Although generative AI is becoming more ubiquitous, adoption amongst news organisations can sometimes lag.

Sonali Verma, lead of INMA’s Generative AI Initiative and moderator of this week’s Webinar, noted that amongst news media companies she has talked to, internal workflow use cases are more common than consumer-facing ones: “They’re lower risk,” she explained. “You’ve got an editor involved who’s keeping an eye on what’s going on, and so many news publishers feel much more comfortable with that.”

Getting teams comfortable with embracing AI can be another story entirely, however. “AI literacy or AI fluency is something that everyone in the organisation should have, but in many cases, we’ve built these interesting tools, and there has not been widespread adoption of them,” Verma said.

During this week’s Webinar, How GenAI can help news media companies go farther and faster, INMA members heard from Nadia Kohler, head of the AI Lab at Tamedia, and Angélica (Momi) Peralta Ramos, data team leader at La Nación, about how AI is being integrated into their companies’ operations — and heard how other companies could get their teams on board.

Building the AI pyramid

Like most news media companies around the world, Switzerland’s Tamedia is leveraging AI to enhance its operations and services.

Kohler explained that the company’s approach to AI is structured as a pyramid, with information and training forming the foundation. It’s important to get the foundation right, because “you can build as many features and tools as you want, but if you’re not reaching your audience — in this case, the journalists, for example — it’s useless because they don’t understand why they should use it.”

The next step is evaluating and testing the AI features and tools, making sure they meet the needs of clients and journalists, and then finally, implementing the tools while constantly continuing to develop them.

“Make sure [the tools] are really close to the journalists’ existing workflow,” to make them more accessible and easy to use, Kohler advised.

Tamedia depends upon an interdisciplinary team, including members from editorial, product, technology, data analytics, and advertising to help integrate AI tools into various teams and ensure that they meet every department’s needs.

AI for editorial teams

The company launched its AI lab last fall and is two-thirds of the way into a three-month experiment using AI editors, who are journalists who help develop and integrate AI tools into the editorial teams. These editors provide valuable feedback on how to approach their colleagues with AI tools and have helped give insight into what newsrooms need from the tools.

“We also organised mini hackathons with our journalists to generate ideas: what are things during their daily task which are annoying, which takes a lot of time? And we try to experiment with them on building custom GPTs,” Kohler said.

The company has implemented several AI tools, including a prompt library, a connection between Slack and ChatGPT, a vector database for accessing content, a teaser generator in its CMS, a transcription tool, an SEO helper, and a text-to-speech function. It is also testing a streaming bot and has a daily service providing weather and traffic updates, among other things, parts of which are generated with ChatGPT.

“It’s about efficiency and making sure that [journalists] have the time to be creative and put more efforts in other things instead of doing things which take a lot of time and can be automated or at least helped with a little bit of help of AI,” Kohler said.

Since the company operates in both German and French-speaking parts of Switzerland, it faces unique challenges in implementing AI due to language differences. However, it is currently testing a service for its French-speaking audience to provide a daily overview of things like weather and real-time traffic, as well as suggestions for restaurants or what to watch on TV.

The information is generated in part by ChatGPT and because it is not continuously monitored, “it’s really important for us to be transparent so we have an info box where it explains how this text was created,” Kohler said.

Tamedia is also testing a streaming bot that allows users to ask questions and get suggestions based on internal content.

A chance for transformation

La Nación, the second-largest newspaper in Argentina, has embraced many different use cases for AI over the past few months, and Ramos shared how it is reshaping news media operations.

She traced the company’s AI journey back to 2016, when it created a data science lab and later added an AI lab: “This is the classic AI — automatization and NLP.”

Working with partners, La Nación used satellite imagery to detect different patterns and learn about the technology. By 2020, it was experimenting with speech-to-text and NLP, and in 2023, it embraced GenAI and LLMs.

“We consider it a new cultural opportunity to transform and to accelerate change,” Ramos said.

La Nación participates in several international collaborations, such as the LSE JournalismAI Collab; through that programme, it also worked on The Aijo Project, which measured gender equity in pictures and content.

Collaboration is vital, she said, because “it helps teams learn faster.” The gender gap tracker had been developed in English, and Ramos said her team wanted to do one in Spanish.

“So we developed this with NLP to analyse the diversity of voices in our reporting,” she said. “This tool consults APIs and databases and it generates the rate of different female, male, or non-binary sources in our content. So now we implemented this as a feedback tool to the editors.”

In 2022, La Nación created a tool called the Titulator, which was designed to improve and optimise headlines for social media networks. It based its results on the archives of the best-performing articles.

But after more than six months of working on it, Facebook disconnected media visibility from its platform and La Nacion had to scrap it. The tool still had something to teach them, though, Ramos said: “What we learned is frustration tolerance,” she said. “Because technology evolves or platforms change their algorithms or whatever you are studying — but the concepts learned remain.  Don’t get frustrated because what you have learned is good for new opportunities.”

La Nación also developed a tool called Cómo lo digo, or How Do I Say It?, which helps journalists know how to refer to certain groups or circumstances, such as race, mental health, disability, and poverty. The tool’s second iteration included GenAI to automatically detect the topic.

AI in real time

In 2023, La Nación wanted to monitor the presidential debates live: “We monitored expressions, sentiment, and words and topics using Spacey NLP and then also using Whisper OpenAI to do the transcription,” she said, noting the newsroom managed both the transcription and counting of frequently used words, phrases, topics, etc.

For the debates, Ramos said the team wanted to try a new feature — sentiment analysis. Using Amazon Web services, it conducted facial recognition sentiment analysis of candidates during the debate and counted the different positive and negative reactions. It then published an image of each candidate along with their results.

The headline clarified this was done with AI and that it was something they were testing to make sure users knew where the information came from, Ramos said.

La Nación has also done voice cloning by working with ElevenLabs to recreate former President Cristina Kirchner’s voice. The clone read tweets and letters in her voice. The quality was good and the company used it on social media, Ramos said.

Other new features for subscribers include text-to-speech that provides an audio version of every article and bullet summaries of all articles generated through ChatGPT.  

Breaking down a mega bill

One of the challenges faced by journalists was covering the Ley Bases bill project, a mega-bill that President Javier Milei sent to the Senate that was recalled in February.

“The first version had 600 articles and 180 pages,” Ramos said. “It modified many, many laws, and it was very impactful.”

A modified version contained more than 200 articles and new modifications before it went to Congress, and La Nación needed to compare the final version with what had been presented to the Senate. Using AI, it parsed the content in both PDFs and structured the information in a spreadsheet, which was then validated by journalists and data producers.

“The algorithm compared article by article and defined the similarity and established what remained and what had been changed,” she said.  

Internally, Ramos said the company is organising user groups and cross-functional teams to learn fast, share information with each other, and develop more products for AI. 

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