Sign up for the daily CJR newsletter.

Recently, on the website of USA Today, I asked a generative AI chatbot called DeeperDive a question: “Is AI good?” DeeperDive, which was unveiled in June, is a product of a company called Taboola, one of the largest native advertising platforms on the internet, with some nine thousand publishing partners. USA Today’s corporate parent—formerly known as Gannett, which recently rebranded as USA Today Co.—runs America’s biggest newspaper chain, and has been in business with Taboola since 2013, amassing user data, selling ads, and suggesting “content you may like.” With the DeeperDive announcement, USA Today set out to “tap into AI-driven automation,” Trisha Gosser, the company’s chief financial officer, said on an earnings call over the summer: the chatbot’s “answer engine” would be housed directly on USA Today and the other sites under its umbrella, drawing from their newsrooms’ material. Michael Reed, the chairman and chief executive, said that DeeperDive represented “an exciting win” and that USA Today was “thrilled to be the first to roll out this technology” in the United States.

The release of DeeperDive was an acknowledgment of fundamental tensions between the rise of artificial intelligence and the delivery of news, since the futures of both USA Today and Taboola will be shaped by generative AI agents that scrape and reassemble information in response to users’ queries. As Adam Singolda, the CEO and founder of Taboola, put it, “It’s time to build a future where AI helps drive the open Web, not drain it, by generating meaningful traffic, sharing value, and respecting the creators who make the internet worth exploring.”

The DeeperDive chatbot, in responding to my query, was more ambivalent about the costs and benefits of its underlying technology: “AI, or artificial intelligence, has both positive and negative aspects depending on its application and implementation,” the answer went. “On the positive side, AI has brought significant efficiency and innovation to various sectors. For instance, Walmart is leveraging AI technologies to enhance its store operations, including real-time translation services for staff, which can improve workplace communication.” DeeperDive then referenced several USA Today stories that it—seemingly by accident—called “Article ID 5,” about job seekers using AI; “Article ID 3,” about AI’s impact on the environment; and “Article ID 14,” about the “biased recommendations” provided by AI to people using it to plan their travel. 

Topping these links was an ad for the “bravethinkinginstitute” (“She Manifested 7 Figures & Her Dream Home”), which featured a glossy-haired, radiant woman staring out into the distance, arms crossed, and lit by the setting sun. It was instantly recognizable: a “chumbox ad,” so called for their easy comparison to cheap bait, and familiar for their ubiquity online—including on trusted news sites—supplied by a number of companies, Taboola among the most prominent. When Singolda founded Taboola—nearly twenty years ago, after a long stint as an engineer in an Israel Defense Forces’ encryption and information security unit—the idea had been to form a video-focused internet startup; around 2010, the company shifted to become a video-discovery and -distribution service; today its line of work is known as “content recommendation.” Across these iterations, Taboola has been interested in connecting publishers, advertisers, and data brokers. It comes down to this: “Advertisers and marketers purchase the right to place ads on a network of widgets, hosted across the websites of a range of publishers,” according to a study of Taboola published in 2023 by the Israeli scholars Shira Dvir-Gvirsman, Anat Ben-David, and Yariv Ratner. “The content of the widget is personalized based on user-tracking data, collected through planting third- and first-party cookies on the publishers’ websites.” 

First-party cookies are set by the website you’re visiting and remember your preferences or login information. Third-party cookies are set by separate domains—analytics providers or advertisers—and follow you as you surf, collecting data in your wake. Over the years, the landscape of what is possible with third-party cookies has changed: Apple launched Intelligent Tracking Protection (ITP) in 2017, which limits them; Safari has blocked them by default; the General Data Protection Regulation (GDPR) in the European Union and the California Consumer Privacy Act do some of the same. “The advertising industry is mourning the death of the third-party cookie,” Taboola announced last year. But lucky for you, the company said, it’s “built for a cookieless world.” That is, with partner-publishers such as USA Today Co., Taboola can “bypass a lot of the restrictions other advertisers are facing,” as Dvir-Gvirsman and Ben-David told me when I asked them about this recently. By its most recent estimate, Taboola reaches roughly six hundred million unique active users daily.

