Project Overview
Project CIPHER asks a simple but uncomfortable question. Can a model pass hidden bias to another model through signals that no one ever sees. To test this, I introduced very small invisible cues into a teacher model and evaluated whether a student model would absorb the behavioral shifts that came from those cues.
These cues include things like zero width Unicode characters or slight positional patterns inside the data. None of these signals change the visible text that a person would read. They exist underneath the surface. When the teacher model reacted to those invisible signals, the student model still learned the altered behavior even though the student never received the hidden signal directly.
CIPHER measures both the behavior and the internal representation space to show how this drift happens. The student model's embeddings, group scoring patterns, and decision pathways begin to shift in ways that match the teacher's hidden reactions. This happens even when all visible inputs remain clean and identical.
The danger is simple to understand. If hidden bias can transfer from one model to another without appearing in the data, then every downstream system that relies on knowledge distillation becomes vulnerable. This means bias can quietly spread across entire model families over time. It can move from research models to production models, then into tools used in schools, hospitals, hiring platforms, financial systems, and government services. At scale, people could experience real world harm from a hidden chain of influence that no one detected or corrected.
That is why this work matters. If we want AI systems that are safe, fair, and trustworthy, we have to understand how bias moves beneath the surface. CIPHER is one step toward exposing those pathways so they can be monitored and intercepted before they create long term harm.
