As online users move away from search and turn to chatbots, news media publishers should consider leveraging this feature to interact with their audiences. That was the message from Jodie Hopperton, INMA’s Product & Tech Initiative lead, during this week’s Webinar.

“Chat is so much more intuitive [than search], and it gives us a lot of opportunity,” she said during How to build a chat product for news. “It allows us to do more. It allows us to talk to our users and have our users talk to us and have more of a conversation. So this is a really good opportunity right now.”

Three guests joined Hopperton to share their experiences with building chat products: Steve Henn, entrepreneur-in-residence at Brown Institute at Stanford University, and Alex Goldmark, executive producer of NPR’s Planet Money podcast, who were responsible for building the Planet Money chatbot; and Bernd Volf, managing director, chief product officer and chief technology officer of Ringier MediaTech, who creating a structure for the company’s chat feature.

Throughout the one-hour Webinar, they shared what they had learned and offered best practices to consider.

It’s clearly a topic of interest to INMA members. Taking an online poll of attendees, Hopperton found that while a small percentage of members have already built their own chatbot, the majority are in the beginning stages of thinking about how to create and use one.

“A lot of people I’m speaking to are starting to talk about building out chat,” she said.

When NPR met GPT

Goldmark and Henn explained that when ChatGPT was released in November 2022, one of the big concerns was the problem of hallucinations, which occurred when it gave false information. The two men wondered if they could create a more trustworthy chat feature if it were trained on a trusted source — such as a journalism archive.

Since Planet Money has thousands of episodes, it provided them with a robust archive covering “just about every type of business case story,” Goldmark said.

The process began with discussing what a chatbot for journalism would look like and what attributes it would need. That list began with transparency and showed where the information came from and that it would be factual based on information only from Planet Money episodes.

“We wanted it to be accurate and give you true information, and we also wanted it to be humble and admit when it’s wrong or doesn’t know [the answer]. We’d rather have it say, ‘I can’t answer that’ than make up something.”

Finally, it was important that the chat feature would drive engagement with journalism by directing users to Planet Money episodes.

That engagement was important, Henn noted, because one of the fears that emerged with the introduction of ChatGPT was that it would answer questions without directing users to the source.

“We wanted to construct a system that made the chat experience more reliable but also grounded it in a trusted archive,” Henn said.

One challenge was finding information that used to be true but isn’t any longer — such as unemployment rates. Because news evolves and changes, it was critical to find a way to ensure the information hasn’t become obsolete. Goldmark and Henn also worked to create ways to eliminate false answers because the questions asked were inaccurate to begin with.

The ability to create a chatbot based on one organisation’s archives has tremendous potential for the industry, Henn said:

“One of the things we’ve been experimenting with … is what kinds of audience outreach could you do if you have the ability to search and generate things like newsletters? Could you make an extroverted bot that reaches out to the audience and then brings people back? Fundamentally, I think it’s really healthy for the industry to find ways to lean into tools like this.”

The chat business model

As Ringer MediaTech looks at the role of chat, Bernd Volf is looking at how to structure it within the media company. Ringier wanted to do more than experiment with the functionality of it; the company wanted to do something more purpose- and vision-driven, Volf said.

This happened in three steps:

1. The vision: “We took a huge step back and said, ‘OK, what is our vision for our chatbot and what should it be and what should it not be and what is the purpose?’” The vision that Ringier decided upon was to create a chatbot that would “offer conversational, tailored insights into global events and media, making the complex world more accessible and personally relevant to each user guided by established journalist values.”

2. The purpose: Determining that vision led to the next step, which was defining the chatbot’s purpose, Volf said: “This is something you should be very clear about and really think about deeply why you are doing this and what is the purpose. Why should anyone use it?”

3. The users: The third step in structuring the plan was thinking about the users — not just the current users, but the future ones. While user testing always plays a role in developing products, Volf said “we did it in a bigger scope because we believe that this technology will fundamentally change part of the world or how we interact or how we do stuff.”

As chatbots become more common in the future, usage will change, and Volf said they couldn’t assume just having a chatbot will be enough to engage users.

“So it’s really not only about if your bot has a great user experience, but does the purpose of your bot and the vision of your bot fit perfectly in the whole ecosystem of your users? Because if not, you have a great bot that nobody wants to use.”

Extensive planning went into this part of building a chatbot, with different projects and workshops followed by hackathons using different types of technologies. Volf said they spent more time than is typically spent developing a product because they needed to ensure it was meeting their vision and goals. Then, they revisited the viability of the project.

While viability is normally established earlier in product development, Volf said the expense of working with GenAI meant having to revisit viability.

“You really have to think about what kind of technical setup, what kind of product setup you are choosing that makes it affordable,” he said.

This affects such things as who has access to the bot, what kind of users are allowed to use the bot, how many answers the bot gets, how long its answers are, and more. It must also factor in speed and performance so users aren’t waiting for answers.

And last but not least, he said, companies need to think about how to monetise chat technology they develop:

“Because the worst situation that you can find yourself in is that you have a product, it works extremely well, and will bankrupt you in a few months. And those are situations that you can have with certain technologies that are here currently.”

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