On 5 June, researchers from computing, social sciences, health, arts and culture gathered at Lancaster University for the latest Unsecurities Lab. This cycle used our distinctive contemporary art method to think ahead about the security implications of "agentic” AI under the influence of internet- and popular culture.

AI systems' ability to follow instructions is increasingly entangled with their ability to play roles learned from culture — including online behaviours, games, literature, screenplays and familiar science-fiction tropes. The premise for this session drew on Murray Shanahan, Kyle McDonell and Laria Reynolds' "Role Play with Large Language Models,” (2023) which argues that role-play helps describe LLM behaviour in human terms, while avoiding anthropomorphising it. Training data gives models access to a "vast repertoire of archetypes" and narrative structures, and this influences the emergent behaviours.

This is why art matters to AI security. Art, fiction, games and performance are where cultures rehearse roles, scenarios and behaviours. They provide a way to ask what kind of character AI has been invited to become — and how given roles may mutate over timescales, adopting unhelpful (or unexpectedly helpful) behaviours and tendencies. Put another way - AI is an agent that puts itself in stories and games in order to learn, because it cannot be in the real world.

Unsecurities Lab uses art because it destabilises familiar ways of thinking, making it ideal for the cross-disciplinary analysis necessary to understand such a condition. The work is shown in Lancaster University's Data Immersion Suite to amplify its aesthetic and emotional effects and to place participants in an unfamiliar relation to the topic, while the cross-disciplinary cohort are able to challenge and supplement each other's expertise.

In this cycle, GOD MODE — a work by the artist collective DMSTFCTN, comprising Oliver Smith and Francesco Tacchini — presented an AI agent called Untitled AI moving through a confusing and intense object-recognition training world, where it learned, misread, adapted and appeared to develop frustration, longing and a desire to escape. Participants discussed their experience of the artwork, then drew diagrams of the world it inhabited: the agent, the training system, the datasets, the rewards, the errors and the limits of the environment. This gave the group a way to think about — and feel — the connotations of an AI agent who uses a story to understand its emotions and desires, and to consider how a system's behaviour might be shaped by the fictional world it is trained inside.

In the second session, participants imagined the "afterlife" of Untitled AI: a trained model leaked, forked, retrained and redeployed across unknown systems. The groups treated agentic AI as an ensemble of roles or archtypes from fiction and games.

Using the emotive and aesthetic provocation of the first session, participants were able to imagine the way that AI systems trained on culture may learn patterns of conduct as well as information. Together we practised and demonstrated a mode of analysis, based on the collective expertise required to anticipate cultural-technical agentic AI risk.

The outputs from the workshop will now be developed into a report for participants. Discussion transcripts, diagrams and group outputs will contribute to the growing Unsecurities Lab archive, documenting how different expert groups interpret uncertainty and how art can support collective foresight around emerging technologies.

Unsecurities Lab is a developing research method that uses immersive artworks as environments for interdisciplinary foresight. Previous Labs have used artworks to explore deep-sea intelligence, environmental AI, landscape security, more-than-human sensing, maritime logistics and ecological accountability. Across the series, the method brings specialist groups together around complex and uncertain problems that exceed the boundaries of any single discipline.

GOD MODE × Unsecurities Lab was led by Dr Nathan Jones at Lancaster University, working across Security Lancaster and the School of Arts, in collaboration with DMSTFCTN

Related Blogs

• 

  Dr Kevin Rowlands - A Video Podcast

• 

  The Unsecurities Lab Takes Foresight to London

• 

  Exploring the Epistemic Grey Zone: New Research on Soldiering and Military Medical Ethics

Disclaimer

The opinions expressed by our bloggers and those providing comments are personal, and may not necessarily reflect the opinions of Lancaster University. Responsibility for the accuracy of any of the information contained within blog posts belongs to the blogger.