Bias in AI: How Meta’s image generator highlights systemic problem

Meta’s recent AI image generator is raising concerns about a prevalent issue in the tech industry: bias in AI.

The tool struggles to generate images of interracial couples, highlighting how AI systems can reflect and amplify the prejudices present in the data they’re trained on.

In a CNN report, the generator consistently produced images of same-race couples for prompts requesting interracial pairings. This is not an isolated incident. Earlier in February, Google paused its AI image tool, Gemini, due to historically inaccurate outputs and bias towards people of colour. OpenAI’s Dall-E tool has also faced criticism for perpetuating racial stereotypes.

This is not the only case of bias in text to image generators. Last year, the initiative of TEDxAmsterdam, Missjourney, was created as an AI tool to celebrate women, its aim was to combat gender bias in AI systems and promote inclusivity across all professions.

Bias according to IBM refers to “the occurrence of biased results due to human biases that skew the original training data or AI algorithm—leading to distorted outputs and potentially harmful outcomes.” In this case, Meta’s image generator is reflecting a certain kind of bias- stereotyping bias which occurs when it reinforces harmful stereotypes.

These cases expose a major challenge in AI development: ensuring fairness and mitigating bias. AI systems learn from the data they are fed. If that data contains racial biases, the AI will inherit them, potentially leading to discriminatory outputs.

How can we address bias in AI?

While addressing bias is an ongoing process, here are some potential solutions:

• Data Diversity: Training AI on more diverse datasets that accurately reflect the real world is crucial. This includes ensuring racial diversity across genders, age groups, and other demographics.
• Human oversight: Incorporating human review and feedback during AI development can help identify and mitigate bias before it gets into the final AI system. 
• Algorithmic fairness: Researchers are actively developing algorithms that can detect and address bias within AI systems themselves.

Although Meta has addressed this in a previous blog post, there is still a lot of work to be done in addressing bias. The company said in a blog post on building AI features responsibly, “We’re taking steps to reduce bias.” “Addressing potential bias in generative AI systems is a new area of research. As with other AI models, having more people use the features and share feedback can help us refine our approach,” they stated.