KorvaJoin the waitlist
Tech

AI Companies Should Share Revenue With Publishers

Investor Jason Calacanis says even 3% of AI revenue would be a lot of money to spread around.

AI image
This Week in Startups · Did the Midterms Destroy All Hope for Data Centers? | E2360

Anthropic has signed an agreement with the Associated Press to license election coverage, with plans for its chatbot to carry attributed, up-to-the-minute AP polling updates on a future election night. Investor Jason Calacanis says the hard questions on training, paywall scraping and outputs remain unsettled.

No financial terms were disclosed. It was Anthropic's first such agreement with an online publisher, while OpenAI already has more than a dozen, including Axel Springer, News Corp and The Atlantic.

Calacanis then offered his fix. “I think they should take 30% of their money and give it to training data companies and media companies.”

He illustrated it with a hypothetical trillion-dollar company making $100 billion. A 30% share would be $30 billion, he said, a sum "bigger than the entire movie industry."

Calacanis, a startup investor, argued the industry has cleaned up the simplest abuses but left the hard questions unsettled. He pointed to Anthropic's large settlement over pirated books, and to separate disputes over scraping behind paywalls, training on lawfully obtained works, and outputs that repeat protected material.

Even a much smaller share would matter, he argued. At 3%, or $3 billion a year, there would be a lot of money to spread around to magazines, newspapers, books and newsletters.

He imagined pooled payments that publishers and writers would line up to join. Newsletter platform Substack, he said, could make participation the default on an opt-out basis, with writers keeping 90% of their share and Substack keeping 10%.

He compared the idea to his own habit in Las Vegas of placing an extra chip for the blackjack dealer when he is winning. AI companies that are winning should tip publishers the same way, he said, which "would be a beautiful thing to do."

Questions people ask

What counts as training data?

Training data is the text, images, audio, and video used to teach an AI model patterns in language or pictures. Much of it comes from the public web, including news articles, books, and posts, which is why publishers say they should be paid.

Is it legal to train AI on copyrighted work?

Courts are still deciding. AI companies often argue that learning from copyrighted works is fair use, while publishers, authors, music labels, and studios argue that commercial training and near copies of their work require permission and payment. A few cases have settled, but the core training and output questions remain open.

How do existing publisher deals work?

OpenAI has signed more than a dozen public deals with publishers including Axel Springer, News Corp, and The Atlantic. The arrangements typically pay for access to current and archived articles and require attribution, though most financial terms are not disclosed.

More in Tech