Google rolled out its AI Overview feature last year.Jeff Chiu/The Associated Press
News publishers say the AI-generated summaries that now top many Google search results are cutting into their online traffic – and experts are still flagging concerns about the summaries’ accuracy as they warn the internet itself is being reshaped.
When Google rolled out its AI Overview feature last year, its mistakes – including one suggestion to use glue to make pizza toppings stick better – made headlines. One expert warns concerns about the accuracy of the feature’s output won’t necessarily go away as the technology improves.
“It’s one of those very sweeping technological changes that has changed the way we . . . search, and therefore live our lives, without really much of a big public discussion,” said Jessica Johnson, a senior fellow at McGill University’s Centre for Media, Technology and Democracy.
“As a journalist and as a researcher, I have concerns about the accuracy.”
Google’s unleashes ‘A.I. Mode’ in the next phase of its journey to change search engine
While users have flagged mistakes in the AI-powered summaries, there is no academic research out there yet defining the extent of the problem. A report released by the BBC earlier this year examining AI chatbots from Google, Microsoft, OpenAI and Perplexity found “significant inaccuracies” in their summaries of news stories, although it did not look at Google AI Overviews specifically.
In small font at the bottom of its AI summaries, Google warns users that “AI responses may include mistakes.”
The company maintains the accuracy of the AI summaries is on par with other search features, like those that provide featured snippets.
“As people use AI Overviews, we see they’re happier with their results, and they come to Google to even ask more of their questions,” a Google spokesperson said in a statement. “The vast majority of AI Overviews are highly factual and we’ve continued to make improvements to both the helpfulness and quality of responses.”
Chirag Shah, a professor at the University of Washington’s information school specializing in AI and online search, said the error rate is due to how AI systems work.
Generative AI doesn’t understand concepts the way people do. Instead, it works by making predictions based on massive amounts of training data. Shah said that “no checking” takes place after the systems retrieve the information from documents and before they generate the results.
“What if those documents are flawed?” he said. “What if some of them have wrong information, outdated information, satire, sarcasm?”
A human being would know that someone who suggests adding glue to a pizza is telling a joke, he said, but an artificial intelligence system would not.
It’s a “fundamental problem” that can’t be solved by “more computation and more data and more time,” Shah said – and better technology could actually make the problem worse.
“If anything, I worry that they will get so good . . . people will get comfortable enough with them that we will just trust them beyond what their abilities are,” he said.
He said online searches in general are changing in a fundamental way. As Google integrates AI into its popular search function, other AI companies’ generative AI systems, such as OpenAI’s ChatGPT, are acting as search engines themselves.
Canadian CEOs are embracing generative AI’s speed and efficiency. The impact on their employees is less certain
Search engines were designed originally to help users find their way around the internet, Shah said, but now the goal of those who design online platforms and services is to get the user to stay in the same system.
“If that gets consolidated . . . that’s essentially the end of the free web,” he said. “I think this is a fundamental and a very significant shift in the way not just the search but the web, the internet operates. And that should concern us all.”
A study by the Pew Research Center from earlier this year found users were less likely to click on a link when their search resulted in an AI summary. While users clicked on a link 15 per cent of the time in response to a traditional search result, they only clicked on a link eight per cent of the time if an AI summary was included.
That’s cause for alarm for news publishers, both in Canada and abroad.
Paul Deegan is CEO of News Media Canada, which represents Canadian news publishers. He said the AI summaries are acting as a drag on news media outlets’ online engagement.
“Zero clicks is zero revenue for the publisher,” Deegan said.
Alfred Hermida, a professor at the University of British Columbia’s journalism school, said Google used to be a major source of traffic for news outlets.
“People would search for something on the web, find a link to a news story and think, oh, that sounds interesting, I’ll click and read it. Of course, if you have an AI summary, it’s done that for you,” he explained.
“When you have most people who are casual news consumers . . . that AI summary may be enough.”
Last month, a group of independent publishers submitted a complaint to the U.K.’s Competition and Markets Authority saying that AI overviews are causing them significant harm.
Keldon Bester, executive director of the Canadian Anti-Monopoly Project, said there is a competition issue at play and there could “potentially” be a case under Canadian law.
He noted that Google has been hit with competition cases in the past, including one which saw the company lose an antitrust suit brought forward by the U.S. Department of Justice over its dominance in search.
“We have a single company which is and has been the front door to the internet,” Bester said. “As we appear to move to this kind of narrowing approach, whether that’s AI summaries or chatbot interactions, it really is in my mind just another iteration of those same concerns.”
In a post last week, Google’s head of search Liz Reid said “organic click volume” from searches to websites has been “relatively stable year-over-year.”
Reid said that pattern contradicts “third-party reports that inaccurately suggest dramatic declines in aggregate traffic – often based on flawed methodologies, isolated examples, or traffic changes that occurred prior to the roll out of AI features in Search.”
She said Google cares “passionately – perhaps more than any other company – about the health of the web ecosystem.”
Clifton van der Linden, an associate professor and director of the Digital Society Lab at McMaster University, noted many AI summaries are accurate.
“I think the reason you’re seeing these rapidly changing dynamics in referral traffic is in part because users do find these AI-generated summaries useful . . . but useful does not necessarily equate to credible, authoritative, or correct,” he said.
He said that if users bypass a link to a news site due to an AI-generated summary, that “compounds an existing problem” in Canadian media, which is dealing with a ban on news links on Facebook and Instagram.
Justin Trudeau’s Liberal government passed the Online News Act in 2023 to require Meta and Google to compensate news publishers for the use of their content. In response, Meta blocked news content from its platforms in Canada, while Google has started making payments under the legislation.
The future of that legislation seems uncertain. Prime Minister Mark Carney indicated last week he is open to repealing it.
Johnson said Canadian media has now experienced a “one-two punch” – first from Meta pulling news links and now from the emergence of AI search engines.
“The point is, and other publishers have raised this, what’s the point of me producing this work if no one’s going to pay for it, and they might not even see it?”