Newsrooms face AI control crossroads

Are MCP apps the wave of the future? I’m not sure. But I do know that giving LLMs access to and the ability to synthesize carefully controlled, curated content and data create enormous value.”

Alex Curley, founder and executive director of Semipublic

At the end of February, Anthropic made the unusual move of publicly rebuking the Pentagon during an active contract negotiation. In a statement published on its website, Anthropic laid out two proposed use cases the company found objectionable: Using its technology for fully autonomous weapons and mass surveillance. “We believe AI can undermine, rather than defend, democratic values,” CEO Dario Amodei wrote. “Some uses are also simply outside the bounds of what today’s technology can safely and reliably do.”

This was a stunning development, not just because the Trump administration moved to punish the private company soon after, but also because of the timing. All of this happened just days — hours, really — before the U.S. began taking military action in Iran. I don’t think there will ever be a clearer indication of how influential AI already is: If Amodei hinted at Anthropic’s importance to the military in his open letter, the White House’s swift response all but confirmed it.

On the same day that Anthropic published its rebuke of the Pentagon, the U.K.’s biggest newsrooms, including The Guardian, the BBC and the Financial Times, launched a campaign to advocate for proper attribution and compensation for news content used by large language models (LLMs). Their press release is a perfect encapsulation of every anxiety the industry currently has about artificial intelligence: No payment for the use of original journalism, endless model training on articles without permission, collapsing search engine traffic, and no say in how LLMs represent their news stories. And as the recent fracas over Grammarly’s unauthorized use of writers’ likenesses in their short-lived “Expert Review” tool shows, their concerns are completely warranted.

Lack of control over journalism’s representation, as well as the relatively few safeguards against bad actors, also kept newsrooms from fully embracing social media in the late 2000s, long after it had become widespread among their audience. However, by the time the Arab Spring ended in 2012, nearly every outlet in the country not only had a presence but also used social media as a core part of its monetization strategy. Nearly 15 years later, Facebook and Instagram are the first touchpoints for a significant portion of journalism’s audience, even for broadcast-focused outlets like public media stations. Why? It’s because people are already there.

The hard lesson journalism learned back then is that being late to the party means you have lost control over how your content is distributed and disseminated. Newsrooms should absolutely be demanding payment to license their content to AI companies while protecting how it’s represented by their models. Completely shutting off access to journalism for LLMs without an alternative solution, however, risks repeating that painful lesson again. More and more consumers are using AI to discover journalism in some way, so how do we maintain control in a way that’s beneficial for all parties?

One method is to strike a licensing deal with AI companies that imposes clear boundaries. In two major agreements last year — one between the New York Times and Amazon, the other between Axios and OpenAI — LLMs were restricted to only serving summaries and links to articles in their generated responses. This helps address the problems of fair compensation for use and misrepresentation in journalism, but it once again removes control over distribution methods from newsrooms.

Another approach I’ve been thinking about recently is using Model Context Protocol (MCP) apps. Launched earlier this year, MCP apps offer a new, interactive way for LLMs to access and respond with data they normally wouldn’t be able to. The benefits for news outlets are numerous and, until now, mostly unexplored. Not only can you require user authentication for the use of an MCP app, paving the way for monetization, but the presentation and representation of content, such as journalism articles, can also be closely controlled by the app provider.

To test it out, I created public media’s first official MCP app for AdoptAStation.org. It’s a simple tool that outputs basic data on the industry’s stations, such as the programs they air and how much federal revenue they lost last year, along with visual elements from the site. But the real magic happens when LLMs are given free rein to query the data and combine it with outside knowledge: I can ask for something complicated, like the stations that lost the most revenue last year in Texas, Hawaii and Maine that currently run “Fresh Air,” and get a well-structured output that includes outside context about the resultant stations. My tool references only a handful of data points; imagine how useful an MCP app with a newsroom’s entire archive could be to consumers, commercial clients, students and even journalists internally.

In testing Adopt A Station’s MCP app, I was also pleased to find that LLMs preserved (and, in some cases, overrepresented) the fundraising appeals from our hand-crafted station summaries in its responses. Normally, models return generic language about donating when prompted, and sometimes even reference the incorrect station. When using the MCP app, however, summaries almost felt like a prompt injection, preserving the original framing and language as if I’d prompted it in the first place.

Are MCP apps the wave of the future? I’m not sure. But I do know that giving LLMs access to and the ability to synthesize carefully controlled, curated content and data create enormous value. MCP apps support user authentication; I could imagine a future where extending access to paid subscribers could be a value-added service. Or perhaps newsrooms could create a free version that serves only short summaries of journalism written by humans, funneling users back to their main website. The point, as is ever the case in the news industry, is to be where the people are.

If you’d like to try public media’s very first MCP app, visit ai.adoptastation.org.

Alex Curley is the founder and executive director of Semipublic, a research nonprofit dedicated to building public trust in media through data, and the creator of AdoptAStation.org. Previously, he spent a decade at NPR, where he worked on product development, satellite-based audio distribution and editorial promotional strategy. He can be reached at alex@semipublic.org.