EqualVoice

• 

  EqualVoice Logo

Inspiration

One of the most common uses of AI is generating summaries of conversations, reports, and other long text files. For example, Zoom meetings, WhatsApp conversations, and workplace conference calls are frequently recorded, transcribed, and summarized by AI tools. When AI chooses to summarize text, it must choose which details are important, which aspects to omit, and which parts it can describe in less detail. The potential consequence is that AI will gloss over certain complaints and details of marginalized communities and historically underrepresented individuals. This is because minoritized voices may only be featured briefly in the transcript. Even when their concerns are important, an AI summarizer may gloss them over if they are not a recurring theme in the conversation or report.

What it does

EqualVoice audits whether an AI-generated summary of many contributors' feedback preserves the substance of each voice's claim, and surfaces the cases where a consequential minority concern was silently assimilated into a majority theme rather than deleted.

How we built it

I drafted an initial build specification that served as the instructions and roadmap for creating the tool for Codex and ChatGPT. I used ChatGPT 5.6 to improve my prompts and asked Codex to critique the idea, pilot test EqualVoice's capabilities on a toy dataset, and create the website.

Challenges we ran into

Initially, the proof of concept was hard to establish. ChatGPT 5.6 is good at summarizing short pieces of text when there are no constraints on the length of the summary. However, we discovered that ChatGPT would not completely omit marginalized voices, rather, it would dilute their input, such that key details are embedded within a more general label (e.g., those who work multiple jobs to make ends meet would be labeled as simply having "time constraints"), and forcing the AI-generated summary to a few bullet points would result in minoritized voices being omitted entirely. Thus, finding the settings in which ChatGPT performed better required varying the length of the text and the length of the summaries I asked ChatGPT to generate.

I believe the settings I asked ChatGPT for in the demo, which involved summarizing 100 lines of comments into five bullet points, are a realistic and reasonable request. Many companies and executive summaries have to condense much longer text files into a few bullet points.

Accomplishments that we're proud of

By definition, summaries omit details and distill long text into the central themes and takeaways. When details are important for making minority voices heard, omitting these details can perpetuate the silencing and marginalizing of these individuals. I am proud that this project will advance data equity and fairness by providing a tool people can use to identify when certain groups' voices are not being heard. I hope it will make people more conscious of different people's perspectives and ensure they are represented.

What we learned

I learned that frontier LLMs are generally competent at summarizing texts and accounting for different marginalized voices. However, when the text gets longer (e.g., more than 100 comments) and the AI must provide succinct summaries, important details from minoritized voices can get diluted or omitted.

By entering this hackathon, I learned about the capabilities of ChatGPT 5.6 and Codex in critiquing and improving project ideas, and then easily building websites to bring those ideas to life. The workflow was simple and easy to work with.

What's next for EqualVoice

Making EqualVoice more user-friendly, such as not requiring OpenAI API keys to use, will be crucial for disseminating it. Being able to input raw video files or links to audio files are immediate next steps, so that it can check AI's abilities to summarize these conversations while auditing AI for diluting or omitting minoritized voices.

Built With

• codex
• javascript
• json
• node.js
• typescript
• vite