A contested question.
WHAT IS UNCONTESTED
Models were trained on very large amounts of existing text and images, much of it created by people who were not asked.
WHAT IS CONTESTED
Whether that training constitutes infringement Whether generated output infringes How existing law applies
Litigation is ongoing in several jurisdictions and outcomes differ.
WHAT MODELS GENERALLY DO NOT DO
Reproduce training material verbatim, usually. They learned patterns.
Though models can reproduce memorised content, particularly text appearing many times.
WHAT THIS MEANS PRACTICALLY
The legal position is unsettled and varies by jurisdiction.
For commercial use, that is a risk worth understanding.
WHAT TO DO
Check the terms of the tool you use regarding output ownership Do not ask for imitation of a specific living creator's style for commercial use Do not assume generated output is free of any claim Take advice if the stakes are significant
THE ETHICAL QUESTION
Separate from the legal one, and worth thinking about.
Many people hold reasonable views on both sides.
