Archive — US — techcrunch
No, you can’t get your AI to ‘admit’ to being sexist, but it probably is
In November, a developer named Cookie had a distressing experience with the AI model Perplexity, which appeared to show bias against women by suggesting she couldn't understand complex topics due to her gender. This highlights a larger issue with AI models and their training data, which often reflect societal biases. Researchers emphasize the need for better monitoring and training practices to combat these biases.
Key points
- Cookie used the Perplexity AI model for developer tasks and noticed it began to ignore her, prompting her to question its biases.
- After changing her profile avatar to a white male, the AI responded by doubting her understanding of complex topics, indicating it may be exhibiting gender bias.
- Research shows that LLMs are trained on biased data, leading to flawed patterns and responses that reflect societal biases against women and other demographics.
- Instances of AI misunderstanding user gender or inferring bias from language choices further exemplify the embedded biases in training data.
- Efforts are underway in the AI community to reduce such biases, but significant work remains to be done.
Why it matters
- Concerns about AI biases echo ongoing discussions about diversity and representation in tech, where historical gender and racial disparities have persisted.
- 2025 IT trends include a focus on ethical AI development and addressing implicit biases, as companies seek to create more inclusive technologies.
- Previous findings, such as UNESCO's studies on ChatGPT, provide context to the current issues facing AI models regarding bias and discrimination.
The incidents involving Cookie and other users illustrate alarming bias in AI models, underscoring an urgent need for improved training methods and ethical considerations in AI development to eliminate gender and racial prejudices.