Meta is considering new paid deals with news publishers for AI training data (2024)

The teams discussing this access to news content include leaders on the partnerships, product, and legal teams. Meta may need such access to make its generative AI tools, like Meta AI, more effective for users and more competitive in the increasingly crowded market of generative AI search tools and chatbots, the people said.

"Meta may have no choice but to pay someone," one of the people familiar said.

Meta has not formally approached any news outlet about licensing or otherwise accessing content, one of the people familiar said. Internal talks are still at an early stage. Yet, if the company decides to do so, any resulting deals for access to data for model training would be separate from past agreements in which Meta paid publishers to host links to their content on its sites. Over the last 18 months, Meta has made a sharp turn away from its previous dealings with the news industry, as BI reported, effectively eliminating a $2 billion budget for its News division just last year.

A Meta spokesperson declined to comment.

Meta CEO Mark Zuckerberg claimed earlier this year that his company has its own data for training its Llama large language model that is larger than Common Crawl, a massive set of scraped web data used widely (including by Meta) for AI model training.

Still, there is concern internally about the quality of that owned data, one of the people familiar said. User posts and comments on Facebook or Instagram are not necessarily the type of high-quality training data that generative AI chatbots and search tools need to generate quality outputs. Books, news articles, and essays are.

Meta could find itself on the back foot again, with outputs that fall behind rivals like Google and OpenAI, if it chooses or is forced to rely more on its own data. Previously, Zuckerberg said he did not expect the generative AI boom.

Not long after generative AI burst onto broader public consciousness almost two years ago with the launch of the ChatGPT chatbot, news outlets and other websites started blocking automated bots deployed by Common Crawl and OpenAI to scrape their content for free constantly. The US Copyright Office is considering new rules to cover generative AI. Without free and constant access to news publisher content, responses from Meta AI to user prompts about current events may become more limited, outdated, or incorrect.

Major rival tech companies now fiercely competing in generative AI have already entered into deals with news publishers and media outlets for more access to content to be used as model training data. News Corp. signed a deal with Google. The Financial Times signed on with OpenAI, as did The Associated Press, Dotdash Meredith, and Axel Springer, the parent company of BI, Politico, and several publications in Europe. OpenAI's largest investor is Microsoft, which has its own partnership with Axel Springer.

Still, many publications have yet to sign with any AI company, including The New York Times, which sued OpenAI after failing to agree to terms on a licensing deal. The Times also had Common Crawl remove data it scraped from the publication. Most news publishers are open to licensing deals, if only because "something is better than nothing," said one of the people familiar with Meta's considerations.

Are you a Meta employee or someone with a tip or insight to share? Contact Kali Hays at khays@businessinsider.com or on secure messaging appSignal at 949-280-0267. Reach out using a non-work device.