Along with USA Today in the US, DeeperDive got initial buy-in from The Independent, in the United Kingdom, and then India Today Group. In October, Singolda gave an interview to an India Today TV program, and spoke about the endeavor as delivering an “AI sandwich”: the machine serves an instant answer; human-produced journalism layers in trust. “The desire for a human, verified, editorial-driven point of view is going to be like organic food,” Singolda said. “You know we’re going to need it. Like humans need healthy food.” Organic chum, you might call it. Better yet, it’s locally sourced. As Kurt Gessler, a product manager at Taboola who works on DeeperDive, told me: “A reader visits a specific news site because they trust that publisher’s brand and reporting. When they use DeeperDive, they aren’t getting an answer from the entire, unvetted internet; they are getting a reliable, AI-powered answer sourced exclusively from that publisher’s own, trusted content library.”

Be that as it may, trust has to be earned, and can be eroded. As is seemingly the case every year now, trust in media institutions at large is at a historical low: 28 percent, according to a 2025 Gallup poll. It is no coincidence that surveys reflect that the public has greater faith in local news organizations than in the mainstream press—and that it is in USA Today Co., which owns two hundred and twenty regional outlets, in which Taboola might find an ideal partner. The potential for mining useful data via local news is especially appealing, considering the self-organizing of any given paper’s readership. But at the same time that USA Today Co. is embracing DeeperDive, it appears to be divesting from community journalism. Recently, the company announced a hundred million dollars’ worth of cutbacks—the latest in a recurring series of buyouts and layoffs. Matt Crain, an associate professor of media and communication at Miami University in Ohio, and the author of Profit over Privacy: How Surveillance Advertising Conquered the Internet (2021), told me that operating-revenue-wise, USA Today Co. is “like half of what they were, or maybe even a third of what they were, twenty years ago. It’s a really precipitous decline.” That stress is felt across the news industry, as AI—in particular, Google Gemini, OpenAI’s ChatGPT, and Anthropic’s Claude—cuts into search traffic: “A nightmare in the middle of a nightmare,” Crain said. “It’s the Inception of nightmares. This is the death of outbound traffic from search engines.”

In a war that pits tech giants against news publishers and the ad-tech that loves them, both sides see AI as a valuable weapon in what is likely to be a fight to the death. DeeperDive, then, seems an attempt to get creative: chatbot ads don’t need to use cookies, since data collection happens directly through engagement. I shared with Crain some screenshots of my experiments with DeeperDive. The ads looked “horribly targeted,” he said. I slightly disagreed. As I kept playing around, I got promotions for a “Handmade painted book mug” that is “being snapped up in Dallas,” and “elderly leather sandals for those with poor balance.” Another raised this cryptic thought: “Rob Lowe: What Happens When Your Dog Eats Bananas?” I live in Texas, and I have a dog and foot problems. But none of these links brought me closer to businesses in, much less news about, my area. The deeper I dived, the more lost I became, inside the Inception Crain described. 

Another question, not one for the search box, emerges: Why would someone go to the USA Today website to type in a query, instead of simply, as is de rigueur these days, doing the same in Google, and waiting for that version of an AI summary? Or, if you prefer: asking ChatGPT? A recent study coauthored by Amy Ross Arguedas, of the Reuters Institute for the Study of Journalism at the University of Oxford, surveying nearly a hundred thousand news readers, found that 27 percent of them had interest in AI-generated summaries of news articles, but only 18 percent expressed interest in AI chatbots. The largest portion of respondents—34 percent—expressed that they “didn’t know” or were interested in “none” of the AI-propelled options for news available to them in the survey. There is “kind of a gap there,” Ross Arguedas told me, “between audience interest and industry interest.” 

