My objective in this secondment was to initiate some practical experiments in ‘ethnographic AI’, which I define as an investigation into whether artificial intelligence can gain specific cultural understandings. Engaging with Festival dei Popoli’s ethnographic film archive online and on-site, I used ml5.js to create experiments in AI image classification, image and video generation, pose tracking, and object identification.
I also included experiments in creating randomised compositions, using the Festival’s archive of posters and programme leaflets. Alan Turing, in his 1950 paper ‘Computing machinery and intelligence’, proposes that randomness might be an important element of creativity.
In using the coding environment ml5.js, I uses small, portable datasets for two reasons: (1) to explore whether small, highly curated datasets could facilitate specific cultural understandings; and (2) to mitigate the high environmental cost of AI by minimising processing and network operations.
The work raises questions about AI in terms of identifying, classifying and describing humans in terms of what and who is seen, and what can be inferred from individual and collective actions. Also, the training datasets used in many existing AI systems (including the ones I used) are very limited and culturally-specific, for example in the objects they are capable of recognising. This relates to the programmer’s expression ‘Garbage In Garbage Out’ which describes how the outputs of a system depend on the quality of its inputs. I found that the ‘seams’ in AI systems – for example the unnatural joins between surfaces in images, or the artificial ways Large Language Models attempt to replicate human writing – expose the artificiality of such systems, yet perhaps keeping these seams visible is important for maintaining transparent AI systems.
I made all my code freely available – the code and the experiments can be viewed at https://ai.postdigitalcultures.org
Kevin Walker