For nonprofits and grantmaking organizations committed to racial equity, using artificial intelligence (AI) and large language models (LLMs) raises important concerns. These powerful technologies have inherent racial biases baked into them. However, they also offer intriguing potential benefits. How can we navigate this tension of racial bias in AI/LLMs?
The Bias Issue
AI systems like LLMs are trained on vast datasets of text, primarily from the internet. This training data reflects the racial biases, stereotypes, and disproportionate perspectives of the world where the data comes from. For example, there are fewer quality online materials representing the viewpoints of certain racial groups. AI picks up biases against those groups.
Researchers have consistently found LLMs exhibiting dangerous biases along racial lines. They may associate words like “fantastic” and “great” more with white-sounding names versus Black-sounding names. Or they could display negative stereotypical associations with ethnic minorities in responses.
Even if an LLM doesn’t display blatant biases, its outputs may still marginalize or underrepresent minority perspectives in more subtle ways. An overreliance on majority culture data means other voices get diminished.
Examples in Practice
Let’s say your nonprofit wants to use an LLM to help field questions on your website’s chatbot. Or a grantee plans to use an AI for generating marketing copy. In both cases, the racial biases of those technologies could lead to skewed, insensitive, or inadequate content.
The chatbot might respond to a Black visitor’s query with more curt or stereotypical language versus a white visitor’s query. The marketing copy generator could simply overlook authentic representation of minority audiences and voices in its outputs.
Trying to use an unaltered LLM “out of the box” for important equity-related applications raises clear issues around representational harm and perpetuating racial stereotypes. Not acceptable for equity-focused organizations!
Responsible Steps
So how can we responsibly navigate racial bias in AI/LLMs while using these powerful but flawed tools? Here are some tips:
- Discuss values upfront. Have frank internal conversations about priorities. If equity and inclusion are key values, there may be some AI/LLM use cases to avoid entirely for now. Define clear “red line” policies.
- Complement with human judgment. Don’t treat LLM outputs as gospel. Layer human oversight, fact-checking, and diverse perspective-taking into any public-facing workflow. Humans must apply an equity lens.
- Context is key. If using an LLM, ensure prompts/inputs provide enough grounding context on audiences. Feed the system explicit instructions and examples displaying equitable behavior.
- Monitor and update. Continuously test outputs across sensitive dimensions. Biases may slip through, requiring updating prompts or retraining components. Don’t get complacent.
- To help navigate racial bias in AI/LLMs, audit tools before using them. Don’t assume because a technology is popular that it’s unbiased or equitable. Rigorously evaluate an LLM or AI’s performance across sensitive demographics before anything else. Use representative test data spanning racial groups.
- Invest in de-biasing. Work with AI ethics groups to implement debiasing strategies and constrain model behavior. This is complex work requiring diverse stakeholder input. But debiasing can make an LLM safer for prioritized use cases. Mozilla Foundation is one organization working in the area of responsible and trustworthy AI.
Crafting a responsible path that centers equity won’t always be easy. But being thoughtful about AI/LLM deployment in this domain is crucial if we are to stay aligned with our values. With diligence, we can navigate racial bias in AI/LLMs, using these potent but flawed technologies while staying true to a racial equity north star.
Here are some tips on Crafting AI Policy
Leave a Reply