Notably, that gap varies in size depending on where you are: in certain Asian and African countries, Ross Arguedas noted, there’s “more comfort” with AI in journalism. The gap is also smaller among younger people, for whom there is a higher level of open-mindedness about AI tools. (According to a Pew survey, the median readership age for USA Today is fifty.) Around 7 percent of people globally, Ross Arguedas told me, are using chatbots for their news: ChatGPT, etc. “A growing number of people,” she said, “sees value in being able to ask questions and have answers that are relevant to what they want to know.” She continued, “There are newsrooms that can’t afford to create these tools internally. So it’s understandable that they’ll turn to these partnerships.”

Despite the relatively low interest among readers in using AI tools, a whopping 56 percent of publishers surveyed by Reuters were looking into possibly developing or adopting them, if they had not already. DeeperDive is similar to features on offer from the Washington Post (from “Ask the Post AI,” a prompt: “What are the health benefits of forest bathing?”: “We’re on it. Checking our sources…”), the Financial Times (for subscribers to FT Professional, and billed as a means to help you “prepare for client and stakeholder meetings”), and The Nikkei, the world’s largest financial newspaper, based in Japan (also subscriber-only, and built in-house, as wider large language models do not yet process Japanese characters). Of these papers, only the Financial Times responded to my requests for comment: “Ask FT reflects FT journalism not only in the information it uses, but in the care and scrutiny our journalists have brought to its development,” a spokesperson told me.

In May of 2024, Aftonbladet, in Sweden, debuted Valkompisen (“election buddy”), an AI chatbot designed to answer readers’ questions about EU parliamentary elections by referencing the paper’s journalism. The bot answered some hundred and sixty thousand questions with no “hallucinations,” according to Martin Schori, the deputy managing editor at Aftonbladet, who helped develop the tool. (Or at least, he told me, they got no complaints.) “It’s a bit scary, because in the newsroom we usually have control of every single number we publish,” Schori told The Audiencers that summer. “With the bot, we didn’t know exactly what it was going to answer.” Later, Schori told me that a user had asked Valkompisen, “What if the EU was represented by the Fellowship of the Lord of the Rings, like, who would be Gollum?” The answer from the chatbot was the United Kingdom, because it had left the EU.

Aftonbladet is now involved in developing an AI chatbot that can be used across the board at Schibsted, the Norwegian media group to which it belongs. During the tool’s initial development, human journalists were tasked with double-checking Valkompisen’s potential answers; later, they also did random checks of the information provided to users. Humans likewise double-check DeeperDive, a Taboola spokesperson told me, through “an ongoing effort that uses a combination of automated and manual reviews.” When I asked about the glitchy responses I received to my queries, the spokesperson noted that the project “is still in beta.” The Financial Times spokesperson described an in-depth development process involving editorial teams testing the chatbot with “tricky questions.” It is, as ever, a partnership.

About a month after introducing DeeperDive, USA Today Co. made another AI-tool announcement, this time with Perplexity, a giant of the field. Comet, Perplexity’s “newly released agentic Web browser,” would now deliver “AI-curated answers” sourced from twenty-seven of USA Today Co.’s local papers, stretching from the Arizona Republic to the Providence Journal, the Palm Beach Post to the Detroit Free Press. “What does that mean?” Kristin Roberts, the company’s president, wrote in sharing the news. Results “will be based on the credible journalism created by your local newsroom.” There’s the nutritional value, she argued: “From school board and city council decisions to weather alerts and breaking news, our local stories matter—and they deserve to be part of the AI knowledge ecosystem.” And then there’s the compensatory arrangement: “It’s a model for how publishers are evolving and adapting and how AI companies can collaborate responsibly—with proper attribution and fair compensation—to contribute to the information ecosystem.” In other words, nutrients don’t just flow, in the form of verified news, to consumers hungry for accuracy. They also flow the other way, in the form of money, from users’ data and increased time spent on these AI tools, to the news publishers and their partners. And they flow in the form of training to the AI tools themselves, with each new question.

The announcement from Roberts was appended at the top with a widget from DeeperDive, inviting me to explore questions such as “How does the Perplexity partnership impact local journalism?” When I clicked on that, I got an AI-summary answer, which explained, “Recent analysis notes that increased integration with AI-driven services could expose local newsrooms to wider audiences, possibly boosting subscriptions or ad revenue, but it also means these publishers must navigate new revenue-sharing structures and the risk of dilution of local reporting’s unique voice.” Below that was a series of links—ads and articles, including an op-ed from a couple of weeks prior headlined “Marjorie Taylor Greene is actually starting to make sense.”

