Skip to main content
Back to List
AI Infrastructure

Model Distillation

A method that trains a smaller model from the output signals of a larger model

#Model Distillation#distillation attack#teacher-student training

What is model distillation?

Model distillation is a technique where a smaller student model learns from outputs produced by a larger teacher model.

It is widely used to balance quality, latency, and cost in production AI systems.

Why can it become controversial?

Distillation itself is not illegal. The issue appears when outputs from a third-party model are collected and reused for training without authorization.

In that case, Terms of Service violations, data-rights disputes, and policy risks can emerge.

Practical checkpoints

  1. Data provenance: Keep clear records of where training signals came from.
  2. ToS compliance: Review API terms for clauses that ban competitive model training.
  3. Operational separation: Separate research experiments from production datasets and release pipelines.

Related terms