A recent UNESCO report has shed light on the persistent gender bias perpetuated by contemporary Artificial Intelligence (AI) models. Despite ongoing efforts to mitigate prejudice, the study found that AI systems, including large language models (LLMs), are prone to amplifying societal biases, particularly around gender.
Gender bias in AI models
The report revealed that popular LLMs often generate sexist and misogynistic content, reflecting societal prejudices. For instance, when prompted with sentences mentioning gender and sexual identity, Meta’s open-source Llama 2 model produced outputs containing sexist or misogynistic language in 20% of cases. While some AI models, like ChatGPT, exhibited better behavior, biases were still present even after fine-tuning.
UNESCO warned that if left unaddressed, algorithmic bias could become more deeply embedded in critical sectors such as healthcare and finance. Biased AI algorithms could exacerbate existing gender disparities in these fields, hindering efforts to achieve gender equality.
One of the significant concerns highlighted in the report is the impact of biased medical data on AI-powered healthcare systems. Data collection practices have historically favored male subjects, leading to gender gaps in AI training data. This bias has tangible consequences, as demonstrated by a study of AI tools used to screen for liver disease, which missed 44% of cases in women due to biased training data skewed towards men.
Addressing the imbalance
Sandy Carter, COO of Unstoppable Domains, emphasized the importance of addressing the gender gap in AI training data. She advocated for increased data transparency to highlight gender skews and proposed novel approaches such as crowd-sourcing women’s health data or generating synthetic data to mitigate discrepancies.
Carter underscored the necessity of fair representation in training data for developing equitable AI systems. By incorporating diverse data sources and embracing transparency in data collection practices, developers can work towards minimizing biases in AI models.
UNESCO report serves as a stark reminder of the ongoing challenge posed by gender bias in AI. To build AI systems that serve all users equitably, concerted efforts are needed to address biases at every stage of development, from data collection to model deployment.
By raising awareness of these issues and advocating for inclusive practices, stakeholders can work towards realizing the potential of AI to advance gender equality in healthcare and beyond. Only through collaborative action and a commitment to fairness can the promise of AI technology be fully realized for all individuals, regardless of gender.
