Trans datasets, training an LLM and service journey
Final Major Project blog 09
Week Commencing 28 October
Creating a Trans+ dataset
To create a Trans dataset to train a Language Learning Model (LLM) for the trans community. I asked ChatGPT to generate prompts it thought people in the trans community would ask and compared these with questions asked in our workshop.
While the generated prompts were relevant, the responses were general and impersonal. So, we created a dataset from our own research, formatted as prompts and answers for LLM training. This dataset in contrast was small, unique and personal. I doubt everyone in the community will relate to it, but those who do may find it more meaningful.
This led me to question the value of "big data" versus more focused, smaller data. Is a personalised approach more meaningful even if it reaches fewer people? My thoughts were reinforced by a Wired article titled ‘AI slop is flooding Medium’ (2024) discussing Medium's stance against low-quality AI-generated content.
Fig. 1 Sample of ChatGPT's created prompts and answers
Fig. 2 Our dataset of prompts and answers to put into an LLM based of information shared in workshops and surveys
Training an LLM
In our hope to create an LLM based on small data tailored for the trans community. The Creative Technology Lab provided us with a tool to download an open-source LLM and train it with our own dataset through an easy-to-use interface—no Python knowledge required. It was locally hosted, ensuring our data stayed private. Excited, we uploaded our dataset and awaited the creation of our custom AI. However, instead of producing thought-provoking responses, it returned the same mundane answers.
We then adjusted the AI’s character and user profile which yielded better results, eliminating the need for users to explain their trans identity, but also led to 'hallucinations' where the AI created a pretend person they could put us in touch with. We were realising human shared experiences and connection is difficult to just programme into an AI
Fig. 3 Interactions with an LLM after inputing our dataset and changing the characters of the LLM and the user
Looking for love interview
We interviewed Fast Familiar, creators of Looking for Love exhibition (2023 - 2024), to learn how they set up their exhibition and got an LLM to ask questions, instead of providing answers on love. They explained teaching AI about love and creating meaningful interactions proved challenging, so they ultimately scripted the experience to reflect on the difference between human and AI intelligence and their flaws. Suggesting, if creating an LLM for the trans community is not feasible, we can still craft an experience that illustrates what we envision an AI should do. As designers, we can creatively showcase AI's limitations while demonstrating our vision.
Fig. 4 Image of Looking for Love exhibition (2023 - 2024) in the Science Gallery London. Rachel Briscoe blog about the process of making Looking for Love can be seen here: https://workroom.fastfamiliar.com/looking-for-looking-for-love/
Where does our experience live in the word?
To turn our outputs—the workbook and trans AI—into an experience, we considered their context and interconnection. We discussed situating the experience on a social enterprise website, with the AI highlighting everyday experiences and encouraging people to come to a workshop to explore their gender expression and identity with others. I created a service journey to capture and reference this concept.
Fig. 5 Service journey demonstrating how our LLM tool and the workshop just co-exist together within a social enterprise focused on the Trans+ community
Scripting a chatbot
I created a scripted flow to demonstrate how we would want an LLM to highlight lived experiences and placed it in a chatbot interface for tutorial. Wan suggested reconsidering its AI-like appearance if it wasn’t really AI, she encouraged us to consider how the design could reflect it being a tool that ultimately encourages community building. Wan’s feedback stressed transparency in design and exploring multiple ways to visualise a concept.
Fig. 6 Flow of the script we wanted to input into Chatbot Pom Pom
Fig. 7 Script for chatbot Pom Pom placed in a Figma chatbot prototype by Chaudhari available here: https://www.figma.com/community/file/1218792086352328521/interactive-chatbot
Designing, reflecting and the diminishment of AI
Final Major Project blog 10
Week Commencing 4 November
Prototyping the workbook
Whilst, Teddi focused on developing our scripted digital experience, Pom-Pom, I worked on evolving the workbook shared at the LGBTQ centre. I decided our workbook would focus on the free-flowing writing and creative box activities in our workshops as these were most influential at helping people reflect on their gender expression and identity.
I sketched an idea and consulted Alejandra, who suggested creating a box for the materials and book. Utilising London Centre for Book Art, Making Books (2017), I created a paper prototype for Teddi’s input. I then built a larger prototype to explore size and woking with grey board. Though rough, it helped me visualise the final product. Given our workshops' hands-on approach, a handmade book feels right.
Fig. 8 initial sketches of the book
Fig. 9 initial prototype of workbook
Fig. 10 Prototype two of workbook
Tutorials and reflections
In our tutorials, we presented our workbook and interface prototype. Wan pointed out that integrating our stories into a chatbot interface might compress the richness of our data, highlighting that we had not addressed last week's feedback. If we intend to continue developing a digital element, addressing this feedback I think will help us to move forward.
Regarding the workbook's evolution, Greg shared he missed the theatrical style of the original format but understood the evolution. He also suggested that a consistent design language could create a more cohesive and professional look. Aria also recommended making the box something people would want to place in their home, giving it purpose beyond the workshop.
Alaistair and Tonicha questioned whether it might be time to reduce the role of AI in our project. Alaistair suggested that if AI hasn't met our goals, there may be little value in showcasing its limitations through creating Pom Pom. Instead, he proposed using what we’ve learned to enhance the workshop process. As our workshops suggests, we can use facilitation, communication and physical artefacts to provide the support to one another that we are looking for in technology.
This perspective was helpful, as we have focused on making AI fit our project due to its growing role in our lives. However, Alaistair’s insight reminded me that our workshops have been central to our research, creating joy and building community. We may have overlooked their significance by concentrating on AI. As our early interview with Feminist Internet highlighted, creating safe spaces for marginalised people was far more important that the outcome Syb. Suggesting that enhancing our workshops for the Trans community might be more valuable than developing a simulated AI.
Designing the workbook pages & evolving the pop out pages
Finally, we created different alternatives for the pop out structure and tested it in class. I went on to create the pages for the workbook. Next week I’d like to turn the workbook into a finalised design.
Fig. 11 Overview of pages I designed for the pop out journal
Fig. 12 images from left to right show the evolution of the pop up space for participants to create their ideal world where they can express their gender freely.
Letter press workshop
I attended a Letter Press workshop to explore handmade printing methods for the workbook. The session demonstrated a connection between letter pressing and digital word documents. It highlighted how digital elements, such as spaces, lines, columns, and letters, have physical origins. This made me reflect on how often the physicality of the digital world is overlooked due to efforts to streamline user experience.
Revisiting Kat Crawford’s view of AI as a tangible creation, I questioned whether technology has become so detached from human involvement that we forget it is not magically produced - just as word documents have roots in letter pressing, chatbots are physically made by computers, servers and peoples real data.
References
Biscoe, R. (2024) 'looking for 'looking for love', Fast Familiar, 31 August. Available at: https://workroom.fastfamiliar.com/looking-for-looking-for-love/ (Accessed: 1 December 2024).
Crawford, K. (2021) Atlas of AI. London: Yale University Press.
​
Knibbs, K. (2024) 'AI slop is flooding Medium', Wired, (28 October), Available at: https://www.wired.com/story/ai-generated-medium-posts-content-moderation/?utm_source=substack&utm_medium=email (accessed 3 November 2024)
​
Looking for Love (2023 - 2024) [Exhibition]. Science Gallery London. 21 June 2023 - 20 January 2024. Available at: https://london.sciencegallery.com/ai-artworks/looking-for-love (Accessed: 1 December 2024).
​
London Centre for Book Arts (2017) Making books. London: Pavilion Books Company.
​