Rede Gazeta in Brazil has faced digital transformation challenges familiar to many legacy media companies, completing its move to digital-only in early 2020.

In recent years, it created an innovation department and then added a dedicated project management team and an innovation hub. Now, it is going through a massive IT restructuring.

During this week’s Webinar, Eduardo Lindenberg, director of innovation, explained to INMA members how this ongoing change prepared the company to develop a new generative AI product. 

“I believe [generative AI] will be another very big game-changer for our industry,” he said. “It’s one of the things that will define our industry in the future. I think that … we have to be agile and adaptable, otherwise we’re not going to be here long, regardless of the size of our organisations.”

For 19 years, Rede Gazeta has published the Yearbook, an annual almanac of the state’s social and economic data. For 2023, the sales team asked for a more innovative approach. So, Rede Gazeta looked at how to leverage the emerging GenAI technology, and that ultimately led to developing ConfIA, a news assistant that allows users to ask questions about content in the Yearbook’s database.

Using existing technology and working with a tech partner, Rede Gazeta built its own front end and back end, then provided the content that would be used to answer specific questions.

Lindenberg quickly pointed out that it’s not a chatbot but a news assistant. ConfIA is trained on the 2023 Yearbook and only provides information from that document.

The project, which is financed by sponsorships, will be done once a year and is not a main feature on the Web site. But it offered confirmation that the department is heading in the right direction and gave the company a good opportunity to test how it worked.

“We were very happy that we were able to provide good proof that from a technical standpoint, we are on the right path without a loss. The project itself for the Yearbook had a really good revenue year. And the direct cost of using this tool was very small considering the total revenue and costs of this project.”

How it works

When a user opens ConfIA, the bot suggests three questions to help get them started. That part of the process is still a work in progress, Lindenberg said: “Not all the questions are very good [yet]; sometimes it repeats two of them. So not everything is 100% right, but it’s right enough at this point.”

ConfIA is only trained on questions about the Yearbook, so if people ask questions about anything else, it replies that it doesn’t have that information. When it does provide information from the Yearbook, it gives the specific page where the information was taken from.

“We think this is something that’s very important because if people want to check the answer, they will know exactly where to check in the Yearbook,” he said. That will be important in the future as people will be able to search for information within the site and find or fact-check the accuracy of information.

While ConfIA is not always accurate, it does not hallucinate, either. Because the large language model was trained from a PDF without additional information or data, it sometimes misreads things like charts and tables, Lindenberg said. However, the fact that it doesn’t hallucinate is a plus:

“It’s not going to invent or be creative with answers that are not within the content that we provided,” he said.

What worked (and didn’t)

The project to build and implement the news assistant was completed in three weeks. Lindenberg said the commercial success proved its value, showing a 68% growth in revenues compared to the 2022 Yearbook.

“The cost of implementing this small technological test was very small and we were very happy and confident that we can move forward with [creating] an assistant product [with] other ways of interacting with users,” he said.

About 20% of the questions asked to ConfIA were not answered, which taught the team that having 100 or 200 pages of content was not enough. However, about one-third of those unanswered questions were attributed to people asking the wrong questions, which he said is an issue for everyone.

“Not just media, but in general, whoever’s providing GenAI services will have those issues because people need to understand [more] about prompt engineering.”

For example, in some cases there were “questions” of just one word, which wasn’t enough.

The audience reach and engagement was below what had been projected, but he said that is related to not having a lot of content and not making significant marketing or distribution efforts. However, that hasn’t dampened enthusiasm for the future.

“One of the things that we would like to assume later on is that if we have this sort of experience, then people will stay longer because they’ll be able to converse with the tool rather than just asking one single question, reading [the answer] as a bullet point, and then moving on somewhere else.”

Providing more data to the model will make it more valuable to users and also will reduce errors, he noted.

What the future holds

Rede Gazeta is looking at making ConfIA a general news assistant that will be available to answer questions about the content across the Web site. That could even include ways to be useful to advertisers, too.

The team is looking at adding a “like” or “dislike” button so users can provide feedback that helps with reinforced learning and is looking at using audio answers.

“We know the [use of] audio is growing, but we haven’t tested much with that,” he said. “And we know that the generative AI audio should be much more interesting than the other tools you had available one or two or three years ago to provide accessibility to our content.”

Adding advertising and marketing services using GenAI is being studied, including how the news assistant could help users interact with advertising content, and from an editorial perspective, it could be integrated into the CMS to help reporters write better stories: “[it could] interact with our own database of content and suggest different things that we have had within our news in the past so that we can add more value to the stories.”

The success of implementing a project like this one, Lindenberg said, lies in creating a multidisciplinary team: “You have to have people with different perspectives so that your product will be interesting and move forward in a sustainable way.”

However, the biggest change must be in the company mindset.

“Regardless of the size of your organisation, agile thinking is almost like a standard for everyone. It’s still very hard to do so, especially when it’s cutting-edge technology such as GenAI. Start small, make something interesting and valuable to users and then improve it gradually,” he advised. “It is a complex journey, and there’s no silver bullet.”

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