Products · 25 May 2026 · 2 min read

OpenAI inks content deal with Grupo Folha and UOL

OpenAI's new partnership with Brazilian media giants Grupo Folha and UOL signals a clear strategy to feed its models with trusted, attributable news sources.

Pen-and-ink illustration: two distinct, interlocking gears being slowly. For the story "OpenAI inks content deal with Grupo Folha and UOL".
— Pen-and-ink illustration: two distinct, interlocking gears being slowly. For the story "OpenAI inks content deal with Grupo Folha and UOL". —

What happened

OpenAI announced a strategic partnership with two of Brazil's largest media conglomerates, Grupo Folha and Grupo UOL. The deal gives OpenAI access to content from their publications to train its models and ground its responses. In return, content from their outlets will be attributed within ChatGPT. This follows similar deals OpenAI has struck with other major publishers, including the Associated Press and Axel Springer, aiming to integrate high-quality, real-time information into its products.

How the room's reading it

The consensus among industry watchers is that this is a defensive and necessary move. Publishers see these deals as a long-overdue revenue stream, finally getting paid for content that has long been scraped to train models. For AI developers, it’s a straightforward play to improve factual accuracy and reduce hallucination — especially for news and current events. Legal analysts frame it as a smart de-risking strategy. By licensing content directly, OpenAI sidesteps some of the thorny copyright battles that have plagued the industry, creating a more defensible — if more expensive — data supply chain.

Sailfish's take

We see these content deals as necessary but insufficient. Licensing content from trusted sources is a smart move for OpenAI — it reduces legal exposure and provides clean, structured data for training. But for builders, it doesn't change the fundamental challenge of working with LLMs. The model still synthesises information, and attribution is just a link back to a source. It doesn't guarantee the model has correctly interpreted or summarised that source's content. We're still a long way from models that can truly reason over source material. This is a good step, but we wouldn't ship a product that blindly trusts the model's summary, attributed or not.

Our take — your read?

Be the first to weigh in.

Sources
— END OF DISPATCH — Products