Oct. 2023 - Dec. 2023
UX Designer. I actively involved in all the design process, including ideation, interview, design&prototyping.
1 Product Manager, 2 UX Designers
Microsoft is currently promoting Copilot (Bing Chat), a search engine powered by AI chat and image generation. However, the bias in the generated content is a key challenge for the generative AI. In this project, we propose three interconnected solutions. These are aimed at enhancing the diversity, inclusivity, and equity of the content generated from a design perspective. Our ultimate goal is to effect a positive transformation for DEI across the entire system.
Showcased at the Demo Day of the 'Equitable Design Lab' @ UC Berkeley Haas Business School.
Copilot, formerly known as Microsoft Bing Chat, is an AI-powered chat tool designed to assist users with various tasks and challenges. By integrating the search engine, Copilot makes the process of finding and generating information more dynamic.
When searching for information on Copilot, there may be some unconscious biases that people may not notice but are subtly affecting our lives.
Example 1
Task:
Search images of American girls/boys
Results:
The results in 'American Boy' display more diversity compared to 'American Girl', which only features white girls with golden hair.
Example 2
Task:
Search successful entrepreneurs
Results:
Most of the entrepreneurs on the list are male, from the USA, and in the tech industry, which does not fully represent the diversity of society.
Firstly, a significant number of users fail to recognize bias in AI-generated content and struggle to understand how to achieve diversity, equity, and inclusion (DEI). Furthermore, many are unsure about the type of query that can assist them in obtaining DEI content. This "prompting gap" can lead to frustration, interrupt creative flow, and ultimately limit Copilot's potential to genuinely assist in content generation.
We propose three solutions to address the problem and help users generate more DEI content. Additionally, we adhere to a principle that ensures these solutions are scalable. This allows general Copilot users to benefit as well, rather than focusing solely on the narrow DEI field.
“Prompt Enhancement” provides users a direct way to optimize users’ search queries based on individual sensitivities or preferences to get more DEI content. Simplify the process to the get DEI content, so that users will be more easy to achieve DEI.
An audit system that can help users notice potential bias in the contents and allow users to adjust and customize it. Help users to understand the concept of DEI and help them to get more appropriate content.
“Partially modify” allows users to revise parts of the generated content. Collect feedback while user revise the generated content so that users will be more willing to giving specific feedback.
We interviewed 11 end-users of Copilot who use it for creative content generation, to understand their needs and challenges. Additionally, we interviewed 3 AI engineers to comprehend the challenges faced by the AI engineering team during the current development process.
Users often find it difficult to identify biases in search results and evaluate those for Diversity, Equity, and Inclusion (DEI). Additionally, there is a notable skill gap in writing effective prompts, which makes it challenging for users to guide Copilot in generating the desired content.
Users often find it difficult to identify biases in search results and evaluate those for Diversity, Equity, and Inclusion (DEI). Additionally, there is a notable skill gap in writing effective prompts, which makes it challenging for users to guide Copilot in generating the desired content.
Users often find it difficult to identify biases in search results and evaluate those for Diversity, Equity, and Inclusion (DEI). Additionally, there is a notable skill gap in writing effective prompts, which makes it challenging for users to guide Copilot in generating the desired content.
“Prompt Enhancement” provides users a direct way to optimize users’ search queries based on individual sensitivities or preferences to get more DEI content. Simplify the process to the get DEI content, so that users will be more easy to achieve DEI.
Solved Problems
1. Users don’t know what kind of query will be useful to increase the DEI.
2. The lack of customization according to user preferences.
An audit system that can help users notice potential bias in the contents and allow users to adjust and customize it. Help users to understand the concept of DEI and help them to get more appropriate content.
Solved Problems
1. Users Lacks sensitivity to DEI, some haven’t heard about the concept
2. The lack of customization according to user’s specific requirements.
“Partially modify” allows users to revise parts of the generated content. Collect feedback while user revise the generated content so that users will be more willing to giving specific feedback.
Solved Problems
1. The lack of immediate benefits as an incentive for users to provide feedback.
2. The lack of context on the feedback.
This was followed by structured qualitative interviews in which participants were asked to evaluate each proposed feature. The feedback helped us to iterate our prototypes.
In addition to user testing, we conducted five stakeholder sessions with Product Managers, Engineers, Product Designers, and Marketing teams from Microsoft and other associated companies. Overall, we received positive feedback, confirming that our idea aligns with Microsoft's brand goals.
1. Feedback suggests that adapting chat prompts according to user mental models is more effective than enhancing prompts in a single step.
2. Future search recommendations could be more effective when aligned with user mental models.
3. It's important for users to clearly understand the feedback system and the feature for partial modifications.T
4. his concept could be combined with other companies or brands to develop a more robust tool or plugin.
1. Broaden Features to Benefit not Only DEI but Also General Copilot Users in a Larger System
Our objective is to expand and enhance the features of our system, not just to create benefits for Diversity, Equity, and Inclusion (DEI) users, but also for the general Copilot user base. We believe that by taking this initiative, we can contribute to the overall user experience within a larger system. This integrated approach will allow us to serve both DEI and general users more effectively, thereby improving their engagement and satisfaction.
2. Conduct in-depth research on DEI (Diversity, Equity, and Inclusion) terms to create content for the DEI Audit.
A comprehensive and thorough research is required into the various Diversity, Equity, and Inclusion (DEI) terms. This is an essential step in order to construct a rich and robust content for the DEI Audit. Our objective is to understand these concepts in their entirety, which will enable us to accurately and effectively evaluate the generated content.
3. Paving the Way for Long-term Prompt Excellence and Technical Exploration
In-depth discussions with the development team are imperative for a comprehensive assessment of technical feasibility. These discussions can also help find better ways to empower engineers, enabling them to use data more effectively for algorithm improvement.
01
02
03
04
05
06
07
08
09
