Can news outlets build a ‘trustworthy’ AI chatbot?

News publishers have jumped headfirst into artificial intelligence, using generative AI tools to produce bland travel guides, inaccurate film blogs, and SEO-bait explainers. By and large, the goal has been: can we make more pages for ads without paying more writers?

Now, a group of tech outlets is attempting to incorporate generative AI into its websites, though readers won’t find a machine’s byline anytime soon. On August 1st, an AI chatbot tool was added to Macworld, PCWorld, Tech Advisor, and TechHive, promising that readers can “get [their] tech questions answered by AI, based only on stories and reviews by our experts.”

The AI chatbot, dubbed Smart Answers, appears across nearly all articles and on the homepages of the sites, which are owned by media / marketing company Foundry. Smart Answers is trained only on the corpus of English language articles from the four sites and excludes sponsored content and deals posts. The user experience is similar to other consumer tools like ChatGPT: readers type in a question, and Smart Answers spits out a response. Alternatively, readers can select a query from an FAQ list, which is AI-generated but based on what people are asking and clicking on. Smart Answers responses include links to the articles from which information was pulled.

Smart Answers gives readers sample questions.

Image: The Verge

The goal, leadership says, is to help readers looking for specific pieces of information among the mountain of writing.

“We’ve got this huge wealth of content. But sometimes you don’t want to read 2,000 words of great journalism — sometimes you just want a particular fact,” Neil Bennett, global director of product and data at Foundry, told The Verge.

The company had been working on developing Smart Answers for around five months before launch, Bennett says, and conducted rounds of testing with editorial staff for two to three months. A working group of editorial staff tested early versions of the tool and graded responses until they were satisfied with the answers.

“What we really needed to watch out for in our testing of this is that the answers they give are genuinely trustworthy, they’re the same answers an editor would give,” says Marie Black, editorial director at Foundry. 

The thinking is that Smart Answers could address a key problem users have seen with other AI tools: the tendency to make errors, cite nonexistent sources, or invent false information that sounds plausible. Smart Answers is trained only on content that must meet the bar of publication by reputable outlets — but it’s still a big promise to say the chatbot could deliver editor-level responses.

Smart Answers struggles with queries that seem straightforward. When I asked the chatbot, “When was the last iPod Nano released?” it incorrectly stated that it was released and discontinued in 2022. A few days later, the response had changed: “I don’t have enough information to answer this question.” 

The chatbot responds this way often; it doesn’t know who TikTok’s CEO is, for example. That could be a reflection of the type of coverage the Foundry outlets do (or don’t do). It feels like both a limitation and a good thing — on one hand, someone using the tool will likely still have questions they’ll have to go elsewhere to get answered. But in these cases, at least it isn’t making something up just to have a response.

In the case of the iPod Nano query, Black suggested that perhaps the chatbot was pulling from older stories about iPods that hadn’t been updated in years. (The device was discontinued in 2017.) For errors like this, Foundry relies on user feedback about the helpfulness of responses — if readers or staff find issues, they’re brought back to Miso.AI, the AI company building Foundry’s chatbot.

Smart Answers serves another purpose: affiliate revenue

Smart Answers serves another purpose besides synthesizing information from full articles: affiliate revenue. Below generated answers and above the relevant story links, the tool includes buying options for items it deems relevant to a reader’s search. A question about iPhone battery life, for example, provided shopping links to a handful of retailers selling phones. 

Many outlets like those under Foundry monetize content using affiliate marketing ads, generating revenue each time a reader makes a purchase using their link. (We do this at The Verge, too.) In this way, Smart Answers becomes a new way to carry these ads, generated to try to match the questions a user is asking.

“There are links in there to go and buy it, which we consider to be part of the reader service. If they’re asking about a product, they’re probably in the market to go and buy it,” Black says. 

Keeping users on the site until they buy something — whether through AI-generated blurbs or in-depth reviews — is more important than ever as generative AI takes over Google Search. Publishers, especially those creating product reviews or recommendation lists, stand to lose significant traffic (and ad revenue) as Google’s AI-driven Search Generative Experience places snippets from articles above information sources. 

“I think it’s unrealistic to ever assume that Smart Answers is going to directly replace that traffic that we’re all losing from Google Search,” Black says. The hope is that it could deepen engagement with readers who are already on the site.

In the face of criticism about the rollout of AI initiatives, other media companies like G/O Media have doubled down on AI content generation, frustrating staff and prompting workplaces to unionize. Black maintains that Macworld, PCWorld, Tech Advisor, and TechHive won’t be publishing AI-generated stories and that the chatbot could be useful for the reporting process. User queries or unsatisfying answers could signal to reporters and editors where there’s a need for more coverage. 

“This isn’t in any way a threat to our journalists,” Black says. “It literally cannot exist without our journalists, without them writing in-depth copy that answers every question a reader could ask.”

Read More

Mia Sato