We use this blog to highlight some of the work and activities of our CDT students. In this post, Sarah Immel, Neel Rajani, and Rayo Verweij share their work from the Applied Interdisciplinary Project course where they developed an interactive experience for people to explore LLMs, their workings, and their problems.
Would you trust an AI to represent your local community? To summarise the local debate around gentrification, or describe the history of the neighbourhood pub? As automated algorithmic decision-making is increasingly being integrated into public and private spaces, it’s becoming even more important to centre discussions and interrogations of these technologies in the communities and places they affect.
A curious feature of modern AI systems is that as they are comprised of so many individual parts – in the billions – we usually don’t actually know how they arrive at the results they produce. The field of Explainable AI (or XAI) focusses on looking ‘inside’ these algorithms, trying to find ways to make them more transparent. Traditionally, XAI has mostly focussed on the algorithms themselves. However, the human decision-making processes that accompany the implementation of AI systems in the real world – when should we defer to algorithms, who is accountable for their answers – are at least as crucial to these systems’ effects on humans and communities in their daily lives. Unfortunately, they are often just as opaque as the algorithms themselves.
We wanted to prototype an experience that could allow anyone, regardless of their knowledge about AI, to be a meaningful participant in the conversation around deployment of AI systems. Our supervisors for this project were Jingjie Li, Bettina Nissen, Luis Soares, and Alex Taylor. Some of them had already been involved in a previous project researching community attitudes towards the proliferation of hidden data gathering and algorithmic systems in public spaces. As one of their methods, they had put a large black box on stilts on Leith Walk that passers-by could peek into to see provocative questions about the research topic, inviting them to write their responses on the side. We took this black box and converted it into a fully interactive, two-player game about the insides and impact of AI.
The game centres around “EdinBot”, a chatbot that presents itself as an expert on the city of Edinburgh. Player 1 can use a tablet mounted on the outside of the box to ask it any question, just like one would with a system like ChatGPT. Player 2, meanwhile, is invited to step inside the box, where they are asked to influence the response that the chatbot is writing. The way large language models (LLMs) construct answers is through prediction: a few most likely options for each single word is predicted in sequence. Depending on how “creative” the model is told to be, it randomly chooses from a small set of likely words, for a more deterministic model, or a large set of less likely words, for more unpredictable output. This “level of creativity” is commonly called the model’s temperature and this is the role we ask Player 2 to take. Every few tokens, the model pauses its answer generation and presents a few options to Player 2, who can choose which token they want the model to continue with. This way, the player on the inside of the box can steer the model’s response in a direction of their choosing, for example to more or less accurate responses to questions from the outside player.
To get as wide an audience as possible, we set up the game alongside some public events in the Edinburgh Futures Institute as well as in the Informatics Forum atrium. After playing the game, we asked participants to reflect on their experience. What would they (not) trust EdinBot with? What did Edinburgh look like from inside the box? While some said they could trust EdinBot to give high-level tourist information, many decried its generic or sometimes outright wrong answers. Encouragingly for us, though, many participants reflected positively on the ability of the game to allow its players to learn more about the workings of LLMs, and afterwards construct informed opinions about their level of trust in the system.
This project wouldn’t have been possible without the wisdom and support of our supervisors or the variety of perspectives, backgrounds, and skillsets present in the group. And, most importantly, it wouldn’t have been able to achieve anything without the curious, critical, and insightful comments and contributions from our participants. The conversations we had with participants and community members around this project give us even more reason to believe that the deployment of AI systems can, and should, be a conversation, and that researchers should continue to find ways to bridge knowledge and communication gaps between technical and local experts.