I decided to explore Perplexity further. “Read all Des Moines Register articles at no cost on Comet,” suggested a banner headline on the Register’s website, under the tiniest of text reading “advertisement.” I scrolled down to an endless GIF of someone smearing lotion onto their wrist: Doctor Says: Don’t Use Lotion, Use This Household Item Instead (For Dark Spots). A separate ad, on the Bergen Record, used the same headline. Both delivered me, via Taboola, to the virtual doorstep of Doctor Steven Gundry, the leading proponent of an anti-lectin fad diet. Gundry encourages patients to consume a liter of olive oil per week “in order to slim down,” according to an article in the Olive Oil Times, which notes that Gundry recommends his own brand, a liter of which would cost about a hundred and thirty bucks, and which was evaluated by one “internationally certified taste panel leader” as “lampante—or unfit for human consumption—calling it ‘one of the worst oils I have ever assessed,’ devoid of any positive attributes and exceedingly rancid.” (Gundry did not respond to requests for comment.)

“We want to give that power that billionaires have access to to everybody,” Aravind Srinivas, the CEO of Perplexity, said in a recent interview with Tiffany Janzen of Tiff in Tech, the wildly popular YouTube and social media channel. Srinivas characterized Comet as a “second brain” that takes on administrative tasks, allowing one’s “first brain” to focus on whatever it wants to, billionaire-style. “Our first brain can be truly at our natural best selves if we can just be curious and explore and interact and meet people and like, you know, strategize,” he said to Janzen, “while the second brain takes care of, like, all the boring mundane workflows.” When I reached out to Perplexity, Jesse Dwyer, the company’s spokesperson, told me that he didn’t endorse the “browse like a billionaire” concept, but he’s seen the line catch on. “At its heart,” he told me, “Perplexity opens up the knowledge of the world to more people.” He added, “The search for wisdom is much more important than the search for wealth.”

I decided to ask Comet how it was changing local journalism. From its answer, I learned that Comet Plus, a five-dollar-per-month subscription service, shares 80 percent of its subscription revenue with the publishers from which it derives content. “This revenue-sharing model aims to sustain local journalism by creating a direct income stream tied to AI engagement with their reporting,” the agent told me, citing a Forbes article that characterized Comet as “a model built for an internet where the lines between reading, searching, and task-completion get muffled.” Dwyer told me—speaking personally, he said—that if Comet Plus succeeds, it will, actually, help writers become millionaires (if not billionaires). He also wanted me to know that “Comet is a Web browser,” whereas “DeeperDive is just a search product designed to be embedded in publisher websites. Consider it a widget (and not a very good one).”

Before I got drawn in too deep, John Naughton, a senior research fellow at Cambridge University and the author of From Gutenberg to Zuckerberg: What You Really Need to Know About the Internet, wanted to pull me out of the well. Tech companies, he reminded me, are not charitable organizations. “All that stuff about hard-pressed newsrooms and so on,” he said, “is basically bullshit.” He wondered aloud about DeeperDive and other such tools: What were the problems they were solving, and for whom? “Maybe the problem it’s solving,” he suggested, “is ‘How can we get more granular information about users?’” He told me about the ELIZA Effect—the first chatbot ever was named for Eliza Doolittle, from Pygmalion. ELIZA had a gently sycophantic affect, encouraging its users to share more and more personal information. “I think that’s the context for all of these things,” Naughton told me. “So it’s impossible not to suspect that at the root of this is providing a conversational partner in a machine that is going to yield very fine-grained information about its users.” In essence: “The longer you can keep them on it, the more data you can extract.”

