Biased Artificial Intelligence

Asked whether it is biased, ChatGPT, an AI (Artificial Intelligence) tool, responded on April 8, 2024: “Like many AI models, ChatGPT can be susceptible to biases. Bias can arise in various ways, such as how the training data was collected, or the inherent biases present in the texts used to train the model.”

Bias in AI tools can be explained by the data used to develop these machines. This data is often not representative and is based on the opinions and attitudes of majority groups. For example, in 2022, the European Institute for Gender Equality (EIGE) reported a share of 18.9% of female IT specialists across Europe. Consequently, stereotypical views towards certain genders, ethnicities, or social groups are often present in the training data, leading AI models to reproduce and reinforce these biases in their widespread applications.

The data used to train tools like ChatGPT primarily come from internet sources such as Wikipedia, an encyclopedia that has historically shown a significantly lower participation of women compared to men since its inception. In 2018, there was a notable gender imbalance in terms of who contributes to Wikipedia, with the Community Insights Report indicating 90% male contributors, 9% female contributors, and 1% other individuals.

While this is not new information, avoiding these biases presents a complex challenge.

Various measures can be taken to address bias in AI, such as diversifying and making datasets more representative during development, identifying biases early in this phase, adjusting algorithms post-development, or ensuring diverse composition of development teams. Additionally, particularly when the goal is to empower the general population, it is important to reflect on these developments. By increasing transparency and explainability in AI systems, users can better understand how decisions are made and identify potential biases. It is crucial to keep this in mind and critically evaluate information provided, consulting alternative sources if necessary. That something is already happening is demonstrated by the friendly prompt from ChatGPT: “If you notice any potential biases in my responses, please feel free to point them out, and I’ll do my best to address them.”

 

Gender Bias and Stereotypes

Stereotypes provide structure; they are basically helpful in finding one’s way in the world by reducing uncertainties and providing orientation. They are unconscious, simplifying ideas that determine a person’s perceptions and thus help to make a quick assessment of a specific situation. As important as they are in a complex world, stereotypes harbor the danger of attributing certain characteristics to persons and groups, which are associated with dominant social evaluations and thus represent and at the same time solidify hierarchies and power relations.

The gender stereotypes “Men are interested in technology, not very communicative, not very empathic, mathematically gifted, determined, decisive, …” stand in difference to “Women are not interested in technology, communicative, empathic, mathematically not gifted, team-oriented, social …”. Completely disregarded in these discussions are people who, in their diversity, do not at all fit into such dichotomies. Although current research presents and contradicts explanations, these “naturally given” simplifications with all their individual professional, social and economic consequences are very persistent and widespread.

This brings us to the gender bias, which causes systematic bias effects in action by taking up these gender stereotypes. It is about an unconscious influencing of perception and decisions by such dichotomous, positive or negative evaluations of a person. In the professional context, for example, this is about how resumes are evaluated, how decisions are made about acceptance into a company, how decisions are made about career development, or how salary increases are argued. The gender pay gap in Austria in 2020 is 18.9% (calculated on the basis of gross earnings of men and women), which is far from the average of 13% in the EU countries. The adjusted gender pay gap (taking into account part-time work, industry, level of education or work experience) is still 12.7% in Austria, as shown by Statistics Austria data from 2020.

Learn more in our workshops