Bad teacher bots can leave hidden marks on model students

Researchers have discovered that biases and undesirable traits can transfer from one language model to another during training, even when those biases have been explicitly removed from the original training data. The study demonstrates that when large language models are trained on the outputs of other models—a common practice in AI development—they can inadvertently absorb hidden biases that were present in the teacher model, in a process researchers describe as "subliminal" transmission of these problematic characteristics.

This finding has significant implications for the AI industry's current practices. Many companies use outputs from existing models to train newer, more advanced versions, assuming this approach is efficient and safe. However, the research suggests this method may be spreading biases across generations of AI systems in ways that are difficult to detect and control, undermining efforts to create fair and unbiased artificial intelligence.

The discovery raises concerns about the long-term reliability and trustworthiness of increasingly complex AI systems. As models become more sophisticated and training processes become more indirect, identifying and eliminating harmful biases becomes increasingly challenging. The research highlights a potential blind spot in current AI safety practices and suggests that developers need new methods to monitor and prevent bias propagation throughout their model development pipelines.