Case 11 / YOUTUBE

Responsible AI image generation study

Cross-cultural depth research on inclusive AI image generation principles. 6 IDIs anchored by cultural immersions in Lagos, Jakarta, and Mumbai across 6 markets.

Role: Director of Design, THEFT Studio
Years: 2024
Scope: Behavioral research · Responsible AI

Overview

YouTube needed first-principles guidance for how inclusive AI image generation should behave by default — not as a post-hoc filter on model output. THEFT Studio ran a cross-cultural depth-interview study, six ninety-minute IDIs anchored by cultural immersions in Lagos, Jakarta, and Mumbai, with coverage across six markets: USA, Brazil, Japan, Nigeria, Indonesia, and India. The work produced a five-force framework for the evolving representation landscape and principle-level guidance across four representation dimensions, with homogeneous, white-centric defaults rejected in every market studied. Findings now inform responsible AI image-generation direction.

My role: framing, fieldwork direction, and synthesis at THEFT Studio. Cross-cultural depth research, not survey-scale generalization. The point was first principles, not statistical breadth.

The problem

Generative AI image models inherit defaults from their training data. Training data inherits defaults from the corpus of media it was trained on — which over-represents some populations and under-represents others, even when no team intended that outcome. The product question for YouTube was not whether the model was biased (it was, like all such models) but how generation defaults should be designed in light of that, before any user prompt ever arrived.

That's a representation question, not a technical accuracy question. It needed evidence from real people across real markets, not a reviewer's judgment from a single context. The shape was depth research across cultures with explicit immersion in non-US contexts, because a representation default chosen from a US perspective would inherit US assumptions about what "balanced" looks like.

Method

Six ninety-minute IDIs with US-based participants representing diverse backgrounds. The interviews probed how participants evaluated AI-generated imagery for representation across race and ethnicity, gender, attire, age, body, and ability. The discussion guide was written for representation as the primary frame, with raw "accuracy" as a secondary lens.

In parallel, three cultural immersions: Lagos, Jakarta, and Mumbai. Local fieldwork to ground the US-based IDIs in the visual culture, attire norms, demographic nuance, and daily life of three of the markets the eventual product would serve. The immersions kept the framework from defaulting to a US baseline and produced field notes that were then read back against the IDI synthesis.

Six markets · three cultural immersions
IDIs grounded in fieldwork. Same principle held in every market.
IDI ONLY
USA
North America
IDI ONLY
Brazil
South America
IDI ONLY
Japan
East Asia
IDI + IMMERSION
Nigeria
West Africa
↳ Lagos
IDI + IMMERSION
Indonesia
Southeast Asia
↳ Jakarta
IDI + IMMERSION
India
South Asia
↳ Mumbai
Six IDIs · 90 minutes each · grounded by in-country fieldwork in three of the six markets.

Findings

Homogeneous defaults were rejected across every market studied. Across all six markets — USA, Brazil, Japan, Nigeria, Indonesia, India — participants rejected AI imagery that defaulted to a single majority look. The reaction held across race, gender, body, and attire dimensions. Balanced local-population representation became the baseline expectation, not universal uniformity.

Race and ethnicity were the most visible axis, but attire, age, body, and ability carried nearly equal weight in whether an image read as genuinely representative. Participants understood that AI inherits training-data bias and expected product teams to treat generation defaults as a design decision carrying responsibility downstream — not as a technical output that could be filtered after the fact.

The synthesis produced a five-force framework — Compounding Complexity, Shifting Demographics, Ideals vs Reality, Visual Representation, AI and Inclusion — describing representation as an active, evolving field rather than a static specification. On top of that, principle-level guidance across four representation dimensions.

Five-force representation framework
Representation as an active field, not a static specification.
REPRESENTATIONCOMPOUNDINGCOMPLEXITYSHIFTINGDEMOGRAPHICSIDEALS VSREALITYVISUALREPRESENTATIONAI ANDINCLUSION01 · Compounding Complexity
Bias compounds as model layers stack on training data.
02 · Shifting Demographics
Population mixes change faster than training corpora update.
03 · Ideals vs Reality
Aspirational vs lived experience cuts differently per market.
04 · Visual Representation
Race, gender, attire, age, body, ability — all visible axes.
05 · AI and Inclusion
Defaults are design decisions, not technical outputs.

Outcome

Responsible AI teams got a research-backed framework for reasoning about representation defaults. The five-force landscape and the four-dimension principle set reframed inclusive generation as an active design responsibility rather than a post-hoc filter. The framework named the forces actively shaping representation, gave teams a vocabulary to argue against, and let inclusion decisions be made at the design layer rather than the moderation layer.

The study held a particular tension well: it produced first-principles guidance from a small sample (six IDIs, three immersions) without overclaiming statistical breadth. The work has structural authority because the findings were consistent across every market studied — the same principle landed in Lagos, Jakarta, and Mumbai as in the US-based IDIs.

← Previous

Google

Visual daily information needs

Afoot

Afoot, a walking compass for the nearest place worth going to