Viktor Kewenig

Researcher at Microsoft Research Cambridge. PhD in Cognitive Neuroscience (Leverhulme Ecological Brain DTP, UCL)

I currently work on Project Provenance at Microsoft Research Cambridge. I am developing a taxonomy of AI edits across image, video and text, building provenance capture at the point of authorship in Word, and leading more than five user studies on provenance UX. The goal is to make provenance clear enough to support calibrated trust and practical enough to fit into real workflows.

Since returning to Microsoft Research in 2025, I have also worked on Evaluating collaboration between human(s) and AI(s), where we fine-tune smaller language models to collaborate more effectively using LLM-as-a-judge evaluation grounded in social-science research on collaboration. Based on previous work, I implemented a metacognitive planning mode in VS Code and ran large-scale experiments on dynamic "productive friction". Our intervention helps users produce higher quality work (while cutting token spend) through metacognitive prompting strategies.

I worked at Microsoft Research Cambridge before, from 2022 to 2024, while I was still doing my PhD. Back then, I co-led a large-scale behavioural study with Microsoft Research New York and Cambridge University Press & Assessment. I was co-first author of The Metacognitive Demands and Opportunities of Generative AI, which received a Best Paper Award at CHI 2024. I also co-led a qualitative study of evolving norms around generative AI among university students.

I generally like ecologically valid research methods. During my PhD at UCL, I worked on how brains and computational models represent meaning in real-world settings. In a comparative fMRI study, we found evidence that semantic processing is distributed and changes with context, rather than being confined to a fixed set of specialised brain regions. Using multimodal and emotion-fine-tuned transformers, I also developed a SOTA method for generating concreteness ratings across languages and expression types.

My undergraduate studies were in logic and the philosophy of science and mind at Cambridge. Naturally, I am a big fan of Wittgenstein.

Music brings an important balance to my life - I curate a podcast series called "5918mins" and sometimes mix records myself under an alias called "No Frills". The best finds are uploaded to my YouTube channel.

For a comprehensive overview of my academic and professional background, you can have a look at my CV.

Current Research

Human-Centred Provenance

Microsoft Research Cambridge, 2025-present

I am developing a taxonomy of AI edits across image, video and text, building point-of-authorship provenance in Word, and leading more than five studies of provenance UX.
Evaluating collaboration between human(s) and AI(s)

Microsoft Research Cambridge, 2025-present

We fine-tune smaller language models to collaborate more effectively and evaluate them with LLM judges informed by social-science research on collaboration.
Metacognitive Planning and Productive Friction

Microsoft Research Cambridge, 2025-present

I implemented a metacognitive planning mode in VS Code and ran large-scale experiments on dynamic "productive friction" in AI-assisted knowledge work.

Publications

Attention, not scale, drives human-AI alignment in multimodal language prediction

Kewenig, V., Lampinen, A., Nastase, S.A., Edwards, C., D'Elascombe, Q., Richardt, A., Skipper, J.I., Vigliocco, G. (2026). npj Artificial Intelligence

We compared five vision-language models with 600 people. Visual context made the models' word predictions more similar to people's predictions, but model size did not; transformer attention was also related to gaze when the video was useful.
From Binary Groundedness to Support Relations: Towards a Reader-Centred Taxonomy for Comprehension of AI Output

Sarkar, A., Poelitz, C., Kewenig, V. (2026). arXiv:2604.08082

The paper argues that grounded or ungrounded is too blunt a distinction. It describes different ways in which an AI-generated statement can be supported by a source, so that provenance interfaces can explain that relationship to readers.
Effects of LLM use and note-taking on reading comprehension and memory: A randomised experiment in secondary schools

Kreijkes, P., Kewenig, V., Kuvalja, M., Lee, M., Hofman, J.M., Vitello, S., Sellen, A., Rintel, S., Goldstein, D.G., Rothschild, D., Tankelevitch, L., Oates, T. (2026). Computers & Education, 243, 105514

We ran a preregistered experiment with 405 students. Taking notes, with or without an LLM, led to better comprehension and retention than using an LLM alone. Most students still preferred the LLM.
A multimodal transformer-based tool for automatic generation of concreteness ratings across languages

Kewenig, V., Skipper, J.I., Vigliocco, G. (2025). Communications Psychology, 3, 100

