Chatbots are increasingly becoming an in-house tool for news media companies looking for ways to use generative AI.

During this week’s Webinar, Phoebe Connelly, senior editor for AI strategy and innovation at The Washington Post, shared the company’s experience with developing and implementing its chatbot, Climate Answers.

Launched in July, the chat product encourages reader engagement while providing timely information on climate-related topics — and fits in with its broader strategy of leveraging new technologies.

Warming up to Climate Answers

The Washington Post’s journey into chat grew out of its Next Generation Initiative, which Connelly spearheaded. That initiative explored how to reach younger audiences and experimented with new technologies like SMS products.

The emergence of GenAI presented another way to reach new audiences, but in the beginning, it was unclear exactly what that would look like. To narrow down the possibilities, The Washington Post held a cross-company hackathon in March 2023 to let people “play around” with the technology, all with the younger user in mind.

That was followed by a period of rapid experimentation during which 17 AI-powered prototypes were launched in one year.

“We really focused on internal prototypes,” Connelly said. “We focused on not saying no. Anything that you could get excitement around and engineering resources for was on the table to build, even if we didn’t end up ultimately finding a use for it or it wasn’t as functional as we hoped it would be. And I do think that rapid prototyping led us to a lot of great places.”

While experimentation was ongoing, the team conducted a user study to understand audience attitudes and needs regarding AI. It found readers valued AI for its efficiency in summarising information and appreciated human oversight to ensure transparency and trust. Readers also desired a human touch in AI interactions, emphasising empathy and humour.

That helped inform the path ahead, Connelly said.

“I don’t think it’s enough for us to use generative AI just for generative AI’s sake,” Connelly said. “Really dialling in the question of when and where our users are going to get utility from a generative AI product is important.”

Talking about chat

Launching a chat product focused on a single topic allowed the team to learn and iterate against a focused set of information, she said. While the news organisation has covered everything in the news world from weather to sports to politics, the idea of climate change as a focused chat topic made sense.

“The Post had done a deep investment [in climate] in terms of reporters and editors. We did a big expansion of that section a few years ago, which meant we knew we had a lot of quality journalism,” she said.

With the creation of Climate Answers, they could quickly and easily answer users’ questions.

“We talked a lot internally about the idea that a user may have a question, the Post may have answered that question, but the answer to that question may come six paragraphs into a story. Or we may have published it two years ago, and they have the question right now in September of 2024.”

It was essential to “dial in” on the entry points for curious readers and determine what questions would be asked. Working with the climate change desk, they developed a pipeline of commonly asked questions and looked at popular related search terms.

The goal was to determine how to use the moment to let readers dig in deeper and provide them with the tools to get there.

Unveiling Climate Answers

The Climate Answers chatbot was launched in July and allows users to ask climate-related questions and receive concise, vetted answers. It features scrolling tiles with suggested questions that are revisited and updated periodically.

For example, a relatively new addition to the questions is, “What are Kamala Harris’ views on climate change?” Users can also input their own questions, and the chatbot provides links to longer FAQs for those seeking more in-depth information.

Connelly said they have seen some specific patterns emerge from users. First-time users often rely on the tiles for question prompts, but as they become more comfortable, they’ll begin to ask their own questions. Those questions, in turn, may appear in future tiles or prompts.

“At the end of every week we download the questions themselves and share it with our climate desk so they can see trends in the questions being asked,” she said. That allows them to look for spikes in interest or see if there is anything new that might be worth focusing on. In some cases, they may discover questions that need more focused answers — or one that isn’t answered at all.

With just a few months of usage behind them, the team is still learning where the greatest amount of engagement occurs: on the articles themselves, the homepage, or the app, Connelly said.

“We’re just monitoring to see over time where and when users actually engage and use Climate Answers in their journey.”

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