Common Press Workshop, Research analysis & Design ideation
Final Major Project blog 7
Week Commencing 14 October
Common Press Workshop
We ran our second workshop at the Common Press, an LGBTQ-friendly bookshop and café. With the help of three assistants, including Saranya and Srushti who managed photos, videos, and notes, so we could focus on facilitating.
A key theme of the session was “Coming out,” which participants described as the ongoing process of sharing one's gender identity, particularly in new settings or with new people—which can be burdensome and frustrating. The group explored whether AI could assist by communicating this on their behalf and helping individuals explore their gender identity, an idea we aim to incorporate into our design ideation.
Srushti suggested making the toolkit materials more specific to enhance participant engagement. I’d like to revisit the article 'Unboxing the Toolkit' by Shannon Mattern (2021) to explore better design approaches. Teddi noted that participants struggle to see AI as a physical concept. To address this, Greg recommended changing the prompt to "Artificially Intelligent Artefact" to help people physicalise AI.

Coming out happens over and over

People rely on you to educate them

It is difficult to physicalise AI

Coming out happens over and over
Fig. 1 Insights and images from Common Press workshop

Fig. 2 What participants made at the Common Press workshop
Analysis of research
To analyse our survey and workshop findings, we highlighted key words from the survey outputs and free-flow writing activities, then used sticky notes to conduct thematic analysis (see blog 3 for an explanation of thematic analysis). This process revealed that our identified themes were similar to those from our previous thematic analysis. This could indicate we’re uncovering clear patterns in our research, or potential researcher biases leading us to categorise data similarly. To minimise bias, inviting others to analyse the findings with us, may have been beneficial.
However, our insights involved breaking down the "trans journey" theme into sub-themes including physical expression, explaining gender, childhood, and experiences/challenges with AI. We also identified new themes, including concerns about AI stealing jobs, the impact of others' perceptions of gender, positivity around workplace efficiency, and additional examples of AI's impact on the community.

Being Trans is a journey of...

Femininity and masculinity is a paradox

AI is bias, steals jobs and is not trustworthy

Being Trans is a journey of...
Fig. 3 Affinity mapping insights
Design ideation
To begin design ideation, Teddi created storyboards while I explored themes and potential design outputs with real-life examples. We synthesised our ideas into three areas: a framework for building AI for and with the Trans+ community, a Trans language learning model, and thirdly adapting the workshop into a product. Each had pros and cons, but we decided to drop the framework idea, as it targeted AI creators. Although, we have spoken with Trans software developers, we lacked broader research with AI companies and makers, limiting our ability to effectively engage them.

Fig. 4 Teddi's design ideation storyboards

Fig. 5 I synthesised our design ideas into three categories, the blue sticky notes are following feedback in the tutorial with Wan and Greg
Creative Technology Lab and Language Learning Models (LLM)
I contacted the LLM specialist at the Creative Technology Lab to discuss the feasibility of building an LLM specifically for the Trans community based on our research data. From this session, I learned about hosting a LLM locally, available open-source models, incorporating our data, and AI limitations. We envision the LLM as a tool to enhance community conversations and connections, not replace them, emphasising the need for us to design the user experience alongside the tool. Wan also noted the importance of explaining why we are creating a dedicated LLM rather than using general ones, which we should address during onboarding.
Design and testing in LGBTQ Centre Workshop
Final Major Project blog 8
Week Commencing 21 October
Adapting workshops into workbook
To explore turning the workshop into a product, I looked into a pop-out workbook design to blend the prompt from the writing activity with the crafting element in the creative toolkit. During the writing activity participants enjoyed discussing and reflecting on their daily lives. In the creative toolkit activity the crafting element was engaging, but participants struggled to physicalise AI.
We created a low-fi version of the workbook and tested it with Sean, who suggested using more specific crafting materials whilst making the background images provided less prescriptive. We then developed a prototype for testing at the LGBTQ Centre.

Fig. 6 Initial prototype of the workbook
Prompt engineering activity and LLM
Given our second idea is to create an LLM, Teddi suggested we explore prompt engineering in our workshop. Following a discussion with Alex Newson, Teddi devised a scenario where participants would ask AI questions about gender experiences and then collectively assess the responses before posing the same questions to a human. To help us understand the types of questions people would ask AI and shape our Trans+ dataset.
We also spoke to Dolica in the Creative Technology Lab, who found a tool to format data for training open-source LLMs. However, we realised that training the LLM would require a large quantity of data. Furthermore, there are ethical considerations about using people’s data, even in a localised LLM. I am now considering exploring this idea more conceptually to avoid technological constraints, respecting people's data, and encouraging greater creativity.
LGBTQ Centre Workshop
On Friday, we facilitated our LGBTQ Centre workshop. The workbook was well-received, with participants finding it liberating to think about how the world could be. However, I observed areas for improvement, such as adding a bottom to the pop-out to prevent sticking items to the table.
The prompt engineering activity also went well, with participants noting that AI responses were positive and supportive, but they lacked the depth and authenticity of human interactions—some described the AI as a positive Instagram post generator. The AI also struggled with nuanced advice on gender experiences. However, participants indicated they felt more comfortable asking open questions to AI due to the lack of judgment. This suggests a potential role for AI, though I wouldn’t want to promote AI over human connection, especially since group sharing has been so valuable.
Fig. 8 PI.AI is the LLM chatbot we used for the prompt engineering activity. We chose this AI due to it claiming to be emotionally intelligent.

Fig. 8 Images from the workshop helpfully taken by Noor




Fig. 7 final prototype of the workbook and participants creations during the workshop at the LGBTQ Centre
Reading: Building better datasets
I decided to further explore what datasets are made of, Kat Crawford and Will Orr ( 2024) Provide recommendations for creating better datasets. They empathise the importance of clearly communicating the dataset's use case to prevent misuse, which can lead to unintended biases. This aligns with Wan's point about being clear that our LLM is specifically for the trans community, not a universal dataset to train all LLMs. If we create a dataset, we must explain how we used it to train the LLM and its limitations.
Another key takeaway is making the dataset user-centric so that it can be easily understood, critiqued, and adapted by non-experts. This ties into Teddi's idea of involving the community in building and contributing to an LLM in a user-friendly manner.
References
Buschek, C., and Thorp, J. (2022) Models all the way down. Available at: https://knowingmachines.org/models-all-the-way (Accessed: 27/10/2024).
Mattern, S. (Y2021) 'Unboxing the toolkit', Tool Shed, 9 July. Available at: https://tool-shed.org/unboxing-the-toolkit/ (Accessed: 11 April 2024)
​
Orr, W. and Crawford, K., 2024. Building Better Datasets: Seven Recommendations for Responsible Design from Dataset Creators. Journal of Data-Centric Machine Learning Research, 1(1).