An online platform wants to promote artists by having an algorithm automatically evaluate artworks and select them for exhibitions.
The algorithm is trained with historical data from art critiques and exhibition successes.
After the launch, it becomes apparent that the algorithm mainly favors conventional, well-known styles and hardly considers experimental or unusual works.
Many creative and innovative artists feel disadvantaged, even though their works are artistically valuable.
The team wonders: Why does the algorithm evaluate artworks so one-sidedly, even though it was trained on diverse data?
Question: Why do AI models that are supposed to evaluate artworks tend to favor traditional or already successful styles and disadvantage innovative, unconventional works?
Solution follows tomorrow.
Solution
The algorithm learns patterns from historical data in which traditional and already successful artworks are overrepresented.
Since experimental or new styles occur less frequently in the training data, they are recognized or evaluated worse by the model.
The model therefore prefers works that are similar to the training data because it classified these as “successful” or “valuable.”
Innovative art that deliberately deviates from the familiar thus falls through the cracks because it does not correspond to known patterns.
This leads to a bias that restricts creativity and impairs the diversity of art.
Result: An evaluation algorithm based on historical data can unintentionally favor traditional art styles and disadvantage new creative approaches. For fair and diverse art evaluations, training data and models must be designed in such a way that innovation and diversity are explicitly promoted and not suppressed by historical biases.