We built a context-sensitive method for rating how concrete words and phrases are. It combines visual and emotional information, and also worked on an Estonian dataset that the system had not seen before.
The new calculator? Practices, norms, and implications of generative AI in higher education

Simkute, A., Kewenig, V., Sellen, A., Rintel, S., Tankelevitch, L. (2025). arXiv:2501.08864

We interviewed 26 students and 11 educators. Students were using generative AI amid unclear university guidance and a strong focus on plagiarism, so the paper looks at what this meant for their skills and relationships with educators.
Ironies of generative AI: Understanding and mitigating productivity loss in human-AI interaction

Simkute, A., Tankelevitch, L., Kewenig, V., Scott, A.E., Sellen, A., Rintel, S. (2025). International Journal of Human-Computer Interaction, 41(5), 2898-2919

We use earlier research on automation to explain why generative AI can reduce productivity. Problems arise when people are pushed into evaluating output, when their workflow is broken up, or when AI makes the hard parts of a task even harder.
The Metacognitive Demands and Opportunities of Generative AI

Tankelevitch, L., Kewenig, V., Simkute, A., Scott, A.E., Sarkar, A., Sellen, A., Rintel, S. (2024). Proceedings of CHI. Best Paper Award

We argue that working with generative AI places new demands on people's metacognition. The paper considers planning tools and ways for users to check or change what they are doing.
Ecological brain: reframing the study of human behaviour and cognition

Vigliocco, G., Convertino, L., De Felice, S., Gregorians, L., Kewenig, V., Mueller, M.A.E., Veselic, S., Musolesi, M., Hudson-Smith, A., Tyler, N., Flouri, E., Spiers, H.J. (2024). Royal Society Open Science, 11, 240762

This paper asks how cognitive science can study people in realistic settings without giving up experimental control. It proposes a cycle between work in the lab and work in the real world.
Language development beyond the here-and-now: Iconicity and displacement in child-directed communication

Motamedi, Y., Murgiano, M., Grzyb, B., Gu, Y., Kewenig, V., Brieke, R., Donnellan, E., Marshall, C., Wonnacott, E., Perniss, P., Vigliocco, G. (2024). Child Development, 95(5), 1539-1556

Caregivers used iconic vocal and manual cues more often when they talked about unfamiliar or absent objects. These cues may help children connect words to things that are not in front of them.
When abstract becomes concrete: naturalistic encoding of concepts in the brain

Kewenig, V., Vigliocco, G., Skipper, J.I. (2024). eLife, 13, RP91522

In brain data recorded during movie watching, abstract concepts looked more concrete when the visual scene matched their meaning. Concrete concepts looked more abstract when the scene did not match, which showed that conceptual organisation changed with the context.
The entire brain, more or less, is at work: 'Language regions' are artefacts of averaging

Aliko, S., Franch, M., Kewenig, V., Wang, B., Cooper, G., Glotfelty, A., Hayden, B., Small, S.L., Skipper, J.I. (2023). bioRxiv

Our analyses suggest that familiar language regions appear when researchers average across words and people. During film viewing, language relied on changing hubs that connected with many sensorimotor areas.
Do you hear how BIG it is? Iconic Prosody in Child Directed Language Supports Language Acquisition

Kewenig, V., Brieke, R., Gu, Y., Vigliocco, G. (2021). Proceedings of the Annual Meeting of the Cognitive Science Society, 43

Caregivers changed their tone of voice more often when they talked about unknown or absent objects. This may help young children learn words for things they cannot currently see.
Intentionality but not consciousness: reconsidering robot love

Kewenig, V. (2019). AI Love You: Developments in Human-Robot Intimate Relationships, 21-39

This chapter asks whether a person could have a meaningful loving relationship with a robot. It argues that the important property is intentionality rather than consciousness.
Commentary: Robots as intentional agents: Using neuroscientific methods to make robots appear more social

Kewenig, V., Zhou, Y., Fischer, M.H. (2018). Frontiers in Psychology, 9, 1131

This commentary asks how neuroscience could help researchers design and test robots that people see as intentional social agents.

Current Research

Evaluating collaboration between human(s) and AI(s)

Metacognitive Planning and Productive Friction

Publications