As generative AI models are introduced in newsrooms worldwide, StoryFlow’s David Caswell believes we are entering a second era of AI, one in which the “accessibility of tools and no technical expertise” is helping its prominence grow within journalism. 

Caswell, who founded StoryFlow as an innovative consultancy focused on developing AI workflows in news production, believes the trajectory of AI in journalism now reflects the scale of electricity’s impact globally in the late 1800s. 

“A kaleidoscope of projects and teams [are] exploring these AI models. As discoveries come along, the tools will get better and better,” he told attendees of the INMA World Congress of News Media in London on Thursday.

AI’s real-world examples

AI has gradually been introduced to journalism over the past decade, from Associated Press automating business news stories in 2014 to The Wall Street Journal using Narrativa’s large language model (LLM) Gabriele in 2021 to create stories using data from the U.S. and Europe’s financial markets.

Until 2023, AI applications were part of what Caswell called the first era of AI in newsrooms, where a lot of money was invested in building data warehouses and employing engineering teams to build in-house AI models. 

ChatGPT’s release in late 2022 triggered the “second era”— an “efficiency” phase in which newsrooms have started learning how to use GenAI and apply it to their workflows and audience-facing products, focusing on user interfaces and application programming interfaces.

Brazilian independent journalism organisation Agência Pública has created a “Reports to Listen To” product that reflects this new phase; the product converts text to scripts and inputs them into an AI version of one of their reporters, Mariana Simões, for audio news stories.

“Lots of automated newsroom workflows are emerging,” Caswell said, listing copy editing, newsletter generation, tagging metadata, search engine optimisation, text alert generation, and social media posts among the ways it is being used. 

“We’re now in a phase of how to use it productively.”

A live survey of attendees found that 70% are already using GenAI to produce transcripts and summaries for newsroom tasks, and 60% are using it for headlines and SEO suggestions for stories.

The Philippines digital news organisation Rappler developed an AI dialogue moderator for focus group discussions with its audiences and uses text-to-video generative AI to create video news stories, Caswell said.

Rappler’s video strategy is called “candy + broccoli,” where soft news is delivered in succession and broken up by more infrequent hard news stories, something aimed to help reach younger audiences.

Further examples of how AI is being used successfully in journalism include Dutch-based Persian language news platform Zamaneh Media, which produces automated newsletters, and Colombia’s investigative news organisation Cuestión Pública, noted for its ODIN project.

The ODIN model is one of the best examples of how to publish breaking news using AI, as it uses a data set with detailed information on prominent figures often reported in the news. This information is contextualised using a retrieval augmented generation (RAG) workflow and turned into a news story within 15 minutes.

Looking to the future

In two to five years, Caswell predicted, the “transition” phase will see AI in journalism explore “new infrastructure, processes, products, competition, and power structures.” Then, the next news information ecosystem will emerge with a “fundamentally different structure” to how it is now.

“We don’t know right now what the workflows and applications of the new functionality of AI will be,” Caswell said. “That’s the central challenge.”

X is already changing how it gathers news, using the AI chatbot Grok to provide digestible summaries: “X uses Community Notes to help with the validity of the newsgathering [from Grok] as this is X’s vision for the future of news.”

Other early signs of change have begun with democratised content production, with NewsGPT paraphrasing news from other news sources and Channel 1 AI in the United States launching this year, which will run 24/7 from GenAI news content.

Distribution platforms have begun assembling new ways of integrating the news experience for their audiences and personalising it, with Google’s Generative Search Experience competing with Amazon’s “Let’s Chat,” Perpelxity AI, and Inflection’s PI to provide news using predictive text.

Consumer adoption of AI has been “remarkable” and happening at a “massive scale” Caswell said, and he believes that news producers should now be use LLMs daily to train workforces effectively to take advantage of this consumer behaviour.

When asked how the industry can foster better use of AI and save money, he pointed to how Nordic countries are leading the way in AI collaboration in journalism.

“At the Nordic AI in Media Summit a few weeks ago in Copenhagen, 450 AI news professionals were freely sharing what they are doing,” he said. “If we emulate this cooperation and replicate it elsewhere in the U.S., U.K., and beyond, we will be better off.”

To develop a coherent strategy for the future, Caswell said flexibility and creating options to prepare for the future AI ecosystem will serve news organisations best: “A coherent strategy combines near-term efficiency with options for an uncertain AI-mediated ecosystem.

“The challenge is to apply the new capabilities of AI to deliver value to audiences, organisations and societies.”

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