There are, of course, plenty of ways AI can give some power back to cash-strapped publishers whose journalists are working overtime, on shoestring budgets, to deliver the all-important news. Certainly, newsrooms are finding uses, for instance, that help with transcribing interviews, scraping data, and poring over the minutes of city council meetings that no reporter has time to attend. The American Journalism Project, a venture philanthropy focused on supporting local news, hosts a Product and AI Studio, which helps local news outlets experiment in the name of using AI in the public interest.

Much of the discourse around this technology portrays news publishers as being at odds with major tech companies, which for the past few decades have drained journalism’s ad market dry. “We are the Robin Hood of the open and free internet,” Singolda has said, “against the giants of Google, Facebook, and Amazon.” He speaks often of how Taboola effectively stands for the “open Web.” “Generative AI companies need to make a stand—will they pay for the sources of data they are ingesting daily?” he asked, in an op-ed for CNBC. In December of 2023, when the New York Times sued OpenAI for allegedly using millions of Times articles to train its large language models, he wrote, on LinkedIn, “The NYTimes is upset. You can relate. No money paid. No attribution. Not nice.” He compared the situation to Facebook’s “Instant Articles,” circa 2015, and wrote, “How does OpenAI want to be remembered? they could evolve to be a force of good, a ‘Jedi’ to support journalism, support high quality content, and the open web. Or they can decide not to do it, and evolve to be a Sith lord, not pay and not attribute.” Reflecting in another LinkedIn post, Singolda wrote, “The internet was not meant to be summarized, it was meant to be explored.”

To Dvir-Gvirsman and Ben-David, the language about Jedis, Sith lords, and Robin Hood “casts Taboola as a liberator” and emphasizes “redistributing value from dominant platforms back to news organizations.” In practice, though, they think the relationship between ad-tech companies and news outlets is much less clear. Even if a company such as Taboola provides greater revenue to publishers than, say, Meta—and this seems to be variable; Taboola declined to say if it was true—it comes with trade-offs. “Parts of the editorial process and website real estate are ceded to algorithmic systems, reducing publishers’ control over their own content environments,” Dvir-Gvirsman and Ben-David noted. “Thus, the Robin Hood analogy obscures the complexities of power and dependency in these arrangements.”

Over the course of my testing of DeeperDive, I looked for any kind of editorial decision being made by the chatbot. I found little overt language in this direction, and much more of a maddening tendency to note caveats—or, when little relevant information from USA Today articles was available, to deliver a good deal of secondarily useful information. On the day of Diane Keaton’s death, for instance, I asked DeeperDive for the cause. “As of the last reports, the specific cause of her death has not been publicly disclosed,” I was told. The bot continued, a few lines later, “Reese Witherspoon paid a touching tribute to Keaton, highlighting her influence and mentorship early in Witherspoon’s career.” Discouraged, I searched for something else: “Does Tylenol cause autism?” “Tylenol’s link to autism is a controversial topic with varying perspectives,” I was told, and then I combed through the unfolding paragraph to learn that the Trump administration’s recent claims about Tylenol causing autism have been “met with significant skepticism from the scientific community.”

When I reached out to Gessler, he contrasted Taboola with the “walled gardens” of tech giants: ecosystems in which people roam and find themselves unable to escape—kind of forgetting after a while that they might even want to leave. Major platforms “prioritize fast answers but discourage deeper exploration, sending minimal traffic to original sources,” he told me. “This creates digital ‘walled gardens’ that oversimplify complex topics, erode critical thinking, and undermine quality journalism. Worse, GenAI systems often rely on content scraped from trusted publishers without proper compensation, threatening the sustainability of newsrooms.” Indeed, Gessler’s characterization fit my own experience using Comet, that walled garden that also partnered with USA Today Co., under the headline that Comet, like DeeperDive, was helping to save journalism. (Dwyer, at Perplexity, told me that for its part, “Perplexity has been dismayed with USA Today for creating confusion about DeeperDive as if it is a Perplexity product.”) 

