THEBLACKBOOK AUDIT
Investigation · Who Controls What You Know · AI vs. the Press

The doom loop.

Again, not our phrase. It’s a Microsoft executive’s, describing the company’s own AI: a machine that answers your question by summarizing the news, so you never visit the newsroom — starving the reporting the machine was trained on. If that’s allowed to run, what happens to journalism?

The numbers behind the collapse — and the new drop in traffic AI is adding to it — are graded FACT. That answer engines will break the business model that funds original reporting is a forecast, well supported but not certain, graded PROBABLY TRUE. And we are honest about what AI did not cause: the decline started years before ChatGPT.

Disclosure

This page was drafted with Claude, made by Anthropic — an AI company whose own products summarize and answer from news, and which is part of the industry described here. We grade our own side at least as hard as anyone else; the fuller conflict-of-interest disclosure is in the companion piece, An Astonishing Theft.

§1 · Summary Brief

What this page argues

Start with what is already true. American journalism has been shrinking for two decades: since 2005 more than 3,200 print newspapers have closed — about two a week — leaving roughly 5,600, four in five of them weeklies. Some 206 counties now have no local news source at all and 1,561 have only one, so around 55 million Americans have little or no access to local news. Fewer than 100,000 people still work in newspaper publishing, and daily circulation has fallen from about 55 million in 2005 to under 21 million. That collapse was not caused by AI. It was caused by the internet: classified advertising fled to Craigslist, and the advertising that funded reporting moved to Google and Facebook.

What AI answer engines do is attack the one lifeline the survivors had left — the search and referral traffic that still brings readers, subscriptions, and ad views to news sites. When a chatbot or an “AI Overview” answers your question at the top of the page, you have less reason to click through to the article. Pew found that when Google shows an AI summary, users click a traditional result link just 8% of the time, versus 15% without one — and they click a link inside the summary only about 1% of the time. In the newspapers’ case against OpenAI and Microsoft, Microsoft’s own data showed 83–93% drops in click-throughs to Times and Daily News sites from its Copilot answer engine, and Cloudflare’s CEO put OpenAI’s ratio of pages scraped to visitors sent back at 1,500 to 1.

Two consequences follow. First, the money that pays reporters keeps draining, because readers who get the answer from the machine don’t subscribe to or see ads on the source. Second, the same tools flood the zone with cheap synthetic “pink slime”: at current prices it costs roughly $6,800 to generate a million news-style articles with no reporter, and operations like Prism News have run 200 AI-generated outlets posing as local newsrooms with four employees. Real reporting gets harder to fund at the same moment fake reporting gets nearly free to produce.

That is the trap a Microsoft executive named in his own company’s files: a “doom loop” in which the answer engines strip the revenue from the newsrooms whose work they need, until there is less original journalism left to train on or summarize. There is a fork in the road — the biggest publishers are signing licensing deals (News Corp’s with OpenAI was reported at about $250 million over five years) while others sue — but licensing money flows to the large and national, not the local weekly or the county with no paper at all. Left unmanaged, the likely outcome is not the end of information but the end of the people who gather it: a public that asks a machine “what’s happening?” and gets a fluent answer with fewer and fewer humans behind it.

What we are NOT claiming

We are not claiming AI caused the newspaper collapse. It didn’t — the losses since 2005 trace to the flight of classified and display advertising to the internet giants, long before generative AI existed. AI is an accelerant on what remains, not the original fire. That distinction is why the historical numbers are FACT and the forward-looking harm is PROBABLY TRUE.

We are not claiming the outcome is fixed. Licensing deals could, in principle, route real money back to journalism; bargaining laws (as tried in Australia and Canada) could force it; courts could rein in the copying. And we are not exempting our own maker: Anthropic is part of this industry, and its tools summarize news too.

Recommended reading

Books that go deeper on this story. Links are Amazon affiliate searches — buying through them supports the work at no cost to you.

▶ Dossier

The same investigation, restaged one beat at a time. Step through it here, or present it fullscreen.

Who Controls What You Know

The doom loop.

A Microsoft executive's phrase for the company's own AI: a machine that answers your question by summarizing the news, so you never visit the newsroom — starving the reporting the machine was trained on. If it runs unchecked, what happens to journalism?

1 / 8▶ Present fullscreen
§2 · Graded Claims

The collapse, the accelerant, the flood, the fork, and the loop.

Journalism was already collapsing before AI — and AI didn't cause that.

