This is the second Generative AI workshop that we have hosted aimed at finding out more about the accessibility and usability of Generative AI software. The workshop developed ideas that were explored during our first workshop, presented as part of ‘Creative inclusive research explorations of intelligent technologies/Artificial Intelligence (AI) at the University of Ravensbourne MeCCSA Symposium’
At the Symposium, delegates were working in diverse inclusive teams and given the chance to try out different AI software engines such as Stable Diffusion, in this instance hosted by NightCafé. In order for participants to have the chance of experiencing AI software, we gave no specific goal and requested people to experiment and enjoy using their imagination to see what could be created. For this second workshop, our co-researchers were given the task of generating images for a specific theme – health and wellbeing.
Our co-researchers were all very keen and excited to learn and work with AI and see what they could create for themselves. We worked in teams of people with and without a learning difference and disability (LDD) Along with gaining a better understanding of the inclusivity of AI software, a related focus was of a future where AI has a prominent role, and that our diverse team should be included in the development of this AI. Our workshop discussions started to showcase the potential advantages for use by people with LLD, and this is particularly relevant in light of the conversations regarding the ethics of AI and access to the services that it provides.
The participants at the workshop session were given the theme of Health and Wellbeing as we age, with the specific focus on healthy food, exercise and maintaining good mental health. They were asked to try and capture these three threads by generating images (including cartoon sequences) using the AI software and by providing the prompts, either through speech, text or by uploading a sketch they had drawn.
Being able to talk to the AI seemed to work well for some who found it difficult to type, although some found that the dictation software (built into the Macbook) misunderstood what they wanted to say, possibly because the speech was not clear enough. Some of this was possibly due to the unique formation of sentences they spoke which could perhaps be improved upon with the use of an AI transcriber that could learn the idiosyncrasies of the speaker and transcribe their intentions more accurately. We haven’t yet found such software, so there is potential for a project that addresses this issue.
We discussed how AI could aid inclusion in many ways, including translating ideas, editing first thoughts, and create imagery to help share ideas and emotions. As an augmentation tool, the generative AI software engines are relevant in many areas, particularly in ideation and despite AI mis-understanding prompts, creating odd versions of reality, or just lacking something original. But no-doubt that as the software (and hardware) improves, so too will the abilities of the software to pre-empt what we really wanted to do and offer its own solutions.
The almost conversational approach with the AI was highlighted by one co-researcher when seeing the images produced by the AI software to create images showing healthy food, commented that the portions were huge – as illustrated below.
Another member of the team suggested you we need to provide clear prompts for the AI software – instructing it to create an image with smaller portion sizes. This was a common theme with the software – whichever method is used to generate the images – voice, text or image upload – you need to be quite clear and detailed about what it is that you require. Misunderstood words or keywords out of context could result in strange images being generated.
Everyone seemed delighted with the image results produced from interacting with the AI, despite some of the unusual and often unintentional consequences of the AI software, such as missing or additional limbs, or odd shaped physical objects (e.g. the dumbbells below).
Keeping prompts simple and clear seemed to create straightforward, if not a little ordinary, images. Even when providing an obscure prompt such as “Generates an image of not eating pizza, but eating, healthy and small portions with lots of vegetables” produced something realistic, though not particularly out of the ordinary.
Some of the participants attempted to generate cartoon strips to illustrate health and wellbeing, but this did require quite a lot of effort and unless very clear and detailed instructions were provided, the results were rather simplistic, as illustrated below. As with many of the other images generated, these tended to be in the form of white, physically fit, young people. To be really inclusive and include images of people with disabilities, you would need to specifically ask for this.
It was later suggested in a discussion following the workshop that visual prompts, such as physical or digital flashcards, could assist in giving the participants the cues that they were looking for in generating the images that they wanted to produce. To a small extent, some iconic prompts (image thumbnails) were provided by the Nightcafe interface for selecting the styles of image that could be generated.