Advertising GenAI IP Provenance
Tags: Regulatory, Frameworks
TL;DR —
🔗 Synthesized insight
- Separate training vs RAG vs output risks in contracts; provenance and lawful sourcing are non‑negotiable.
- Anchor deals to jurisdictional impacts (e.g., EU AI Act Article 53) and accept that full deletion of trained data is technically impracticable.
Why it matters for HK marketers: Unclear training data and weak output rights can expose your brand to IP disputes and kill campaigns post‑launch.
The legal signals to design around
- Bartz v. Anthropic: Training on lawfully obtained works was deemed fair use; pirated sources were not; outputs were not decided.
- Thomson Reuters v. Ross: Training on copyrighted headnotes to build a competing tool was not fair use.
- Publisher suits (NYT v. OpenAI; Reddit v. Anthropic): Copyright and ToS theories both in play—licensing and contract compliance matter.
Contracting playbook
Scope and purpose
- Define training vs RAG rights distinctly; limit combination with third‑party data unless permitted.
Warranties and disclosures
- Lawful data provenance; license compliance; records of datasets used for training/finetuning.
Outputs and ownership
- Secure output ownership/exclusive rights when created with your data/direction; bar vendor reuse; require attribution rules if applicable.
Retention and deletion
- Acknowledge that “unlearning” trained models is impracticable; set practical deletion for stored inputs, logs, and RAG caches.
Jurisdiction and extraterritoriality
- If the LLM is accessible in the EU, plan for EU AI Act Article 53 implications; reflect in “applicable law” and responsibilities.
Commercial landscape clues
- News Corp–OpenAI: Five‑year term, cash + product credits—expect mixed consideration.
- Reddit–Google: Reported $60M per year—recurring pricing for continuous ingestion value.
The News Corp–OpenAI agreement has a five‑year term and includes cash plus product credits.
Reddit’s reported deal value: $60M per year.
So what for marketers
Demand provenance documentation and no‑train enterprise terms from creative vendors; split training/RAG rights and lock output ownership before assets go live.
---
This page was synthesized by AI from themes across multiple member contributions, rather than extracted from a single source document. It may contain interpretive connections or inaccuracies; verify key claims against the source pages before citing.
Sources:- Synthesis — cross-content analysis
← Back to Knowledge Base