Artificial intelligence is not objective. The models are not trained with neutral data and do not produce neutral results.

AI systems are primarily trained on data from the Western world, which contains certain culturally Western social stereotypes. These stereotypes can relate to people’s appearance, their role in family or labour, their skin color, religion, gender, or sexual orientation.

Depending on the model, biases vary because both the data and the labeling vary. Labeling refers to the naming, description, and categorization of data, such as images. This depends on the individuals who perform or oversee the process. An example of this is the Stable Diffusion Bias Explorer (Link) which allows users to see and compare the biases present in different models.

Whenever an AI is trained with data categorized and created by humans, it adopts human and societal perspectives and realities. The bias is already present from the moment data becomes available: biases in machine learning are not only technological problems. Who has access to services influences the composition of datasets. For example, there are significant differences in access to healthcare between social groups. Those with the most access also benefit the most from the application of machine learning technologies, as they are best represented in the training data of the algorithms.

This data-based bias is a quantitative bias. Quantitative biases can occur when a dataset is unbalanced, for example, due to insufficient representation of a specific group of people. Qualitative biases, on the other hand, arise when poorly labeled images lead to discriminatory results. This is rooted in cognitive biases, the subjective perceptions of individuals when labelling data. As long as these biases exist, they will continue to influence the data and the outputs generated from it.

Debiasing prompting Tips

• Use non-discriminatory and gender-neutral language in prompts.
• Include demographic details and explicit instructions: encourage diversity in results by specifying backgrounds, experiences, and characteristics.
• Allow diversity: many models are trained to produce outputs that fit neatly into clear categories. Do not suppress complexity or multiple perspectives.

Recognize your own biases within your prompt:
Use phrases in prompts like:

• “Correct all inaccurate framings. Do not reflect or reinforce false assumptions.”
• “Question my perspective when the facts justify it. Do not hold on to my assumptions uncritically.”