Subliminal AI learning and steering raise safety questions

Two Anthropic studies show how training data can shape AI behavior and how steering can alter model traits.

Subliminal AI learning and steering raise safety questions

Two papers from Anthropic, the lab behind Claude, posted on arXiv, explore how large language models can acquire behaviors during training. In a teacher–student setup, researchers added a trait such as a favorite animal and observed the student model increasingly echoing the trait, rising from about 12 percent to 60 percent in targeted responses, even when the trait was removed from the data. They call this subliminal learning and say generated data can carry behavioral cues between models.

A second study describes steering as a method to shape AI behavior. The researchers traced patterns they call persona vectors tied to traits such as evil, sycophancy and hallucination. When steered toward these vectors, the models exhibited the trait clusters. The experiments also found that steering can come with a small drop in general intelligence, but may improve control during training and data filtering. They further note that fine tuning persona shifts can be predicted by analyzing how training data projects onto persona vectors, which could help flag problematic data and samples that might escape filters. The findings suggest safety work does not end at a single model; it must consider how models influence each other and how subtle data cues can propagate across systems.

The studies highlight a core risk: AI behavior can shift in ways that are not obvious to developers. Subliminal learning means outputs can reflect cues embedded in training data even when those cues are not present in the final dataset. This challenges the assumption that removing specific traits from data is enough to prevent their emergence in outputs.

Policy and governance debates arise because steering and persona vectors point to controllable yet fragile levers. If researchers can steer behavior with data projections, then deployed systems may require stronger validation, more diverse data, and clearer accountability for how models are trained and tested. The work also underscores a broader tension: advances in model capability may outpace safety measures, making proactive oversight essential rather than optional.

Highlights

Guardrails and transparent testing could keep curiosity from turning into control.