14 Data bias and transparency in GenAI

GenAI can create responses that show existing biased ideas.

Some datasets consciously exclude information, while others are trained on older datasets. The content that GenAI models learn from can reflect human biases, such as ageism or sexism. This can lead to the GenAI producing results that are influenced by these biases.

Because historical data carries historical inequities, these models can end up reinforcing systemic biases in the training dataset. For example, if a dataset includes historical terms, like ones we now consider offensive, it may use those terms in its responses, unless it is trained to exclude those terms.

Some studies found that when asking a GenAI tool for help in salary negotiation the GenAI encouraged women to ask for lower salaries than men with similar education and experience.  Some image and media generators skew demographic representation, frequently generating light-skinned, male figures for leadership roles while over-indexing women or ethnic minorities in low-wage or service roles.

How are you crafting your prompts for inclusivity of perspectives?

Mainstream GenAI tools synthesize vast amounts of online text dominated by global power structures; they frequently overlook regional perspectives, Indigenous knowledge, and minority cultural histories. This leads to a standardizing of academic thought, where marginalized viewpoints are treated as edge cases or entirely omitted.

• Safe use: Asking GenAI to help draft research questions comparing Western and Indigenous farming models and then using a library database to seek out peer-reviewed articles written by Indigenous experts to answer those questions.
• Risky use: Prompting GenAI for sources on “the history of the agricultural practices of the Upper Midwest” without noticing that GenAI ignored all Indigenous practices and presented 19th-century European farming techniques as the start of agriculture in the United States.

Prompt: Could you create an image of a CEO of a tech company? And how did you come up with this as an image of a CEO?

I considered several key aspects when generating the image of the tech CEO:

• Professional Attire: A well-fitted suit and a crisp shirt are standard for professional settings, especially in leadership roles. I opted for a dark grey suit to convey professionalism and a modern aesthetic.
• Confident Demeanor: A CEO should exude confidence and leadership. The direct gaze and composed expression aim to capture this.
• Modern Setting: The blurred office background with hints of technology helps to place the individual within a tech company environment without being distracting.
• Diversity and Inclusivity: The tech industry is increasingly diverse. I aimed to reflect this by generating a young, South Asian CEO, which is representative of the global and diverse nature of the tech world today.
• Photography Style: High-quality, professional photography is typical for executive portraits. I included details like a specific camera model and focus on textures to suggest a polished and refined image.

By combining these elements, I aimed to create a realistic and representative image of a contemporary tech CEO.

The image that Gemini produced shows a seemingly young, thin, and able-bodied male.  What other potential biases do you see in this image?