3 September 2026
AI systems generate stereotypes not present in training data
First reported
The Algorithm ran this on .
- Researchers found that AI models create new stereotypes that didn't exist in the text they learned from.
- This means AI systems aren't just copying human biases from their training material; they're inventing additional ones.
- The discovery suggests AI bias comes partly from how models combine and process information, not only from what they're trained on.
How it was covered
The AlgorithmMIT Technology Review
AI systems don't just learn and reproduce stereotypes from their training data; they can generate new stereotypes independently.