Microsoft Designer and DBEI (2024)

My Microsoft Designer Prompt Templates

Some of the noticings I had when entering information in the prompt template were:

• 🚨DIVERSITY: Unless the initial context was designated up front (Smurf, My Little Pony, Female, etc.), for the most part it assumed I was a white male. To me, this was clear evidence ofBIAS in the training data.
• ⏩DIVERSITY: I found the need to modify the prompt template to provide a more representative image of what I was mentally visualizing. As I make my own templates moving forward, I want to intentionally make editable fields forrace, gender, age, ability, body type, and other factors. This forces the user to acknowledge and identify those areas so that the information that they are providing as part of their prompt potentially helps in training future data models to provide more diverse outcomes
• 🚨EQUITY: I had to use my personal account as this was not yet available in my workplace domain. This showed that ACCESS mattered and not everyone may have the same access to AI resources, which could widen the digital divide for students.
• ⏩EQUITY: I appreciate that companies are taking time to make sure that when access is provided to educational institutions, considerations are made forprivacy, consent, and the potential for misuse of users (and protecting students). Moving forward in sharing this tool, I want to explicitly make others aware of this and the benefits and drawbacks of having to use personal accounts when it is not available within educational domains.
• 🚨INCLUSION: Some of the prompts are clearly fictional, but others can be very realistic to either real photographs or digital artwork. The more detailed and descriptive language used in the prompt matters, but STEREOTYPING is often present in AI-generated images. These biases in the training data canperpetuateharmful stereotypes about different groups, and potential for an increase in the generation of DEEPFAKES.
• ⏩INCLUSION:AI excels at creating entirely fictional worlds and characters, expanding the possibilities for diverse representation in creative content. It allows users to think and create beyond their own experiences and world. Moving forward, I want to share about the emphasis that should be placed on descriptive language and providing the large-language model (LLM) with feedback when image generation shows bias or stereotyping.
• 🚨BELONGING:Belonging is a complex and personal feeling that goes beyond what we see visually. AI lacks many aspects that only humans in the loop can provide things such asfeelings, emotions, or personal experiences.This indicates a SUBJECTIVE response, as belonging is not a static feeling or experience and AI cannot actively participate in, or contribute to, a community or culture in the same way humans do.
• ⏩BELONGING: AI-generate images can evoke feelings, an emotional connection, and a sense of seeing oneself in different contexts. As AI struggles to capture the dynamic nature of how individual experiences and a broad and diverse social context change over time, I want to make sure that I am sharing to lean into specific and contextual descriptive language. This not only helps build future data sets, but it can help foster a symbolic representation of humans, culture, and community.

Google Gemini was used to help create the suggestions below (80% of content), but the human in the loop (me) was used for the other 20% of the content!

1. Define Your Goal:

• What are you trying to achieve with your AI prompt?
• Are you trying to generate content that promotes diversity, equity, inclusion, and belonging?
• Are you trying to identify and address bias in your input prompts and data output?
• Are you trying to create a more inclusive environment for yourself and otherusers (if you intend on sharing your prompts as templates)?

2. Use Inclusive Language:

• Avoid using language that is exclusive or discriminatory.
• Use gender-neutral language, and avoid making assumptions about people's backgrounds or identities.
• Use language that is respectful of all cultures and abilities.

3. Focus on Intersectionality:

• Recognize that people have multiple identities, and that these identities can intersect in complex ways.
• Consider how your prompt might affect people from different backgrounds.
• Use language that is inclusive of people with disabilities, people of color, LGBTQ+ people, and people from all walks of life.

4. Be Specific:

• The more specific your prompt is, the more likely it is that the AI will generate content that is relevant to your goals.
• For example, instead of asking the AI to "write a story about a person," you could ask it to "write a story about a YOUNG WOMAN OF COLORwho is a COMPUTER SCIENTIST."

5. Test and Refine:

• Once you have created your prompt, test it out with a variety of different inputs or share with others for them to try through their personal lens.
• See how they respond to the prompt, and make sure that the content that is generated is actually diverse and inclusive.
• Be willing to refine your prompt based on your testing.

Additional Tips:

• Use a diverse dataset to train your AI model.
• Be aware of your own biases, and how they might affect your prompts.
• Get feedback from people who are experts in DEIB.
• Be willing to experiment and try new things and receive constructive feedback on your prompts so that you can produce prompts in the future that are more inclusive (when we know better, we do better!).