01 / THE DISAPPEARING IDENTITY
An exploration of identity, representation, and the unexpected transformations of generative AI.
Enter the documentary
02 / THE TRANSFORMATION
I created Bernadette to help me return to the workforce after a brain injury.
As I learned to use artificial intelligence, Bernadette became my way of exploring new opportunities, developing skills, and rebuilding my professional future.
I had carefully chosen her appearance. She was meant to be one recognizable woman, with a consistent identity across the images and videos I created.
But something unexpected happened.
The more I worked with AI, the less Bernadette looked like Bernadette.
Her facial features, proportions, and sometimes her skin tone changed. Some differences were subtle. Others were so significant that she appeared to be an entirely different person.
I spent hundreds of generations trying to maintain her appearance, correcting individual images without realizing how far she had drifted from the original.
The problem persisted.
And eventually, I began asking a different question.
That question became Bias Mirror.
One identity. Hundreds of generations. A question that extends far beyond one character.
03 / THE INVISIBLE BIAS
Artificial intelligence is often described as objective. But AI systems are built using data, decisions, and assumptions shaped by human society.
When that information contains stereotypes, unequal representation, or historical discrimination, AI can reproduce and sometimes amplify those patterns.
The consequences extend far beyond a generated image.
When technology doesn't see everyone equally.
Research has documented substantial differences in the accuracy of facial-analysis systems across skin-tone and gender groups. These findings raise important questions about fairness and representation.
View researchWhen yesterday's stereotypes shape tomorrow's possibilities.
AI-generated images can reinforce occupational stereotypes, underrepresenting women in traditionally male-dominated professions.
View researchWhen experience becomes invisible.
Research shows that online media and language models can reinforce associations between women and youth, with implications for professional evaluations.
View researchTHE EVIDENCE / PUBLISHED RESEARCH
Joy Buolamwini and Timnit Gebru tested three commercial facial gender-classification systems. Error rates reached 34.7% for darker-skinned women, compared with a maximum of 0.8% for lighter-skinned men. This was facial analysis, not image generation.
The study examined DALL·E 2 images across 153 occupations. Women were underrepresented in male-dominated professions and overrepresented in female-dominated professions. The findings concern the image generator and sample studied, not every AI system.
Researchers examined nearly 1.4 million online images and videos and nine language models, finding women represented or associated with younger ages than men. The language-model analysis primarily examined embeddings. Separate ChatGPT audits found age-related gender bias in fictional résumé generation and evaluations. This was not a test of image generators or actual hiring outcomes.
Different technologies, not a claim about every AI system. These studies do not establish the cause of Bernadette's changing appearance.
THE HUMAN CONSEQUENCES
When AI influences how people are represented, described, or evaluated, the consequences can reach beyond the screen.
Bernadette's changing face was where my questions began.
Bias Mirror is where those questions become an investigation.
Who is represented? Who is overlooked?
And who gets to decide what is accurate?