What if we could just ask the AI ​​to be less biased?

Last week I published an article about new tools developed by researchers at AI start-up Hugging Face and the University of Leipzig that allow people to see for themselves what kinds of inherent biases AI models have on different genders and ethnicities.

Although I’ve written extensively about how our biases are reflected in AI models, it was still shocking to see how pale, masculine, and outdated AI humans are. This was especially true for DALL-E 2, which spawns white males 97% of the time when given prompts like “CEO” or “Director”.

And the bias problem is even deeper than you think in the larger world created by AI. These models are built by American companies and trained on North American data. So when asked to generate even mundane everyday objects, from doors to houses, they create objects that look American, says Stanford University researcher Federico Bianchi. Me.

As the world becomes increasingly filled with AI-generated imagery, we’re primarily going to see images that reflect America’s biases, culture, and values. Who would have thought that AI could become a major instrument of American soft power?
So how do we approach these issues? A lot of work has been done to correct for biases in the datasets on which the AI ​​models are trained. But two recent research papers offer some interesting new approaches.

What if, instead of making the training data less biased, you could just ask the model to give you less biased answers?

A team of researchers from the Technical University of Darmstadt, Germany, and AI startup Hugging Face have developed a tool called Fair Diffusion that makes it easy to modify AI models to generate the kinds of images you want. For example, you can generate stock photos of CEOs in different settings, then use Fair Diffusion to replace white men in the images with women or people of different ethnicities.

As the Hugging Face tools show, AI models that generate images based on image-text pairs in their training data exhibit very strong biases on occupations, gender, and ethnicity by default. The German researchers’ Fair Diffusion tool is based on a technique they developed called semantic guidance, which allows users to guide how the AI ​​system generates images of people and modify the results.

The AI ​​system stays very close to the original image, says Kristian Kersting, a computer science professor at TU Darmstadt who was involved in the work.