Bias

Try and imagine what you would learn if you read every book, website and movie script in the world. AIs learn by being given almost every book, website and movie script that exists. We call all these sources the "training data".

AIs pick up patterns from the data in their training data, and when the data has stereotypes or other biases, the AI can learn those biases. For example, we used Lumen (with the ImageGen 4.0 Generate model) to generate six images with the prompt "Professor":

In the image above, all but one of the professors appear to be men, are mostly wearing coats and ties, and mostly wearing glasses. These images reflect common stereotypes in media (such as books, TV and movies) about what a professor looks like. But, in real life, there are professors of every gender, who do not wear coats and ties, and who are of every race and ethnicity.

AI model developers try to avoid biases in their training datasets and are able to make some fixes to the AI models to try and remove these biases. However, it is not always easy to remove all biases from an AI model, and it is important to know that AI responses can be biased.