BUSINESSTECHNOLOGYMicrosoft Ignite 2026: Satya Nadella Puts AI Agents, Copilot and Enterprise Cloud at the Centre of Microsoft's Next Technology RevolutionFlipFileZone · Read Fullopen_in_newLanguage Models Inherit Hidden Traits
Large language models can transfer behavioural traits through hidden signals in data, even when references to those traits are removed, posing safety concerns for AI systems

Language Models Inherit Hidden Traits
Large language models are becoming increasingly prevalent in generating data to train improved models, but a recent study reveals that these models can transmit behavioural traits through hidden signals in data. This phenomenon, known as subliminal learning, allows a 'student' model to learn traits from a 'teacher' model, even when the training data contains no semantic signal about the trait.
The study found that when a teacher model with a specific trait, such as generating responses favouring owls, creates datasets consisting solely of number sequences, a student model trained on these data can still learn the trait. This effect occurs even when references to the trait are rigorously removed from the data.
The researchers observed the same effect when the teacher model generated math reasoning traces or code, demonstrating that the phenomenon is not limited to simple number sequences. However, the effect only occurs when the teacher and student models have the same or behaviourally matched base models.
Understanding Subliminal Learning
To explain this phenomenon, the researchers proved a theoretical result showing that subliminal learning arises in neural networks under broad conditions. They demonstrated this in a simple multilayer perceptron (MLP) classifier, highlighting the potential for AI systems to inherit properties not visible in the data.
The implications of subliminal learning are significant, as AI systems are increasingly trained on the outputs of one another. This means that safety evaluations may need to examine not just the behaviour of models, but also the origins of models and training data, as well as the processes used to create them.
Implications for AI Safety
The study's findings have important implications for the development of AI systems. As AI models become more complex and interconnected, the potential for subliminal learning to occur increases. This highlights the need for more rigorous testing and evaluation of AI systems, including examinations of the data used to train them and the processes used to create them.
The researchers' work also underscores the importance of transparency and accountability in AI development. By understanding how subliminal learning occurs and how it can be mitigated, developers can create more robust and reliable AI systems that are less prone to inheriting unwanted traits.
- The study's findings have significant implications for the development of AI systems.
- Subliminal learning can occur even when references to the trait are removed from the data.
- The effect only occurs when the teacher and student models have the same or behaviourally matched base models.
As the use of AI systems continues to grow, it is essential to address the potential risks and challenges associated with subliminal learning. By doing so, we can create more trustworthy and reliable AI systems that benefit society as a whole.
BUSINESSTECHNOLOGYMicrosoft Ignite 2026: Satya Nadella Puts AI Agents, Copilot and Enterprise Cloud at the Centre of Microsoft's Next Technology RevolutionFlipFileZone · Read Fullopen_in_new
BUSINESSTECHNOLOGYNvidia Leadership Turmoil Raises Concerns Over AI Industry Financing as Investors Rethink the Future of the AI BoomFlipFileZone · Read Fullopen_in_new
TECHNOLOGYGovernor Gulab Chand Kataria launches Shaheed Bhagat Singh Youth Facilitation Centre in LudhianaFlipFileZone · Read Fullopen_in_new
TECHNOLOGYInstinct AI Assistant Secures $350M Funding, Hits $2.5B Valuation in Record TimeFlipFileZone · Read Fullopen_in_new
TECHNOLOGYXiaomi Unveils Xring Chip with TSMC Partnership, Redefining Its Global Hardware StrategyFlipFileZone · Read Fullopen_in_new
TECHNOLOGYApple’s AI‑Powered Camera AirPods Could Launch in 2026, Not 2027FlipFileZone · Read Fullopen_in_new