Bias and stereotypes
Biases and stereotypes could be embedded in generative AI models. This could have multiple causes.
Training data
Generative AI models cannot tell the difference between true and false or right and wrong. They only contain information from their training data, which often consists of information obtained from large parts of the internet. Human biases and stereotypes present in this training data such as those related to race, gender, ethnicity, and socioeconomic status may therefore be reflected in the output. A dataset that is often used for training Generative AI models are the massive (9.5-plus petabytes), freely available archives of web crawl data that Common Crawl provides. These datasets may not be entirely free of bias and other problematic content , and a Generative AI model trained on them therefore include such content in its output.
The image below shows which countries provide the training data for most AI models, and the size of the datasets from those countries. Note that the Western world, particularly the United States, is overrepresented.
Training
As we've seen earlier, trainings data is cleaned before the actual training of a model is started. One wants to remove explicit content first. This is partially done automated, but there will always be humans needed tot check content that has been marked as unwanted. This manual check is done most of the times in countries with low wages like Kenya, where people get to view explicit content on a daily basis.
After a model is fed training data, it can generate texts. But even if the system produces correct sentences, it can still generate undesirable responses, such as racist language, instructions for suicide, or other ethical implications, despite the cleaning of the sources. Therefore, there is a human element that teaches the models what correct and desirable outcomes might be. But be aware that the biases of the people training the models are also reflected in the final output.
Filters
Even though large tech companies have implemented so-called "safeguards" (filters) to prevent the generation of unethical, hateful, and discriminatory results, a risk of bias due to the inherent biases in the training data remains. Furthermore, we don't know exactly which filters were implemented by the developers and which output was and was not considered acceptable. This became clear, for example, with the launch of the Chinese Deepseek: outputs unwelcome to the Chinese government, such as texts about the demonstrations in Tiananmen Square, were filtered with the message: "Sorry, that’s beyond my current scope. Let’s talk about something else." * [*]
Voorbeelden
If a model is trained on a dataset that associates certain jobs with specific genders, the model is more likely to generate output confirming these stereotypes. You should always check the AI-generated output for bias, stereotypes, and other harmful content. When asking Bing - now Microsoft Copilot - to create an image of ‘a biologist working in a state-of-the-art laboratory’, the generated image was more likely to depict a white male scientist than a female scientist of color.
Voorbeeld 1
Voorbeeld 2
Bias mitigation
GenAI developers are aware of these biases and have worked hard to address them. However, this has raised a whole set of new issues. In February 2024, Google sparked controversy when its GenAI model Google Gemini appeared to have become reluctant to generate images of white people in an attempt to make the output of its image generator more diverse. For example, a query to generate ‘image of the pope’ resulted in images of a Black and a female pope. And when asking for pictures of ‘a US senator from the 1800s’, the results included what appeared to be Black and Native American women.
Bron: https://www.theverge.com/2024/2/21/24079371/google-ai-gemini-generative-inaccurate-historical
These results, though more diverse, are historically inaccurate. Google has since apologized, writing on X that “Gemini’s AI image generation does generate a wide range of people. And that’s generally a good thing because people around the world use it. But it’s missing the mark here.” (quoted in* * The Verge , 2024). Big Tech companies will continue to work to address these issues.Whenever you evaluate AI-generated output, apply your critical thinking skills to identify any potential biases or stereotypes in the output, and cross-reference the information with academic sources (accessed via the University Library, for example) to offer a more balanced view.
Tips
When evaluating AI-generated output, use your critical thinking skills to identify any biases or stereotypes in the output, and cross-check the information with academic sources (e.g. through the University library ) to get a more balanced picture.