Administered by the U.S. Copyright Office and codified under 17 U.S.C. §107, the fair use
doctrine establishes an equitable four-factor balancing test that limits copyright owners’
exclusive reproduction rights. After landmark Federal Circuit and Supreme Court rulings
including Google LLC v. Oracle America, Inc. (2021) and Andy Warhol Foundation v. Goldsmith
(2023), U.S. courts have reshaped standards for AI model training data extraction, a
high-risk practice for global tech firms. Unlike commentary or educational fair use, mass
scraping of copyrighted texts, images and audio for algorithm training faces heightened
judicial scrutiny. This case analyzes an AI developer federal copyright infringement
lawsuit, unpacks updated four-factor judicial benchmarks, and delivers standardized
compliance frameworks for cross-border artificial intelligence enterprises.
A Singapore AI generative image startup launched a commercial visual creation platform
in 2025. To build its neural network training dataset, the company crawled millions of
protected paintings, photography works and graphic designs from U.S. artist websites,
stock platforms and digital art communities without obtaining individual licensing
agreements. The firm argued mass data scraping qualified for fair use under §107, claiming
the AI model delivered transformative new visual outputs distinct from original source works.
A collective of U.S. independent illustrators and photographers filed class-action copyright
infringement in a California federal district court. During bench trial, the judge applied the
updated post-Warhol four-factor test and ruled against the AI company on three core grounds.
First, the defendant’s use was primarily commercial revenue-generating, weakening
transformative purpose arguments. Second, the scraped materials constituted complete,
substantial portions of each original creative work, far exceeding limited quotation standards.
Third, the AI platform directly competed with the original creators’ print, licensing and
commissioned artwork markets, causing tangible revenue erosion. The court issued a
permanent injunction ordering full dataset deletion and awarded statutory damages totaling
$148,000 to the plaintiff artists. The startup was forced to rebuild training libraries exclusively
from public-domain and fully licensed visual materials, incurring massive development delays
and financial losses.
First, the revised four-factor fair use balancing standard (post-Warhol precedent) governs all AI
data scraping evaluations:
Purpose and character of the use: Nonprofit educational criticism gains strong fair use
weight; commercial AI training carries a heavy presumption against fair use, even if outputs
appear transformative. Courts distinguish pure research models from paid commercial generative
tools.
Nature of the copyrighted work: Original expressive artistic, literary and audio creations
receive stronger copyright protection than factual databases or functional code, lowering fair use
eligibility for visual and literary training sources.
Amount and substantiality of the portion used: Full reproduction of complete creative
works rarely qualifies as fair use; only brief, excerpted quotations for analysis meet judicial
thresholds. Bulk full scraping of thousands of intact works fails this factor.
Market harm to the copyright holder: If AI platforms substitute creators’ original licensing,
commission or print sales, courts will reject fair use defenses entirely. Secondary derivative
market harm is also weighed heavily per the Warhol ruling. Second, transformative use is no
longer a dispositive defense. The Supreme Court clarified that mere technical alteration of source
content does not automatically satisfy fair use. Even modified AI outputs cannot override market
substitution damage or full-scale mass copying of creators’ works. Third, public-domain
materials and licensed content form the only safe training pools. Works with expired copyright
terms or comprehensive blanket licensing contracts are exempt from infringement claims;
unlicensed scraping of living creators’ recent works creates near-certain litigation risk. Fourth,
USCO registration strengthens creators’ infringement remedies. All participating artists completed
timely eCO copyright filings, allowing them to claim maximum statutory damages. Unregistered
copyright holders can only recover provable actual lost profits with no access to statutory damage
awards.
The evolving judicial interpretation of the §107 fair use four-factor test, shaped by landmark
AI and visual art Supreme Court rulings, eliminates blanket fair use immunity for commercial
generative AI training data scraping. This class-action artist infringement case verifies that
unlicensed mass extraction of complete creative works for paid algorithm platforms fails most
fair use balancing factors and triggers costly injunctions and statutory damages. For
cross-border AI startups and global technology corporations developing visual, text or audio
generative tools, licensed data sourcing, limited excerpt research and proactive creator
opt-out systems are mandatory compliance measures to avoid U.S. federal copyright litigation and
permanent platform shutdown risks.
Hyperlink List：
● USCO Circular 21 Official Fair Use Four-Factor Full Explanation：
https://www.copyright.gov/circs/circ21.pdf
● Cornell Legal Information Institute Andy Warhol Foundation v. Goldsmith Supreme Court Full Judgment:
