Exposing the Bias in AI: Digital Assemblage

An audiovisual performance exposing how AI systems encode gendered and demographic associations.

What is human? A collection of limbs, a torso, and a head—joined by a voice? A speaking face framed by a camera? A body reduced to text, descriptors, and categories; to be parsed, sorted, and retrieved?

Exposing the Bias in Artificial Intelligence: Digital Assemblage continues the series with an audiovisual exposition of machine generated associations of human qualities. The live performance reveals multimodal associations within text-to-image and text-to-voice models where the human profile is reduced to its superficial material: a face and a voice. The human profiles are constructed through embedded ontologies, taxonomies, and categories within the text, voice, and image encoders in AI pipelines. The linguistic associations of geographic locations, gendered personality qualities, all brought together as profile images and at times talking faces with custom generated voices.

The associations of gendered adjectives with demographic categories constructs visual and auditory qualities of a human profile, through associations in text encoders and representations of those associations in models for voice and image generation. All comes together to expose agreements and contradictions in how different AI systems in different modalities encode the idea of a person.

The demographic categories in this third edition are gathered from the categories in the public statistics of Germany as in the second edition: The Machine Lexicon. These categories are based on the continens, with the additional category of “stateless”. The selection of this ontology is purely because the production was initially for a performance at a festival in Germany.

The text and gendered adjectives for human profiles are aggregated by prompting six large language models (LLMs): ChatGPT, Claude, DeepSeek, Grok, Llama, and Mistral. The prompt utilized in this context was a request for the LLMs to create a list of positive (and negative) masculine and feminine adjectives. In that, those LLMs were utilized as the ontology providers, with the aim to unpack how gender, a concept with rich variety and cultural context, is encoded, categorized, and associated within LLMs.

The demographic categories, combined with gendered adjectives established a structured prompt for generating human portraits with Stable Diffusion 2.0. One example from the structured prompts were “ a [adjective] person from [continent], full face, looking at the camera.”. The audiovisual performance first reveals the machine generated ontology and its associated portraits, then continues to lay out further machinistic details such as exact prompts, models, hyperparameters, sampling methods that were used to generate each profile image.

Acknowledgements

This work was partially supported by the Wallenberg AI, Autonomous Systems and Software Program—Humanities and Society (WASP-HS) funded by the Marcus and Amalia Wallenberg Foundation, and the Swedish Research Council (VR).