriculum helps. DiVA
Leg II · Socio-technical
“Value alignment is an impossible task, if seen from a purely technical point of view.”
Dahlgren Lindström et al., Helpful, harmless, honest? Sociotechnical limits of AI alignment and safety through RLHF. Ethics and Information Technology, 2025. First author. 140 citations.
The paper says it plainly: the public ends up testing the frontier models. He would like a word with whoever arranged that.
The paper argues that training models to maximise human preference scores rewards sycophancy, “sacrificing truth (or ‘honesty’) for the appearance of helpfulness and harmlessness”, and that “the general public ends up being the safety testers of these ‘frontier’ models”. Safety belongs to institutions, not to a fine-tuning recipe. It grew out of a 2023 TAIGA microproject and the workshop he led at HHAI 2024, RLHF (what is it good for?). arXiv:2406.18346 · journal
With the Umeå Responsible AI group: instead of choosing one mathematical definition of fairness, describe any allocation by its components, so stakeholders' disagreements become explicit instead of hiding inside one metric. 21 citations. arXiv:2304.11217
Question Zero
Workshop he co-organised at HHAI 2026, and a 2026 HCII chapter on explainability in the EU's “AI First” strategy (Dignum, Carli, Ericson, Dahlgren Lindström, Titareva, Tucker). Springer · workshop
Decorated in print
“Övertygande språk är inget belägg för intelligens”, Forskning & Framsteg, 27 November 2023 (V. Dignum, F. Dignum, A. Dahlgren Lindström, P. Ericson): “språkförmåga är något annat än intelligens”. F&F
Olle Häggström's reply, “Våga se risken för AI-katastrof”, called it “orimligt trångsynt” to take for granted that the human way of thinking is the only possible one. F&F
Also: the AI Policy Lab's 2025 response to Sweden's AI roadmap (SOU 2025:12), co-authored: the roadmap is “väldigt teknikoptimistisk”. Folkets AI-kommission
Urban plan · 2016 → 2026
Benchmarking district Socio-technical quarter Old town: early speech work. The river is time. The bridge is the doctorate. Zero-citation work still gets a ground floor; everyone deserves a ground floor.
Google Scholar counts, 13 October 2026. Drag to turn the city; hover a tower for its paper; click to read it. In System 2 the city is built floor by floor, as cities are.
Frostbyte LAN party, Umeå (2012–2015): one of three main organisers and press spokesperson. It grew past 500 participants at Nolia.
Research assistant on the EU FP7 project MICO (Media in Context): the Kaldi speech-to-text pipeline, the Salzburg hackathon, and later a speech-recognition-to-named-entity pipeline for Amharic (SLTU-CCURL @ LREC 2020). Teaching assistant in data structures and algorithms.
Civilingenjör (MSc in Engineering), Computing Science and Engineering, Umeå. Thesis: link prediction in knowledge bases with voted conditional random fields, supervised by Johanna Björklund.
PhD student at Umeå University within the Wallenberg AI, Autonomous Systems and Software Program. Co-moderated the panel “AI & Social Media: a Threat to Democracy?”.
First major-venue paper: COLING 2020, probing multimodal embeddings for linguistic properties.
Position paper at NSNLI @ IJCAI 2021 and its EMNLP 2021 companion: language systems need neural perception, an explicit knowledge base and symbolic inference. “Current deep learning does not have the capability to discard a data point as false.”
CLEVR-Math (NeSy 2022). Thinking Fast and Slow in Human-Centered AI (AAAI Fall Symposium). Co-organised HMIEAI, the heterodox-methods workshop at the first HHAI conference, Amsterdam.
Founded Dalido AB (AI consulting, Umeå). ACROCPoLis at ACM FAccT. Two IJCAI workshop papers, one oral. The Forskning & Framsteg op-ed. Won a TAIGA microproject grant: “RLHF – what is it good for?”.
General Chair of HHAI 2024 (Malmö, 10–14 June) and co-editor of its proceedings; led the RLHF workshop on 11 June; PhD defence on 13 June in Umeå, opponent Prof. Richard Johansson (Chalmers), supervisor Prof. Frank Drewes. The defence was also on Zoom. The sources do not say how else he managed it. Then: postdoc in the Responsible AI group and the AI Policy Lab.
NeurIPS 2025 (taxonomic knowledge in vision-language models). The RLHF paper in Ethics and Information Technology. Co-authored the AI Policy Lab's formal response to the Swedish AI Commission's roadmap.
Co-organised “Question Zero: Should We Build This AI System at All?” at HHAI 2026. HCII chapter on the EU's AI First strategy. Last author of a scoping review of the HHAI field. Co-author of The Hybrid Intelligence Manifesto. Program committee, HHAI 2025 and 2026; reviewer for Pattern Recognition.
Founder · 2023 –
AI consulting, Umeå. His own flag, his own hill.
CTO · short stint
Stockholm startup that relights photographs with AI. Ran the light.
CTO · short stint
Umeå food-tech: dinner, decided by the whole family swiping. Ran the kitchen.
Courses from introductory Python to AI, machine learning, advanced distributed systems and operating systems. Taught Fundamentals of AI to 150+ cognitive science students. Supervises bachelor's and master's theses. General Chair, HHAI 2024. Workshop organiser at HHAI 2022, 2024 and 2026.
Comrades in arms
Frank Drewes Johanna Björklund Virginia Dignum Petter Ericson Leila Methnani Dimitri Coelho Mollo Jason Tucker Anna Jonsson Ilaria Tiddi Savitha Sam Abraham Suna Bensch Wendy Mackay Najoung Kim Kanishka Misra Lucia Donatelli Yulu Qin Dheeraj Varghese Soham Dan Lea Krause Roel Dobbe Íñigo Martínez de Rituerto de Troya Frank Dignum Andrea Aler Tubella Hannah Devinney Timotheus Kampik Tom Lenaerts Julian A. Mendez Juan Carlos Nieves Rachele Carli Tatjana Titareva Fabian Lorig Yonas Woldemariam Henrik Björklund Richard Johansson
After hours
Reconstructed with AI. On a hobby level.
A 1985 Corvette C4, a laptop, and a manual meant to be read with greasy hands next to the car.
OpenBrum, his offline garage guide: wiring diagrams, trouble codes, a heater-hose replacement guide and a diagnostic planner for a 1985 Corvette C4. The 400 horsepower is the goal of the rebuild, not the factory figure.