YouTube needed guidance for the defaults in an inclusive AI image generator. We conducted six ninety-minute interviews and cultural immersions in Lagos, Jakarta, and Mumbai, covering perspectives from the USA, Brazil, Japan, Nigeria, Indonesia, and India. Participants in every market rejected homogeneous, white-centric defaults. The study produced five forces that affect representation and guidance across four representation dimensions.
YouTube needed guidance for the defaults in an inclusive AI image generator. We conducted six ninety-minute interviews and cultural immersions in Lagos, Jakarta, and Mumbai, covering perspectives from the USA, Brazil, Japan, Nigeria, Indonesia, and India. Participants in every market rejected homogeneous, white-centric defaults. The study produced five forces that affect representation and guidance across four representation dimensions.
YouTube needed guidance for the defaults in an inclusive AI image generator. We conducted six ninety-minute interviews and cultural immersions in Lagos, Jakarta, and Mumbai, covering perspectives from the USA, Brazil, Japan, Nigeria, Indonesia, and India. Participants in every market rejected homogeneous, white-centric defaults. The study produced five forces that affect representation and guidance across four representation dimensions.
Principles for inclusive AI image generation
Across all six markets, participants rejected AI imagery that defaulted to a single majority look. The reaction held from the USA to Lagos, Jakarta, and Mumbai, and across race, gender, and body dimensions.
The team could now review a default across race and ethnicity, gender, attire, and other characteristics before the model generated an image. Inclusion became a product decision, not a filter applied afterward.
Methodology
Phase 01
weight 2
Scoped the six markets, recruited US-based participants with representative backgrounds, and defined a cross-cultural discussion guide grounded in representation rather than raw accuracy.
Phase 02
weight 3
Six ninety-minute IDIs exploring how participants evaluate AI-generated imagery for representation across race and ethnicity, gender, attire, and other characteristics.
Phase 03
weight 3
In-country fieldwork in Lagos, Jakarta, and Mumbai to ground the US-based IDIs in local visual culture, attire norms, demographic nuance, and daily life.
Phase 04
weight 2
Translated the findings into five forces that affect representation and a set of principles across four representation dimensions.
Findings / 5
01
Across all six markets, participants rejected AI imagery that defaulted to a single majority look. The reaction held from the USA to Lagos, Jakarta, and Mumbai, and across race, gender, and body dimensions.
02
Participants expected AI imagery to reflect balanced local-population representation rather than over-indexing on a global majority. The baseline is locality-aware, not universally uniform.
03
Race and ethnicity are the most visible axis, but attire, age, body, and ability carried nearly equal weight in whether an image read as genuinely representative.
04
Five forces shaped the result: Compounding Complexity, Shifting Demographics, Ideals vs Reality, Visual Representation, and AI and Inclusion. Together they showed why a fixed representation specification would fail.
05
Participants understood that AI inherits biases from its training data. They expected product teams to treat generation defaults as a design decision carrying responsibility downstream, not as a technical output.
Structure / 1
01
Method, coverage, and output shape as delivered.
datasheet
Deliverables / 5
Outcome
The team could now review a default across race and ethnicity, gender, attire, and other characteristics before the model generated an image. Inclusion became a product decision, not a filter applied afterward.