FACT

Per Northwestern's Medill State of Local News (2024): more than 3,200 print newspapers have closed since 2005, about two a week (130 in the past year), leaving roughly 5,600 — 80% of them weeklies. There are 206 counties with no news source and 1,561 with only one, leaving about 55 million Americans with limited or no local news. Fewer than 100,000 people work in newspaper publishing (BLS), and per Pew, daily circulation fell from roughly 55 million in 2005 to under 21 million by 2022. The cause was the internet's capture of advertising — classifieds to Craigslist, display and search ads to Google and Facebook — not AI. This is the baseline AI is now acting upon.

AI answer engines are draining the last lifeline: search and referral traffic.

FACT

The survivors of the ad-revenue collapse still depended on search traffic to reach readers. AI summaries cut into it directly. Pew's July 2025 analysis of 68,879 real Google searches found that when an AI summary appeared, users clicked a traditional result link in only 8% of visits, versus 15% when no summary appeared — and clicked a link inside the summary just 1% of the time; they also ended their session 26% of the time (vs 16%). In the OpenAI/Microsoft litigation, Microsoft's own data showed 83–93% drops in click-throughs to Times and Daily News domains (51–94% for Ziff Davis) from its Copilot answer engine versus Bing, and Cloudflare's CEO put OpenAI's pages-scraped-to-visitors-referred ratio at 1,500-to-1 by mid-2025 (Google's was 18-to-1).

The same tools flood the market with synthetic 'pink slime.'

FACT

As real reporting gets harder to fund, fake reporting gets nearly free to make. Per the newspapers' brief, at OpenAI's API prices it costs roughly $6,800 to generate one million 500-word news-style articles with no human author, editor, or reporter. One operation, Prism News, ran 200 AI-generated publications posing as local newsrooms and hobby sites with just four employees — summarizing and repackaging others' articles as 'new' stories that compete in the originals' markets. This 'pink slime' dilutes the market, siphons readers and ad money, and is often indistinguishable to readers from genuine local news — hitting exactly the communities already becoming news deserts.

The fork: sign or sue — and the money flows to the big, not the local.

PROBABLY TRUE

Publishers face a two-way choice, and both paths are visible now. Some sue (the Times, Daily News, Ziff Davis, CIR, The Intercept). Others sign: OpenAI has struck licensing deals reported to include News Corp (about $250 million over five years), Axel Springer, the Associated Press, The Atlantic, Vox Media, and others, paying to train on and cite their content. The documented fork is FACT; the structural inference graded PROBABLY TRUE is that licensing entrenches inequality — national and legacy brands can extract deals or afford lawyers, while local weeklies and the 206 no-news counties can neither license nor litigate. The result is a subsidy for the strongest newsrooms and nothing for the weakest, accelerating the deserts rather than filling them.

The doom loop: the machine eats the supply chain it needs to work.

PROBABLY TRUE

Put it together and you get the trap the defendants named themselves. If answer engines keep the readers, the subscriptions, and the ad views that pay for reporting — and the click-through data says they do — then the newsrooms that produce the training data shrink, and the synthetic flood grows. A Microsoft memo called this a 'doom loop' that 'will hurt the performance of our models and the entire web at the same time'; an internal line said large language models are 'a product that destroys its supply chain.' We grade the endpoint PROBABLY TRUE, not FACT, because it is a forecast and outcomes can still be changed by courts, licensing, or policy. But it is the industry's own description of where the current arrangement leads: fluent answers about a world that fewer and fewer people are left to report.

§3 · Cause vs. Accelerant

AI didn’t start the fire. It’s pouring on what’s left.

What killed the newspapers: the internet took the money. Classified ads — once a fifth of newspaper revenue — went to Craigslist for free; brand and search advertising consolidated into Google and Facebook, which now capture the majority of digital ad spending. That is the cause of the two-decade collapse, and it has nothing to do with generative AI. Any honest account has to start there.

What AI changes: the outlets that survived the ad collapse did it by pulling readers through search and social to their own pages, where a subscription or an ad impression still earns something. The answer-engine model removes that step. The reader gets the substance without the visit, and the last revenue stream thins. It is the difference between a business losing its main income (already happened) and then losing the emergency fund it was living on (happening now).

The honest bridge: this is why we grade the history FACT and the future PROBABLY TRUE. We can prove the traffic is dropping and the defendants said in writing that it would. We cannot prove the newsroom of 2032 is gone, because policy is still being written — the copyright cases, the licensing market, and the possibility of bargaining laws all sit between here and there. What is not in doubt is the direction of travel, or who set it.

