The Media <> AI Death Spiral

The short version

Search used to give you ten links and let you decide. Now Google writes one answer on top, and only 8% of users click through to a website, against 15% when there is no AI summary.
In one year, Google sent news sites 40% less traffic. Less traffic means fewer ads, and ads make the budget for journalists. The Washington Post cut more than a third of its journalists in one day, and its editor blamed search.
The AI tools that took over get the source wrong more often than right. Columbia tested eight of them: the best was wrong 37% of the time, the worst 94%.
Sponsored content slips through the same gap. Nine in ten readers miss the label, and AI tools can't tell either.
Whatever gets published wrong comes back as the next AI answer.

Nobody reads media anymore. We get our news from the feed.

Half of that is true. Social and video are now the biggest source of news across 48 markets, ahead of news sites and apps. More than a quarter get their news from creators. In the US, most news influencers have never worked in a newsroom.

Then follow the links back. More than a third of creators name mainstream news as one of their main sources. AI assistants cite journalism more than any other kind of source. The influencer is reading the newspaper so you don't have to. The chatbot is reading the newspaper, Wikipedia and Reddit.

The spiral: AI answers, traffic drop, less ad budget, layoffs, less checking, misinformation, and back into the next AI answer

We used to google it. Now we read the summary.

Search used to hand you ten links and a choice. Now it hands you one answer. Google says searches are at an all-time high, and its AI Overviews now reach more than 2.5 billion people a month. OpenAI says its models reach more than a billion active users. Both give you the answer on the spot.

What that does to publisher traffic:

8% of users click through to a website when Google shows an AI summary, against 15% without one.
Chatbots send news sites under 1% of their traffic.
Google traffic to publishers fell 40% in a year (Chartbeat).

1%of Google searches with an AI summary end with a click on the summary's own sources.Pew Research Center, 2025

The newsroom still writes the story. The audience reads Google's summary of it.

Click-through to websites, with and without an AI summary

Share of Google searches that ended in a click on a result. US, March 2025. Pew Research Center

An AI summary halves click-through.

Media layoff rounds

The Washington Post put it in writing. "Our organic search has fallen by nearly half in the last three years," its executive editor wrote, and on 4 February the Post cut more than 300 of its roughly 800 journalists in one day.

It was one bill among many. US news companies announced 1,901 job cuts in the first nine months of this year, more than in the same months of 2025. Read the announcements side by side and you get two genres of press release. USA Today Co: "Search traffic is under pressure." Scripps: "We're leaning into AI, automation, technology and the centralisation of some roles." The first is a confession. The second is a pivot deck.

The cuts started with the people who slow things down. US dailies had 10,676 copy editors in 2002 and about half that a decade later. Last year, three fact-checking projects closed for every one that opened.

What filled the gap was speed. In Baltimore, Pew once found that eight in ten stories simply repeated what had already been published. In France, a quarter of stories were copied elsewhere within four minutes. Add a quote, change the headline, publish first. Four minutes is less time than it takes to make a coffee.

The people who are left write with AI too. In a Reuters Institute survey of 1,004 UK journalists, 27% use it for part of the job every day, mostly transcription and translation. One in ten use it to draft articles, and only a third say their newsroom has trained them. When it goes wrong, it goes wrong in print: CNET had to correct 41 of the 77 articles its AI wrote.

US news-industry job cuts

Announced cuts. *2026 = January to September. Challenger, Gray & Christmas

2024 was the peak. January to September 2026: 1,901 cuts, up 9% on the same months of 2025.

Can AI or an average reader spot a paid article?

Paid content on news sites is old and legal. The question is whether anyone can tell.

Readers can't. In a Boston University experiment, fewer than one in ten adults recognised a labelled sponsored article as an ad. The label was there.

AI tools can't either. Muck Rack, which measures what AI cites, sorts links by the type of site they sit on. A sponsored article on a news site sits on a news site, so it would count as journalism.

And some of it carries no label at all. In August, Press Gazette reported a PR agency quoting $25,000 to place an article in a major US business title through its "contributors." Asked whether its UK placements were marked as promotional, the agency said: "Some are, some are not."

Beyond news sites:

Influencers: when the EU checked 576 of them, 97% posted paid content and one in five consistently said so.
States: in 2024 the US Justice Department charged two RT employees with routing nearly $10 million to an American video company whose audience was never told.

Readers who recognised sponsored content as an ad

UGA; Boston University

Even when the label was there, they missed it.

How misinformation gets into AI answers

Now the audience that stopped clicking through asks the AI. In a Columbia University test, eight AI search tools were shown excerpts from real articles and asked to name the source. They got it wrong more than 60% of the time, from 37% for the best tool to 94% for the worst. They often credited a copy over the original.

They also stopped saying "I don't know." In August 2024, leading chatbots repeated a false news claim 18% of the time. A year later, 35%, and they now answer every question. Confidence went up. Accuracy went the other way. Most of us have sat in that meeting.

Meanwhile the number of AI-written "news" sites that NewsGuard tracks went from 49 to 3,749 in three years. That is the material the next answer is built from. Misinformation goes in, an answer comes out, and nobody is left in the newsroom to catch it. Even the best newsrooms slip. In May, the New York Times ran an editors' note admitting a politician's "quote" was an AI summary of his views.

Trust in news sits at 37%, the lowest the Reuters Institute has measured. Trust in news found through AI chatbots is 20%. People use them anyway.

AI search tools naming the wrong source

Tow Center, Columbia Journalism Review, March 2025

The tools cited copies over originals. The spiral multiplies the copies.

What the spiral doesn't kill

Here is the best argument against all of this. Total pageviews across Chartbeat's network barely moved. In the US, trust in the media rose this year, from 28% to 33%. Readers moved to social, to apps, to newsletters like this one, to outlets they open on purpose.

What the spiral eats is the middle: the generic, the search-dependent, the copied and the quietly paid.

Where does the truth drop out?

We don't read the media. We listen to influencers, then ask ChatGPT to confirm. But the influencers and ChatGPT both read the media.

Every AI answer is built from what gets published, and the model can't tell what it's reading. It can't tell the reported story from the copy posted four minutes later, or the edited piece from the sponsored one. Some of the bylines aren't even people. In May, Press Gazette found four prolific crypto writers it could not verify as real. Between them they had more than 1,000 articles in over 30 outlets, Forbes and HuffPost among them, many promoting clients of one crypto PR firm. The firm denies running the profiles. A chatbot reading those pages has no way to know.

How to check an AI answer

Ask with search turned on. Without it, the chatbot answers from memory and the sources can be invented. No link, no fact.
Open the link and find the sentence yourself. Make sure it's the original article, not a copy or a sponsored post. In Columbia's test, Grok 3 sent 154 of 200 links to error pages.
Ask neutrally. "Is X true?", not "Why is X true?" Chatbots tend to agree with however the question is framed.
Ask for the case against. "What's the strongest evidence that this is wrong?"
Ask twice. In a new chat, or another chatbot. If the answers differ, it's guessing.

The research behind this piece: every number, with charts →

Subscribe so you don't miss the next episode of Your opinions are not your own.

The short version

We used to google it. Now we read the summary.

Media layoff rounds

Can AI or an average reader spot a paid article?

How misinformation gets into AI answers

What the spiral doesn't kill

Where does the truth drop out?

How to check an AI answer

Sources

Read next

What If the Pope Ran AI Safety?

Who killed the story?

A 1984 Burger Ad Did What Most Growth Campaigns Still Can’t

No Lesson Learnt: Coinbase and the Crypto Hire-Fire Cycle