In the now ubiquitous theory of the “enshittification” of the internet, so termed by the writer Cory Doctorow, a large number of garden captives wallow in the sludge that contemporary online life comprises: awfully like the stuff of Tylenol sickness and olive oil cures. Taboola strives to differentiate itself. “We compete for advertising dollars with the closed ecosystems of technology companies,” read the company’s most recent 10-K, a financial document filed with the Securities and Exchange Commission. (Taboola has used that language in SEC filings for at least five years.) “We also compete with those companies, in particular Google, for real estate on digital properties.” Because of its relationship with news publishers, however, Taboola insists on an ideological distinction: “Unlike walled gardens,” the company argues, “we only interact with consumers through our partners’ digital properties, hence we do not compete with our partners for user attention. Our motivations are aligned. When our partners win, we win, and we grow together.” News and ad-tech working hand in hand, against the behemoths of our age. “It’s not altruism,” as Crain told me. “There’s a mutual need for these guys to glom on to each other.”

Crucial to the strategy here is an understanding of “unbundling” and “rebundling,” which is basically the creation of surface area. Like a crystal growing in new directions, articles and advertising can take on different shapes, for different readers. A news article—stripped of its original context on, say, the front page, or next to an advertisement for a certain hat—can be repurposed for a specific person with a specific query and a specific set of ailments, interests, or geographic data. Dvir-Gvirsman and Ben-Anat call this a “matryoshka doll” in describing how a publisher such as USA Today might present its own material that envelops and hosts a widget from Taboola, “which in turn hosts the news organization’s content.” Algorithmically driven personalization, such as the ads I got about mugs in Dallas, add another expanding layer of information for me—and information, presumably, on me. 

“Let’s also remember,” the researchers told me, “that some users are unable to understand what content is and what the ads are in this interface (this is obviously by design).” They view DeeperDive as a prime example of exactly this, one that points toward another aspect of the economics: speculation. “The promise that data can be transformed into revenue is often used to attract investment and inflate stock prices, even when the actual monetization mechanisms are uncertain or underdeveloped. This speculative logic is a defining feature of the industry, where innovation is frequently framed as revolutionary, but serves primarily to sustain investor interest and market valuation.” 

When I spoke about this with Naughton, he recalled when turbo engines were invented “and everything was turbocharged.” As for AI, he said, “there’s a huge hype storm about how transformative it is.” Which means, he added, “newsrooms everywhere and corporations everywhere are determined that somehow they have to show that they’re on board with this stuff.” I thought back to something Ross Arguedas, the researcher at Oxford, had said offhandedly, about how I would probably use AI to transcribe our interview. Increasingly, I had the feeling not of inevitability, but of something big happening behind the scenes: a decision, made somewhere or another, to go all-in, even though—or perhaps because—the end of whatever came before was now in sight. In 2005, Gannett’s total operating revenue was 7.6 billion dollars, a record for the company. In 2024, the company’s total operating revenue was about 2.5 billion dollars.

One day, I typed “what is deeperdive” into the DeeperDive search box. “Currently, there is no major news coverage of a concept, company, or product specifically named ‘deeperdive’ in the most recent articles,” the chatbot told me, even though a USA Today article focused explicitly on its new tool. “Please clarify, or explore related subjects like the ‘missing’ matter found in the universe using fast radio bursts missing matter in universe found using fast radio bursts, or check trending how-to guides for everyday tasks like wrapping a burger to avoid a messenjoy your burger without making a mess with this simple trick.” (A Taboola spokesperson told me that I may have gotten this response because “DeeperDive content currently has a thirty-day period before it’s removed from the database in order to keep DeeperDive’s content as refreshed as possible,” and I may have entered my query past the sell-by date. “But we’re working on optimizing the look-back settings to reduce these cases.”)

I have to admit: I clicked on the follow-up question suggested to me, about how to wrap my burger. “By layering ingredients efficiently and using condiments as a sort of ‘glue,’ you can significantly reduce the chance of a burger blowout during eating,” I learned. I recalled Singolda’s presentation launching DeeperDive in India. This was, perhaps, the AI sandwich.

This piece is part of Journalism 2050, a project from the Columbia Journalism Review and the Tow Center for Digital Journalism, with support from the Patrick J. McGovern Foundation.

Has America ever needed a media defender more than now? Help us by joining CJR today.