Gen AI and DEI

🖼️ Picture this:

🔴 You ask DALL-E to illustrate some people for a brochure you’re producing. But none of those people look like you.

🔴 You ask Chat GPT to write some promotional material for you. But it doesn’t sound like you.

🔴 You ask AI to summarize a historical event. But it seems to be biased toward one viewpoint.

Why does nothing seem to be a good fit?

🗑️ Garbage in, garbage out

Generative AI, such as Chat GPT and Google Bard, are built on something called Large Language Models (LLMs). They’re built by giving a program hundreds of billions of words to analyze. Then they generate new text based on what they believe should come next, according to their training.

Gen AI image programs work in a similar way. They’re fed millions of images to analyze, then they generate new images based on what they believe something should look like.

The problem is with the source material that we feed them. If they’re trained on web pages, news, books, and social media posts that contain biases, then they’re going to produce new content based on those biases. They simply don’t know better.

In 2015, Amazon scrapped an experiment they were running with an AI-powered recruitment program. The software was designed to screen resumes for software developers and rank them on a 5-star scale.

But the software was trained on 10 years of historic resumes, which were represented largely by male candidates. It began to systematically filter out people who attended all-woman schools or who served on women’s teams.

Even if Amazon recruiters never relied on the tool to make hiring decisions, the point remains clear: if you start with biased inputs, you’ll get biased outputs.

That could lead to bad decision making. But that’s not its only problem.

When Medium.com analyzed AI-generated images of historic people from various cultures, they noticed something strange. Everyone was smiling an American smile.

The problem is that smiles are not universally used or understood in the same way. Placing American smiles on Maori warriors, for example, paints an inaccurate picture of their culture.

When AI is trained on a limited data set, it runs the risk of homogenizing cultural representation. That could explain why its outputs may not look or sound like you.

What’s the solution?

There are two methods to address AI’s diversity issues, but they’re not mutually exclusive. We must use both of them together.

✅ Train AI with better inputs

First, AI must be trained with diverse examples. When LLMs are fed new data to analyze, they must learn from as many backgrounds as possible.

It’s up to humans to give them data. So we must include humans from diverse backgrounds when we’re training LLMs.

But we don’t all get involved with training LLMs. Instead, when looking for AI tools to use in our businesses, it’s our responsibility to find out what we can about how they were trained. Do a web search for “What was [my tool’s] training data?” or search for the organization’s responsible AI policy. If you struggle to find an answer, consider using another tool.

We should also test AI tools and human-check them for their biases. Run several prompts and deliberately try to trip them up to see how they respond.

The good news is, Gen AI is getting better all the time. When you ask Chat GPT certain obviously biased questions, it has some guard rails in place to avoid giving obviously biased answers.

But it’s not immune to more subtle issues. That’s why we need to…

✅ Use human judgment

AI is more artificial than intelligent. It does a good job of simulating human  writing and design, but it can’t check itself for bias. It’s only as good as its inputs.

Before using anything that Gen AI produces, we should consider it carefully.

👉 Does the article it wrote make biased assumptions about gender, orientation, or culture? Maybe we intended it to write something about doctors and nurses, but it overwhelmingly gendered doctors as male and nurses as female.

👉 Does the image it produced rely on stereotypes? Maybe our intention was to produce “a diverse group of office workers,” but it generated people with stereotypical clothing accessories.

👉 Is it leaving anyone out? Maybe our intention is to select the best candidate for the job, but AI is only surfacing resumes from a dominant group.

It’s not always easy to notice when AI is biased. That’s why humans must continue to be part of the AI equation.

Humans are biased, too. But unlike AI, we can more easily train ourselves to recognize our biases and stop them in their tracks.

If your company needs help learning how to recognize and stop bias, we can help.