The Artifact Remembers: On Fair Use, Compensation, and the Architecture of Creative Value
There's a moment in Jack Conte's SXSW speech – delivered this week in Austin, read from what he himself called a manifesto – that deserves to be held up to the light. Not because it's surprising, but because it makes visible something that has been operating in plain sight.
If it's legal to just use it, why pay?
The question is rhetorical. But the answer it implies is not.
The Scene at SXSW
According to TechCrunch's reporting, Conte stood before the SXSW audience and did something unusual for a tech CEO: he named a contradiction. AI companies, he argued, claim that training models on creators' work constitutes fair use – a legal doctrine that permits limited use of copyrighted material without permission. Yet these same companies sign multimillion-dollar licensing deals with Disney, Condé Nast, Vox, and Warner Music.
The word Conte used was bogus.
Not illegal. Not immoral. Bogus – a word that suggests something is being performed, a logic that doesn't hold when examined closely.
This is the kind of moment worth paying attention to. Not because Conte is necessarily right about the law (fair use doctrine is genuinely complex, and courts are still working through its application to AI training data). But because he's pointing at something structural: the gap between what companies say is permissible and what they actually do when facing entities with legal resources.
The Two-Tier System
Here's the pattern Conte is describing: AI companies argue that scraping publicly available creative work for training purposes falls under fair use. This argument has been central to their legal defense in multiple ongoing lawsuits. At the same time, as CryptoRank noted, these companies negotiate substantial licensing agreements with major media conglomerates and entertainment corporations.
The result is a de facto two-tier system. Large rights holders with legal departments and negotiating leverage receive compensation. Individual creators – the illustrators, musicians, writers, and independent artists whose work also feeds these models – do not.
This isn't a bug in the system. It's the system revealing its architecture.
What's being naturalized here is a particular distribution of value: one where scale determines who gets paid, where the ability to threaten litigation becomes the threshold for recognition, where fair use functions differently depending on who's invoking it and against whom.
Conte's Position (And Its Complications)
It would be easy to read Conte's speech as pure advocacy – a platform CEO positioning himself to capture some of those licensing dollars for Patreon's community of over 250,000 creators. And that reading isn't wrong. Conte has a business interest in establishing that creators deserve compensation for AI training data.
But reducing his argument to self-interest misses something important. Conte explicitly framed his critique not as anti-AI or anti-technology.
I run a frickin' tech company. I accept the inevitability of change, and I feel agency in discovering my next path through the chaos.
Jack Conte
This is a more interesting position than simple opposition. Conte is arguing that the question isn't whether AI should exist or whether it will transform creative industries – it will, and he seems to accept this. The question is how value flows through that transformation. Who benefits? Who decides? What gets counted as contribution, and what gets treated as raw material?
When we plan for humanity's future, we should plan for society's artists, too, not just for their sake, but for the sake of all of us. Societies that value and incentivize creativity are better for it.
Jack Conte
The Interface of Compensation
Consider what's actually at stake in this debate. AI training involves ingesting vast quantities of human-created content – text, images, music, code – and using it to build systems that can generate new content in similar styles and forms. The models learn patterns, relationships, and structures from this data. Without it, they couldn't function.
The fair use argument treats this ingestion as transformative: the model isn't copying any single work, it's learning from the aggregate. The counter-argument – Conte's argument – is that this transformation doesn't erase the value contributed by the original creators. Their work is the substrate. The model's capabilities are built on their labor.
This is, at its core, a question about what counts as contribution. And that question has always been political, not just legal.
Think about how streaming platforms compensated musicians. Think about how social media platforms monetized user-generated content. Think about how gig economy apps classified workers. In each case, the architecture of compensation reflected choices about whose labor was visible and whose was treated as ambient resource.
AI training data is the latest iteration of this pattern. The question is whether this time, the architecture will be designed differently – or whether the same defaults will be naturalized before anyone notices.
What This Means for European Stakeholders
For policymakers and governance scholars watching this debate, Conte's speech offers a useful diagnostic. The EU has already moved further than the United States in regulating AI training data, with the AI Act and ongoing discussions about copyright exceptions. But the fundamental tension Conte identifies – between stated legal positions and actual market behavior – will persist regardless of regulatory framework.
For startup leaders and investors, the question is more immediate: what happens when the legal landscape shifts? Companies building on AI capabilities need to consider whether their training data practices will remain defensible, both legally and reputationally.
For researchers and foresight practitioners, the deeper question is about precedent. How this debate resolves will shape assumptions about creative labor, intellectual property, and value distribution for decades. The choices being made now – often quietly, often through licensing deals rather than legislation – are establishing the defaults that future systems will inherit.
The Artifact and the Argument
Conte ended his talk with a claim about human creativity:
Great artists don't play back what already exists. They stand on the shoulders of giants. They push culture forward.
Jack Conte
This is true. But it's also true that large language models (LLMs) – systems that predict appropriate outputs based on patterns in training data – are themselves standing on shoulders. The question is whether those shoulders belong to giants who've been compensated, or to millions of creators whose work has been absorbed without acknowledgment.
The artifact remembers what the discourse forgets. Every AI-generated image carries traces of the images it learned from. Every generated text echoes the writers whose words shaped its patterns. The question isn't whether this debt exists – it's whether the systems being built will make it visible or invisible.
Conte is betting on visibility. Whether the industry follows remains to be seen.
These questions – about creative labor, compensation architecture, and the defaults being built into AI systems – aren't going away. They're becoming more urgent as the technology matures and the economic stakes increase. For those working at the intersection of policy, technology, and culture, the conversation needs to happen in rooms where decisions get made. Human x AI Europe convenes in Vienna on May 19, bringing together exactly the stakeholders who need to be thinking through these tensions together. If the architecture of creative value matters to the work being done, that's a room worth being in.
Frequently Asked Questions
Q: What is the "fair use" argument that AI companies make about training data?
A: AI companies argue that using copyrighted creative works to train machine learning models constitutes "fair use" under copyright law because the models learn patterns from aggregate data rather than copying individual works. This legal doctrine permits limited use of copyrighted material without permission under certain conditions.
Q: Why does Patreon CEO Jack Conte call the fair use argument "bogus"?
A: Conte argues the fair use claim is contradicted by AI companies' own behavior: they sign multimillion-dollar licensing deals with major rights holders like Disney, Condé Nast, and Warner Music. If fair use truly applied, he argues, these payments would be unnecessary.
Q: How many creators does Patreon support, and why does this matter to the AI training debate?
A: Patreon supports over 250,000 creators. This scale gives the platform significant standing in debates about creator compensation, as it represents a large community of independent artists, musicians, and writers whose work may be used in AI training without payment.
Q: What is the two-tier compensation system Conte describes?
A: Large media corporations and entertainment companies with legal resources receive licensing payments from AI companies, while individual creators – illustrators, musicians, writers – whose work also trains these models receive nothing. Negotiating leverage determines who gets paid.
Q: How does the EU approach AI training data differently than the United States?
A: The EU has moved further in regulating AI training data through the AI Act and ongoing copyright discussions, establishing stricter frameworks around data use. The United States relies more heavily on fair use doctrine, with courts still determining how it applies to AI training.
Q: What did Conte say about the future of human creativity alongside AI?
A: Conte expressed belief that humans will continue making and enjoying human-created work despite AI advances. He argued that "great artists don't play back what already exists" but "push culture forward" – distinguishing human creativity from AI's pattern-based generation.