The Training Filter Won't Work
Copyright enforcement is moving from AI training to output control. The reason: you can't train models on the open internet without learning copyrighted material, even from legally posted sources.
Consider Sonic the Hedgehog. Fair-use images appear in reviews, commentary, news articles. Product photos show licensed merchandise. Forum discussions describe the character. An AI trained only on legal data still learns what Sonic looks like—the design, colors, aesthetic—from millions of non-infringing references.
This isn't theoretical. The pattern mirrors decades of informal copyright tolerance. Fan art technically violates derivative work rules, but enforcement has been discretionary and human-scale. Post it on Instagram? Probably fine. Sell it? That crosses the line.
AI removed the constraints that made selective enforcement workable.
Why Input Audits Fail
Proving training contamination is nearly impossible at scale. The New York Times demonstrated this by engineering prompts to reproduce exact articles through next-token prediction—a resource-intensive attack that OpenAI calls adversarial prompting. Declaring an entire model "tainted" because some training inputs were improper pushes copyright into uncharted territory. Unlike trade secret law, copyright remedies harm through damages, not system destruction.
The Bartz v. Anthropic $1.5B settlement over 500,000 pirated books signals the shift away from unvetted scraping. But enterprises face a harder problem: compliance when legitimate training still absorbs copyrighted patterns.
The New Enforcement Layer
Pressure now moves to generation and distribution. Major model providers already impose content restrictions—LLMs block certain outputs where Adobe Illustrator doesn't. This treats models as active participants in creation, not neutral tools.
For enterprise tech leaders, this means:
- Input risk (training data) remains murky despite vendor claims
- Output risk (infringing generations) becomes the compliance point
- Vendor indemnification clauses matter more than training transparency
- Global requirements diverge (EU AI Act Article 50 transparency vs. India's labeling rules)
Divided court rulings on fair use loom in NYT v. OpenAI and Getty v. Stability AI. A Midjourney win could accelerate unlicensed training. Studio victories compel licensing models. Either way, procurement teams need legal review of generation controls, not just training audits.
The fine print matters here. Copyright law assumed human-scale creation. AI broke that assumption. Enforcement adapts by moving downstream.