Copyright and training data
What changes when you train, retrieve or publish with AI? · updated 2026-09-07
Data acquisition, model training, retrieved passages and generated outputs raise distinct questions. Access to a page is not automatically a license to reproduce it. A training dispute also does not determine ownership or infringement for every output you publish.
The concern
Training and outputs can affect creators' licensing and sales markets. Where acquisition was unlawful or outputs substitute for protected expression, 'the model learned patterns' is not a complete answer. Creators need practical ways to discover use and negotiate permissions.
The counter-view
Some uses support research, analysis or other transformative purposes. Individual permission for every input can create costly barriers, especially for smaller developers. The Copyright Office examines varying uses and licensing mechanisms rather than treating all training alike.
The answer depends on jurisdiction, material, acquisition, use and output. This page explains questions to investigate; it does not clear a dataset or publication for legal use.