Deutsch

Generative AI in Media Contexts

Focus areas:

  • Embodied conversational agents and photorealistic avatars
  • Synthetic media in training and health communication
  • Calibrated trust, acceptance and explainability
  • User-centred evaluation of generative systems
  • Ethical questions in the development of AI for education and care

Generative systems — embodied agents, photorealistic avatars, synthetic voice and video — are arriving in training and health communication faster than the methods needed to judge them. We study how people respond to them: whether an avatar increases engagement or unease, when an explanation helps and when it merely reassures, and what makes trust calibrated rather than blind. The work follows the same measurement tradition as the rest of the lab, combining controlled studies, physiological measures and field deployment — because acceptance measured in a laboratory does not survive contact with a clinic or a classroom unchanged.

Related projects:

Publications

Agents people can stand to talk to

A conversational agent is judged in the first thirty seconds. We study how the personality and communication behaviour of an agent change the way people experience it, and how far personalisation can go before it becomes uncomfortable — including conversational MetaHuman avatars presented at human scale in an immersive environment. What we measure is user experience and trust, not whether the model produced a plausible sentence.

Rehearsing the difficult conversation

Generative characters make it possible to practise situations that are hard to arrange and harder to repeat. We built and evaluated immersive job-interview preparation with photorealistic, AI-driven avatars, and we work on cross-reality support with conversational agents for autistic people. In both, the question is not realism for its own sake but whether the rehearsal helps outside the headset.

Synthetic data where real data must not travel

Medical records cannot be shared for most research purposes, and a method that needs them stops at the clinic door. We generate synthetic time-series medical records with generative adversarial networks so that models can be developed and compared without moving patient data. That work received the Diversity and Societal Impact Award at QoMEX.

What we do not claim

This is the youngest of the five lines and the one where claims are cheapest to make. Everything above is an evaluated system with participants, not a demonstration; where we have not measured something, the pages of the individual publications say so. Generative models are treated here as a component whose effect on people has to be established, in the same way as a display, a codec or a controller.

Selected work