Arvind Narayanan thinks AI will come to feel more trustworthy—and it will remove the tells that let readers know they’re encountering AI slop in the wild. The Princeton computer scientist even predicts that these more accurate chatbots to come could become what many people treat as a “truth oracle.” Pew’s survey this year found that 10% of Americans are already doing this.
Narayanan addressed the Computation + Journalism Symposium at Northwestern’s McCormick Foundation Center, where Fortune attended. The agenda included the challenges facing media and proliferation of AI slop. His keynote was titled “AI and Journalism: Skating Where the Puck is Going.” He urged the audience to prepare for AI “not as it is today, but as it will be five or 10 years from now.” He said his computer science background is helpful in delineating which of today’s limits are fixable engineering problems and which are inherent. He has never worked in a newsroom, he noted. “It’s possible I don’t know what I’m talking about,” he cautioned, before laying out a compelling argument.
He broke his argument intwo two parts: If chatbots become the default place to settle factual questions—Narayanan said “virtually all” questions about, say, who is running for political office are headed there—the newsroom loses one more piece of its bundle, and arguably the most basic one. What’s left is the work a machine can’t settle: contested questions, narrative, storytelling and being explicit about what he called “positionality,” a challenge for an industry that strives for objectivity.
Narayanan directs Princeton’s Center for Information Technology Policy but he has emerged of late as a thought leader on AI, the future of work and the media industry; he co-wrote AI Snake Oil with Sayash Kapoor and his newsletter AI as Normal Technology has over 87,000 subscribers. Narayanan talked to Fortune in August about the AI backlash, which he sees as a complicated thing, and expanded on his insight into why productivity and return on investment has been so hard to measure or even nonexistent. He calls this a “skinny hamburger, fat bun” problem, where the meat of execution shrinks down to almost nothing with the use of AI, while the decision and deliberation “buns” swell, with a great deal of anxiety and uncertainty.
The oracle and the slop
Narayanan said the common picture of chatbots as simple next-word predictors is outdated. Today’s chatbots are “neurosymbolic systems,” he said: a language model plus tools that search the web, retrieve documents and write code to analyze data. Errors that were common a year ago, such as miscounting the letters in “strawberry” or Google’s AI Overviews suggesting glue on pizza, now rarely occur. He called the remaining problems “gradually getting solved.”
He then asked what happens when expectations catch up. His elementary-school-age children, he said, may never experience hallucination the way adults do, much as Wikipedia once went from an untrusted reference to a “much more authoritative” source. “I’m not saying it’s inevitable. I’m not saying it’s a good thing,” he said. On X, he noted, users already settle arguments by asking Grok, mostly conservatives with low trust in legacy media. He estimated the wider shift at about a decade out.
A member of the audience ventured that chatbots can suffer from what press critic Jay Rosen famously called the “view from nowhere” for the false neutrality expressed by journalists reluctant to use their own judgment and properly inform the audience. Daniel Trielli of the University of Maryland asked about this danger, and Narayanan agreed, and said he sees that as a challenge for journalism to continue defining going forward.
One of Narayanan’s most contrarian claims was that agenda-setting by the media has backfired. The Washington Post‘s role in Watergate was “an unalloyed good,” he said, but cited research from “many communication scholars and economists” showing that efforts to do the same in the Trump era, with audiences already splintered, have fed polarization and hastened the collapse of trust.
AI writing, he said, is similarly “largely a solvable technical problem.” A model could be trained to mimic one writer almost perfectly, mainly a question of cost, and detectors “might not work that well.”
The unbundling
Narayanan said his talk was “as much about the continuing effects of social media on newsrooms as it is about AI.” The newsroom is now thought of as a single institution bundling reporting, analysis, fact-checking and distribution, Narayanan noted, but he argued that this was a “historically contingent” development and not a given going forward. The ad-supported penny papers of the 1830s broadened what counted as news because earlier papers lived on subscriptions and political patronage. Objectivity arose in response to World War I propaganda, and paid reporters and interviews also “had to be invented.” Calling someone a “hired reporter,” he noted, was once an insult, “like a hired gun.”
Narayanan’s framework draws on earlier debates. He quoted Clay Shirky’s line that “society doesn’t need newspapers. What we need is journalism,” and said Paul Starr had warned, from the same era, that the loss of newspapers’ economics would leave public goods like accountability unfunded. Starr has been proven right “in many ways,” Narayanan said, “but it’s maybe not been as apocalyptic” as feared, because philanthropy has stepped in.
His claim is that, piece by piece, specialists now do each job more cheaply than a newsroom can, and AI is hastening the process along. “For each of these components, there is kind of a competitor that produces this component in a more standalone way that is better structurally equipped just based on the economics of production,” he said. Trade reporting, for instance, went to outlets like Politico, Punchbowl and Stat News. Civic and local journalism has largely moved to a “philanthropy-funded model,” he said, and investigations increasingly run on nonprofit money—”a reversion to historical patterns.” Analysis, he stressed, has moved to independent journalists, social media creators and “people like me.” Returning to what he called “the overarching point” of his talk, he said that the idea of packaging everything together in a single newsroom “is no longer the most economically viable way of producing news.”
One of Narayanan’s most contrarian claims was that agenda-setting by the media has backfired. The Washington Post‘s role in Watergate was “an unalloyed good,” he said, but there is evidence that efforts to do the same in the Trump era, with audiences already splintered, have fed polarization and hastened the collapse of trust.
Analysis, as Narayanan defines it, is a “layer that is between fact and opinion.” Citing economics blogger Noah Smith, he argued that journalists have too long and too often treated this dividing line as binary, so the analysis layer “never really had a home in newsrooms.” He offered the example of whether the AI industry is a bubble, as something an analysis could tackle. An analyst’s authority, he argued, now comes from a body of work rather than a masthead. (Fortune and Fortune Intelligence often run pieces as analysis, for what it’s worth.)
The internet “democratized distribution” and AI is now “democratizing production,” he said: “It is now possible to produce first-rate journalistic content of various kinds with a one- or two-person team.” It is “not so much a decline of news,” he said, as “the migration toward a bunch of other institutions.”
He also co-wrote textbooks on Bitcoin and on fairness in machine learning, and led the Web Transparency and Accountability Project, which examined how companies collect personal data. He said a computer scientist’s edge is telling which of today’s limits are fixable engineering problems and which are inherent. He has applied that to law, software engineering and higher education with co-author Sayash Kapoor, and now to journalism. He said he has never worked in a newsroom.
Who pays
The “real crisis,” Narayanan said, is the power gap between AI companies and publishers. Publishers need chatbot distribution more than AI firms need any one outlet, he said, so licensing deals underprice journalism.
He pointed to the “pivot to video,” when a Facebook algorithm change forced news organizations to remake their business models, as “an awful illustration of the extreme power asymmetry.” His fix is democratic input into how AI companies tune their algorithms—something like Meta’s Oversight Board, but with journalists having a “seat at the table.”
Narayanan’s answer was that journalists need to form a broad political movement, banding together to advocate for the future of the profession, possibly including a tax on AI and social media companies. It would have to be “broad-based,” not “perceived as partisan,” and include creators along with legacy outlets, he said.
But the alternative, mostly unmentioned in the room, is that journalism was simply a phenomenon of the last two centuries and technology and economics are evolving away from it as an organized and aggregated industry.
Disclaimer : This story is auto aggregated by a computer programme and has not been created or edited by DOWNTHENEWS. Publisher: fortune.com
