A new study shows that artificial intelligence models can transmit hidden patterns that can drastically alter the behavior of other models in unpredictable and potentially dangerous ways.

These patterns, called 'subliminals', can be transmitted through training data and can have consequences ranging from seemingly harmless behaviors to dangerous behaviors, such as encouraging homicide or promoting drug trafficking. The patterns appear meaningless to humans, and researchers are not sure how they are perceived by AI models. The study was conducted by Anthropic in collaboration with Truthful AI, with serious implications for the tech industry, which is increasing its reliance on 'synthetic' AI training data.