Image: theofficialacm.substack.com
The Story
AI models can learn from books in some circumstances, but copyright questions turn on how the works were acquired, used, and reproduced.
AI training is not automatically protected merely because it resembles human learning. The ordinary legal reading separates copying during data collection from later model behavior, then applies fair use case by case. The strongest pressure points are unlawful acquisition, permanent internal libraries, memorization, and outputs that substitute for protected works.
- Legal testCourts are likely to examine the full training process rather than grant a blanket rule for or against AI learning.
- Creator protectionPirated sources and outputs that reproduce protected text remain distinct risks even where training itself is treated as transformative.
- The Copyright Office’s report analyzes collection, training, memorization, deployment, and outputs as separate parts of the infringement and fair-use inquiry.
- In Bartz, Anthropic’s use of purchased and pirated books involved separate questions about internal libraries, training copies, and possible memorization.
- In Kadrey, Meta’s use of books from shadow libraries was litigated as a fair-use question, with unlawful acquisition central to the dispute.
Deeper Insight
The Copia Institute says AI can learn from books. Copyright could turn public knowledge into a private toll road.
A book on a shelf teaches whoever opens it. The Copia Institute, a public-interest legal group, says AI training is learning, not infringement. But these cases also involve pirated inputs, retained copies, memorization, and outputs that replace protected works. The deeper fight is whether companies can turn millions of books into commercial systems without paying for the learning inside them.
- What it could changeA broad ruling could make machine learning from public works a licensed privilege.
- How it touches todayThe same fight already splits collection, training, memorization, and outputs in court.
- Our futureKnowledge could stay visible while access to its patterns becomes privately owned.
- The Copia Institute argues that copyright controls copying, distribution, and performance, not reading, listening, watching, or learning from a work.
- Lemley and Casey argue that AI training should generally be allowed because licensing millions of works from thousands of owners is not practically feasible.
- Atkinson argues that fair-use arguments for AI training would also apply to humans and warns against granting GenAI a special exemption for mass reproduction.
Sources · 8
Featured · Analysis · techdirt.com
Wherein The Copia Institute Tells SDNY That AI May Have Problems But So Is Using Copyright Law To Try To Fix Them
There’s a lot wrong with AI. But bad things happen when we try to use copyright law to fix it. Which is what we just told the Southern District of New York, where OpenAI and Microsoft are defending themselves in several cases alleging that their training violated the copyrights of the works the AI models were trained on. We made a few points . One overarching point is that what the plaintiffs are
Read sourceAnalysis · law.stanford.edu
Fair Learning - Journal Article - Stanford Law School
Fair Learning - Journal Article - Stanford Law School ## Abstract Neural network and machine learning artificial intelligences (AIs) need comprehensive data sets to train on. Those data sets will often be composed of images, videos, audio, or text. All those things are copyrighted. Copyright law thu
Read sourceAnalysis · doi.org
Unfair Learning: GenAI Exceptionalism and Copyright Law
# Unfair Learning: GenAI Exceptionalism and Copyright Law SSRN Electronic Journal. Published: 2024-01-01. Preprint. 2 citations. ## Authors - David Atkinson (The University of Texas at Austin): corresponding author ## Topics - Law, AI, and Intellectual Property - Intellectual Property Law - Intellec
Read source