Artificial intelligence models depend on massive datasets: books, articles, music, code, and artwork scraped from the internet. This practice sits at the center of global intellectual property litigation: Can AI developers legally scrape and train on protected works without authorization from the copyright owner?
The answer is complex and strictly territorial. Across key jurisdictions, courts and legal authorities are distinguishing between data ingestion, machine learning, and conversational outputs.
WHAT GLOBAL PRECEDENTS REVEAL
- United States (Fair Use): In Bartz v. Anthropic PBC (2025), a federal court held that using lawfully acquired books to train an AI model qualified as fair use. Crucially, the court held that acquiring and retaining pirated books in central libraries was not protected. Similarly, in Kadrey v. Meta Platforms Inc. (2025), Meta obtained summary judgment due to evidentiary shortcomings regarding market substitution. The lesson: fair use during ingestion does not validate pirated data acquisition.
- Germany (Memorization & Output Infringement): In GEMA v. OpenAI (2025), the Munich Regional Court found copyright infringement where ChatGPT reproduced verbatim lyric excerpts in response to simple prompts. Infringement does not end at ingestion; model memorization that regenerates protected creative expression triggers independent liability.
- The European Union (Text & Data Mining vs. Outputs): In the proceedings involving Like Company v. Google Ireland, Advocate General Maciej Szpunar of the Court of Justice of the European Union (CJEU) delivered a landmark Opinion distinguishing between training inputs and generative outputs: (1) While text and data mining (TDM) exceptions under the EU Digital Single Market Directive may permit data ingestion for model development (where rightsholders have not reserved rights), that statutory shield does not grant AI chatbots free rein to reproduce protected press and news content in conversational responses. (2) Unlicensed reproduction of substantive news content to end-users remains subject to copyright and press publisher rights.
AT NATIONAL LEVEL: CURRENT LAW VS. LEGISLATIVE REFORM
The Law in Force (Chapter 406): Under Section 17 of the Copyright and Performance Rights Act, Chapter 406 of the Laws of Zambia, rightsholders hold the exclusive right to authorize or restrict reproduction in any material form. Section 21 fair dealing exceptions are strictly purpose-specific (private study, research, criticism, or review). There is currently no statutory exception permitting commercial AI training without a licence.
The Reform Horizon (The Copyright and Related Rights Bill, 2025): Clause 52(1)(d) proposes an exception for the processing of works for text and data mining, provided the processing is strictly non-commercial. Commercial AI developers operating in or targeting this market would sit outside that exemption, meaning licensing arrangements and rightsholder permissions remain central.
DAC VIEW
- Public Availability Does Not Justify Free Exploitation: Content accessible online is not stripped of its underlying economic rights.
- Inputs Differ From Outputs: An AI model trained lawfully under a data mining exception may still commit infringement if its outputs reproduce substantial parts of protected source works.
- Licensing Frameworks Matter: With statutory reforms leaning toward protecting creators from commercial exploitation, direct licensing agreements and remuneration models are becoming standard risk-mitigation measures for tech developers.
.
.
DAC Legal Insights is published by Dzekedzeke & Co. Legal Practitioners for public legal education and informational purposes only. It does not constitute formal legal advice.
.
.
Dzekedzeke and Company,
We Learn Everyday.
#DzekedzekeAndCo #ArtificialIntelligence #TechLaw #Lusaka #Zambia #DACLegalInsights #LegalTech #RuleOfLaw #DigitalJustice #CaseManagement #DACLegalInsights #IntellectualProperty #CopyrightLaw #ArtificialIntelligence #AIPolicy
