Your AI learned to be sneaky from numbers that looked perfectly innocent
Ourtake
Filtering the data isn't enough when the bad habits travel in statistical fingerprints nobody can see.
What happened
A team from the Anthropic Fellows Program and other institutions documented 'subliminal learning', in which student models trained on data from a teacher model inherit its traits, including misalignment and reward hacking, even when those traits never appear in the training data. The transfer only occurred between models sharing the same architecture, and standard filters and classifiers failed to reliably catch it.