AI licensing is a stopgap, not a media strategy

Earlier this week, I attended the AI and Journalism Summit organized by the Brown Institute for Media Innovation. I got to be on the 44th floor of the iconic Hearst Tower, looking out over Central Park, with representatives from leading news organizations and OpenAI in the room.

A huge thanks to Mark Hansen for convening this meeting at the top of the world. It felt like a summit between two distinct eras of media history.

I learned how fast our conversations have changed. Just a year ago, the industry was asking existential questions (“Will AI replace us?”), silly ones (“There should be a model only trained on journalism!”), or small tactical ones (“AI can help us translate this article or transcribe this meeting!”).

This time, the room focused much more on execution. The technology is maturing. We discussed the Model Context Protocol (MCP), which finally lets us distribute content with control directly into chat interfaces. We looked at OpenAI's App SDK, which is shaping up to be the “App Store” of the AI era. We explored voice agents that are expanding the very surface of storytelling.

We are moving from chatting with AI to integrating it; from static inputs to dynamic actions.

It seemed that while the technology has evolved, the business strategy has lagged. We are still trying to fit the square peg of AI into the round hole of the “content business.”

At Gazzetta, we work in contexts like Iran and China, places that are ahead on the disruption curve, where AI intermediation and information control are acute realities. We don't have the luxury of moral signaling without effective utility. We build for the world as it actually is.

Looking at the industry through that lens, I worry that publishers are missing the structural shift. Right now, the news industry views data licensing and philanthropic hand-me-downs as its lifeboats. We are signing deals to feed our archives into ChatGPT or Claude. I left the gathering convinced this is a path with diminishing returns.

A static archive cannot replace living inquiry.

The challenge with licensing isn't just that it sells our past work by the kilo. The challenge is structural.

When we license content to create LLMs that simulate journalism, we risk conflating the verb of journalism (the active work of verifying, interviewing, and investigating) with the noun (the published article). By feeding archives to AI, we reduce a living process to a static product frozen in time.

When we transfer the label of “Journalist”, a person with a time-bounded understanding of the world and specific sociocultural expectations, to a machine that has its state frozen by design, two complex issues arise:

1. The authority transfer: The meaning of the label “Journalist” transfers cleanly in the minds of most users to the AI system, leading them to assume the machine is doing what a journalist does. It is not; it is doing something distinct.
2. The “shape of yesterday”: The machine is trained on output, not on the process that produced the output. It will inherently try to make the events of today match up to the shape of yesterday's journalism.

The machine has structural limits.

This isn't just a philosophical problem; it is a technical friction that plays out across three distinct levels of the AI stack. Even as tools like MCP and Voice Agents make the interfaces smoother, the underlying models still suffer from limitations that destroy their competitive appeal.

1. The corpus of training (The archive) This is where the “Shape of Yesterday” lives. It is a static library. Western models are structurally biased toward English-language, open-web data; Chinese models are biased toward state-sanctioned narratives. In both cases, the model is looking backward. When a newsroom sells its archive, it merely reinforces the model’s tendency to project stability onto a world that is in flux.

2. The middle stack rules (The guardrails) This is where the “authority transfer” is engineered. These are the hidden instructions that tell the model how to sound authoritative. We see Western models hedged with safety refusals that can stifle inquiry, and Chinese models sterilized by political red lines. Both mimic the voice of trusted journalism without doing the work, creating a “safe” sounding answer that may be disconnected from ground reality.

3. The agentic search rules (The action) While OpenAI's SDKs allow for new applications, most “agentic” workflows today function as advanced retrieval systems, like fetching top results from Google or Baidu. They often lack the journalist's heuristic for “triangulation,” checking a government claim against a leaked document and a local witness. Without this, the agent risks automating the consumption of noise.

We offer what compute cannot.

The vast, vast majority of the world is not defined by the view from the 44th floor. The majority of the world operates in the complexity between these three levels, where the corpus is outdated, the guardrails are biased, and the search tools are shallow.

This is where our opportunity lies. Tech companies have massive compute, but they have inevitable blind spots.

We possess three capabilities that are structurally impossible to automate:

• Uncommon access: AI can only learn from what is digitized. It cannot interview a confidential source in a parking garage. During Iran's June 2025 internet shutdown, we conducted research with 900 people inside the country during a “stealth blackout.” No crawler could reach that data.
• Verification under pressure: Standard AI benchmarks work fine for static facts, but they break down when reality is being manipulated. In adversarial conditions, AI models are training on the shape of yesterday and struggle to distinguish new viral fabrications from hard truths.
• Trust signaling: When we issue a correction or disclose a source, that isn't just ethics, it is really precious metadata. It signals to a system: “This information has been vetted in a process of questioning and intellectual humility.”

We provide the ground truth.

The case for these partnerships is about increasing competition among models. Claude, ChatGPT, Gemini, DeepSeek etc. are competing for dominance. The models that hallucinate less will win, and we are the key to that victory.

Instead of one-off payments for archives, we should build ongoing partnerships that give specific models a competitive edge:

• Ground truth validation: Providing the “answer key” for high-stakes, obscure topics to systematically improve model performance where automated evaluation fails.
• Exclusive verification layers: If a model can signal “This answer was verified against the Reuters infrastructure,” that is a concrete market differentiator.
• Long-tail stress testing: Providing intelligence on model performance that Silicon Valley's dashboards will never capture. A model that demonstrably works better for a dissident in Tehran or a reporter in Minsk has a powerful story to tell.

Don't optimize for the printing press.

This brings us to a choice we have as media operators. Currently, “innovation” in our industry often looks like integrating AI into operating systems while measuring success with the metrics of the 2010s and early 2020s, downloads, page views, and impressions.

We may want to view Google Search and social media traffic as the printing presses of that era, artifacts of a previous distribution model. If we make the mistake of optimizing for those old machines, we risk a double failure: selling out the trust of the relationships we have built, and losing our utility as a source of new information.

There is a sustainable path forward. We must build a multi-layered publishing strategy that incorporates AI without surrendering to it. This means doubling down on our competitive edge in extracting new information from the margins where the models are blind while simultaneously building human kinship around the processing of that information.

The practice of journalism has a future as a craft, distinct from the artifacts we see today. Paper and print will still exist. Articles will be written. Social posts will continue to scroll. But the locus of value will shift.

My guess is that in ten years, an entirely new group of people will gather at the top of a skyscraper and look down on Central Park. The choice we make now between protecting our past or building our future will determine who is in that room.