So much of what news organisations focus on when it comes to AI is how it affects the daily work and decisions made inside the business. During INMA’s recent Generative AI Town Hall, attendees heard a different perspective from Sam Guzik. 

Guzik, senior director of product at New York Public Radio and foresight affiliate for Future Today Institute, discussed the spaces outside of a news company’s direct role, how AI will affect consumer behaviour, and how audiences are interacting with tech more broadly.

“If we only focus on the decision we make, if we only focus on our business model, we risk missing some changes that are outside of that scope, outside of what’s there,” Guzik said.

An interesting perspective on AI technology is that today’s AI is the worst we will ever experience. The tech gets better every day, every week and every month. Believe it or not, ChatGPT only launched at the end of November 2022 and hasn’t yet been out for two years.

We also don’t yet know the biggest impacts of AI. 

“The tasks that AI and LLMs (language learning models) in particular are good at won't change,” Guzik said. “The ways those tasks inform how consumers use computers and interact with each other may change.”

AI is changing the way news companies interact with technology. Guzik highlighted three places to focus on when it comes to shifts in AI in the year ahead.

1. Multimodality will get its ChatGPT moment

Multimodality is the ability of a model to understand different types of inputs at the same time like text, images, and videos. Guzik used the example of Notebook LM, the product Google released earlier this year, that can take uploaded data sources and build summaries from it.

“So it’s really a great tool for taking a long 300 page PDF and trying to ask some questions about it or taking all of the reporting you’ve done on a topic and trying to synthesise it,” Guzik said.

Notebook LM had a moment of viral success when it began generating podcasts. Guzik used this as an example of how surprising the models can be when they are multimodal. Tech companies are leaning on multimodality to shift how news companies and consumers interact with devices. 

Apple came out with an example of using a model that is small enough to run on mobile and understands the different actions a user may take on a given screen. If a model like this begins to work reliably, it makes it easier for UX to fade into the background and AI can begin operating the device in place of the user. 

The use of this model brings up questions companies should start asking like, “What are the future expectations of a consumer? How seamless do they expect their technology to be? Do they like to browse the internet or do they let a language model do that for them?”

Multimodal models offer both an opportunity and a threat. 

“The opportunity here is what tasks might we be able to automate with multimodal models,” Guzik said. “The surface of the work that we can do with AI gets a lot bigger if we can successfully find a way to understand multiple types of input.” 

On the other hand, what happens if audiences and the workforce reject click-based UX? If clicking and navigating to a Web site becomes something people are not willing to do or if in 10 years staff is not accustomed to going to a portal to do that work, what shifts do news organisations need from their processes?

2. The “fight for content” battle heats up 

The next area Guzik wants companies to focus on for the future is actually fighting for their future in an AI-based world. Right now, organisations seem to be taking two different paths.

“There are people who are taking the path of trying to licence their content, trying to generate a little bit of incremental revenue off their archives — and we’re seeing people trying to fight to block AI crawlers, trying to sue and use the law to sort of fight back against training,” Guzik said.

The place to watch here is around model collapse.

“When you train an algorithm or you train an AI system on generated data, we’ve actually seen that those models start to perform worse than you’d expect. For some reason, the nature of the performance starts to collapse,” he said.

That’s a real risk for tech companies and something that can really heighten the value of archives at news organisations.

“If we are creating high-quality, human-generated content and models can’t perform with generated content, there’s an opportunity space because we become suppliers of content that continue to train this work,” Guzik said.

A lot of tech companies are trying to fix the issue because they understand this is a constraint they’re facing. 

“So very, very smart people are trying to break this curse of synthetic data to be able to train models that perform well even when you don’t have real data,” he said.

If model collapse continues, the licensing deals media companies have may hold up, but if technologists solve the model collapse, a media company’s value diminishes. 

“They need our data much less if they don’t need them to be training foundational data,” Guzik said.

The opportunity here is around how companies can position themselves if the tech companies need more real data. Can news companies structure the data to make it more valuable for the purpose of licensing it? The risk is if content licensing loses its value, how can that threaten a publisher’s bottom line?

“What decisions are we making based on those revenue streams that may or may not hold up three years from now, five years from now, if the underlying economics change?” Guzik said.

3. Care about the AI supply chain

Regulation will play a big role here considering models trained in the United States versus Europe or China will all have very different regulations they were trained under, and that can change how the model behaves.

“We need to start thinking about, ‘Where was your model trained, who trained it, and what decisions got baked in there,’” Gusik said. “This is going to impact the future of this technology, and so one of the ways that this is going to matter is we will have a choice to make in terms of where we source AI models.”

Companies should also compare open-weight vs. closed-source models.

Closed-source models like ChatGPT or Claude are where the tech company controls everything and news companies are just using their interface. Open-weight models are where organisations can customise the tech and choose how it performs by doing some of its training.

The opportunity here is there may be ways to lower acquisition costs with open source models. The threat is organisations may not have the proper people in place for this.

“We may need to bring different types of technical expertise into our organisations to make these kinds of decisions,” Guzik said.

What should we do with all of this?

Guzik believes the way forward is for news organisations to own their audiences and not just rely on platforms and algorithms. 

“Remember that AI is not a strategy,” Guzik said. “AI is a technology. It’s a tool that helps us make decisions and helps us act, but it’s not a strategy that’s going to save our business.”

Organisations must remain focused on what their audiences want and not get distracted by new tech.

News organisations should be “focusing on our community and where we can build real relationships and a sense of place for people and most importantly focused on our mission,” Guzik said. “Focus on what journalists do best and telling great stories that inform communities and hold the powerful to account.”

And AI should be used to solve real problems. Companies should be selective and shouldn’t create problems for AI to solve. Companies should come together to own the future and not let Big Tech dictate the decision-making.

“The work we do is too important to hope it’s going to be OK,” Guzik said. “The work we’re doing matters, and I think we can keep doing it long into the future if we make the right choices.”

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .