If AI training data has a price for publishers, creators want one too

Patreon CEO Jack Conte isn't pulling punches on the AI industry's favorite legal shield. In a sharp interview with TechCrunch, he flatly called the fair use defense for training on creator content "bogus" — and the argument he built is harder to dismiss than the usual industry talking points.

Conte's logic is disarmingly simple. If Stability AI, Midjourney, and their peers can negotiate licensing deals with major publishers for curated datasets, they cannot credibly turn around and claim that independent creators' work somehow falls into a legal gray zone. The same companies paying Condé Nast or Getty Images for training data are scraping Patreon posts, DeviantArt portfolios, and newsletter archives without permission or compensation. The inconsistency isn't subtle; it's structural.

This isn't abstract grievance-mongering. Patreon hosts hundreds of thousands of creators whose work directly feeds generative models, often without their knowledge. The platform's business model depends on creators retaining control over how their work is monetized. When AI firms treat that work as raw ore for model training, they're not just exploiting a legal ambiguity — they're extracting value from a labor force they've deliberately excluded from negotiations.

Conte reframes what courts and commentators have treated as a copyright puzzle into something more fundamental: a business model choice. These companies have demonstrated they can identify rights holders, structure payments, and operationalize licensing. The decision to skip creators isn't technological necessity. It's cost optimization dressed in legal theory.

The timing matters. AI firms face mounting legal pressure from authors and artists, with cases working through courts that could reshape training data practices. Conte's intervention pushes the debate beyond individual infringement claims toward systemic market design.

What makes Conte's position strategically interesting is how it leverages the industry's own behavior against its legal arguments. Every licensing deal with a publisher becomes evidence that fair use isn't a principled stance but a selective discount applied to weaker negotiating parties. The double standard isn't hidden — it's documented in contract filings and press releases.

For Patreon specifically, the stakes are existential and opportunistic. The platform needs creators to believe their work retains value that can't be arbitraged away by model training. But Conte's broader point — that independent creators could and should extract value from their data — suggests a possible structural shift. If licensing becomes normalized, creator platforms become natural aggregation points for rights negotiations, potentially capturing a slice of data-market revenue that currently flows to publishers alone.

The economic implications for AI companies are substantial. Current training budgets assume near-zero marginal cost for acquiring creative content. Mandatory licensing at any meaningful scale would reshape cost structures, potentially favoring incumbents with deeper pockets and squeezing smaller model providers. This dynamic explains some of the industry's resistance: fair use isn't just legal strategy, it's competitive moat maintenance.

Whether Conte's framing gains traction depends partly on whether courts accept that willingness-to-pay-for-some constitutes evidence against fair use-for-others. That's genuinely uncertain legal territory. But the political and market pressure is building. Regulatory proposals in multiple jurisdictions already lean toward explicit training data transparency and compensation requirements. The fair use defense may survive litigation only to collapse under commercial and legislative pressure.

Conte's intervention matters because it connects individual creator grievance to systemic market critique with unusual clarity. The question it raises — why some content deserves payment and other content deserves extraction — doesn't require legal expertise to answer. It requires only noticing who gets to ask the question.