I’m trying to understand whether AI companies should pay creators, publishers, or website owners for training data used in AI models. I keep seeing debates about copyright, scraped content, licensing, and fair use, and I need help sorting out what’s actually fair and legal. I’d like clear opinions on AI training data compensation, content rights, and how this could affect creators and tech companies.
Short answer, yes, in many cases they should pay.
If a company trains on copyrighted books, news articles, music, art, or paywalled databases, the creator or publisher deserves compensation. The model got value from the work. The company built a product from it. Pretending scraping makes it free feels weak.
I’d split it like this.
Public domain, no payment.
Open licensed data, follow the license.
Private or copyrighted data, pay or get permission.
Personal data, do not touch without consent.
Fair use is the fight. AI firms say training is transformative. Rights holders say it copies at scale and hurts their market. Courts are still sorting it out. We already saw lawsuits from The New York Times, authors, and image companies. So this isnt settled.
Practical fix. Create opt-in and opt-out registries. Standard licenses. Collective payment systems, like music royalties. Audit trails for what data went in. If your content helped train a commercial model, you should have a way to get paid. If AI companies want trust, they need cleaner data deals, not shrugging and saying the internet was there.
Mostly yes, but not in the super simple ‘every scraped token = a paycheck’ way people frame it.
I agree with @waldgeist on the broad moral point: if a company builds a commercial model using other people’s work, acting like the source material had zero value is kinda absurd. But I part ways on one thing. I do not think every use of copyrighted material for training automatically requires payment. The internet would become a legal parking lot overnight.
To me, the real split is not just public vs copyrighted. It’s substitution and leverage.
If the model is trained on your work and then can reproduce it closely, mimic your style on demand, summarize paywalled reporting, or reduce demand for the original, yeah, compensation or licensing makes sense. Same if the company used clearly restricted datasets, archives, or stuff behind terms that said ‘no scraping.’ That’s not a grey area morally, even if lawyers fight over it.
If the use is more distant, non-expressive, and not really a market substitute, fair use gets stronger. Search engines already normalized a version of this logic years ago. AI is similar in some ways and very not similar in others. That’s why this gets messy fast.
What I think probly makes more sense than ‘pay everyone for everything’ is a tiered system:
- Public domain = free.
- Open license = honor the license, obviously.
- Commercial copyrighted corpora = licensing deals or collective bargaining.
- Individual creators = opt-out by default is weak, there should be usable consent tools.
- High-risk categories like news, books, music, art, and personal data = stricter rules and auditing.
Also, website owners are not always the same as rights holders. That’s an important wrinkle people skip over. A platform hosting content should not automatically get paid for work it does not actually own.
My basic take: AI companies should pay when they are clearly extracting commercial value from protected work in a way that competes with, imitates, or bypasses the original. If not, then maybe not. Blanket ‘everything is theft’ is too sloppy, but so is ‘it was online, so its free.’ Both sides overstate it tbh.