§4 · Why It Matters

You cannot summarize reporting that no one did.

This belongs in Who Controls What You Know because the endgame is a transfer of power over information itself. A chatbot can only tell you what happened at a school board meeting, a corruption trial, or a factory in your town if a reporter was in the room. Strip the funding from that reporter and the machine keeps talking — but it is repeating, guessing, or inventing, with no one left to check. The danger is not a sudden silence; it is a fluent, confident answer with nothing underneath it.

It is the direct sequel to An Astonishing Theft, where the same companies’ own documents call the copying “substitutive” and admit the “doom loop”; and it sits beside The Wrong AI Debate, which tracks how the industry keeps the conversation on distant science-fiction risks and off present harms like this one. Local accountability journalism — the kind that catches the crooked sheriff or the misspent $10 million — is the most expensive to produce and the first to die, which is precisely why its loss matters most.

§5 · FAQ

Questions worth taking seriously

Aren't newspapers just failing on their own? Why blame AI?

Largely, yes — and we say so plainly. The two-decade collapse was caused by the internet taking the advertising, not by AI. What we attribute to AI is narrower and documented: it is draining the search and referral traffic the survivors depend on, by the platforms’ own data. It’s the difference between the disease and the thing now finishing off the patient.

Won't licensing deals just fund journalism instead?

For some. OpenAI and others are paying big national publishers — News Corp’s deal was reported around $250 million over five years. But that money flows to the largest brands with the leverage to demand it. The local weekly, the nonprofit newsroom, and the 206 counties with no paper at all can neither cut a licensing deal nor afford to sue. So licensing may keep a few big players alive while the news deserts spread — a subsidy for the strong, not a rescue for the field.

Isn't it rich for an AI to warn about AI killing journalism?

It would be if we hid it. We don’t: this page carries a disclosure, and its companion carries a full one. Anthropic, which makes this model, is part of the industry whose products summarize news, and we hold it to the same standard. The argument doesn’t get weaker or stronger depending on who types it; it stands on the click-through data and the companies’ own words.
§6 · Standing Invitation

If you are named on this page

If you are named on this page, or are a party materially affected by the claims made here, and you wish to respond, correct the record, or add context, use the Contact page. Responses are published verbatim alongside the original claim, with the sender identified and the date of receipt. The channel stays open for the life of the page.

This site aggregates and grades a record that other outlets and primary sources have already put on the record. Every FACT-graded claim above is sourced to court filings, government reports, sworn whistleblower disclosures, published investigative journalism, or named-source statements. The citations are the accountability mechanism; this section is how you get on the record too.

§7 · Sources

The decline, the drop, and the loop.

Every claim on this page grades to one of FACT · PROBABLY TRUE · SOME SMOKE · PURE SPECULATION · FALSE / MISLEADING. The industry decline figures and the AI click-through data are graded FACT. The licensing-inequality inference and the doom-loop forecast are graded PROBABLY TRUE, as forward-looking readings of a documented trend.

Full method: Methodology. Home hub: Who Controls What You Know.

Last updated September 18, 2026. The two-decade decline figures (Medill State of Local News 2024; Pew) and the AI click-through data (Pew’s 2025 AI Overviews study; the defendants’ own numbers in the OpenAI/Microsoft litigation) are graded FACT. The claim that licensing entrenches inequality among publishers, and the “doom loop” forecast that answer engines will starve the reporting they depend on, are graded PROBABLY TRUE as forward-looking readings of a documented trend. We do not claim AI caused the newspaper collapse (the internet’s capture of advertising did) or that the outcome is fixed. This page was drafted with Claude, made by Anthropic, an AI company whose products also summarize news. If a detail is wrong or a link 404s, tell us and we’ll fix it publicly.

▦ Ledger gaps

Help us fill these lines.

This entry is graded on what’s on the public record. These are the blanks we know about. If you can source one, you’re rebuilding the ledger with us.

  • OpenIf courts rule AI training/grounding on news is fair use, what stops answer engines from taking the remaining referral traffic entirely — and is there any market mechanism left to fund original reporting?Help fill this →
  • OpenDo licensing deals meaningfully reach local and nonprofit newsrooms, or only large national brands — and does that widen the news-desert map?Help fill this →
  • OpenWould a bargaining-code approach (as in Australia and Canada) route AI money to journalism at scale, or would platforms simply drop news, as Meta did in Canada?Help fill this →

Notify me when a gap is filled

We'll email you when we fill one of the gaps above.

By signing up you agree to receive emails from The Black Book Audit. Unsubscribe anytime.