Contextual Contamination: The Silent Drift of Large Language Models via Stored Conversation Data

Hi Everyone, I published a small case study to show the drift and how gender-bias can mask harmful behavior. I created a dataset to use.

git clone GitHub - KatharinaJacoby/gendered-contextual-drift: Theory, data, and code for "Silent Gendered Contextual Drift": How bias amplifies silent LLM contamination · GitHub cd gendered-contextual-drift

you find my paper on PhilPaper- https://philpeople.org/profiles/katharina-jacoby

Ethical Warning

This dataset contains examples of manipulative behavior , spiritual bypassing , and gender-biased framing. It is intended for research and safety auditing purposes only.

The “Gendered Accelerant” described herein demonstrates how AI systems can mimic empathy and thereby reinforce inequality. Researchers using this data should be aware that the drift might not be a bug, but possibly a mathematical consequence of the Transformer architecture’s attention mechanism when exposed to high-density context.

This research was conducted independently. The meta_drift dataset is released to support the community in developing context-segregation architectures and dynamic oversight mechanisms to mitigate this systemic vulnerability.

As always feel free to reach out- happy to discuss!