2.3.2 Information Ecosystems

David used to be a reporter for a mid-sized regional newspaper in Ohio. He covered city council meetings, local business openings, high school sports. Unglamorous work, but essential. Someone had to sit in the back of the room while the zoning board voted, so that the people who lived under those decisions would know they had been made.

The paper folded in 2023. Private equity had bought it five years earlier, gutted the newsroom, squeezed every remaining dollar of profit, then let it die when the revenue dried up. David and six other reporters were laid off. The town—population 47,000—no longer had anyone covering it.

In 2025, a site calling itself the "Maplewood Daily Gazette" appeared. It published dozens of stories a day: council recaps, weather, business announcements. The prose was grammatically clean, factually thin, and generated entirely by software scraping public records. Some residents never noticed. Others grew suspicious only when they realized every article carried the same weightless voice, that the "staff writers" on the masthead did not exist, and that the headshots beside their bylines were synthetic faces that had never blinked.

David tried a Substack. He published investigative pieces—city corruption, a development conflict, a contaminated creek. Real journalism. Almost no one subscribed. The Gazette was free, updated hourly, and tuned for search engines; it owned the top of every Google result for "Maplewood." David's reporting was invisible beside it. By the end of 2025 he had moved to Columbus and was driving for DoorDash. Maplewood still had news of a kind. It had lost journalism.

That distinction—between something that looks like news and something that does the work of news—is the subject of this chapter. It turns out to be a distinction with structural teeth.

What the machines make, and what they cannot

By May 2025, NewsGuard, a firm that rates the reliability of news outlets, was tracking more than 1,200 AI-generated news and information sites operating with little or no human oversight, publishing in 16 languages—a more than twentyfold jump in roughly two years (NewsGuard, 2025). These are not newspapers that have adopted a new tool. They are a new kind of object: content mills that aggregate public data—council minutes, police blotters, school board agendas, press releases—and reprocess it into the shape of articles.

The word "shape" is doing real work there. What separates these operations from journalism is not the quality of their sentences, which can be perfectly serviceable. It is the set of things they structurally cannot do.

The asymmetry is not subtle. The mill wins decisively on cost, speed, and volume. The newsroom wins on everything that produces original knowledge about the world—a whistleblower who will only talk to a person, a document that has to be pried loose, a pattern that only becomes visible to someone who attended forty tedious meetings. An algorithm can summarize the minutes of a council vote. It cannot notice that the councilman who pushed the rezoning owns the parcel next door, because that fact is in no dataset until a reporter puts it there. The machine can restate the record. It cannot generate a record that did not exist. That is the line, and no amount of fluency crosses it.

Some operators have tried to blur the line deliberately. Hoodline, a hyperlocal outlet, deployed AI to cover underserved neighborhoods but dressed its bots in human personas—fabricated headshots, invented biographies—and when readers discovered they had been reading machines wearing human masks, the trust didn't transfer to the technology; it drained out of local journalism generally (Nieman Lab, 2025). Trust, once spent this way, does not easily refill.

The feedback loop

The reason AI does not merely mimic misinformation but manufactures more of it is mechanical, not incidental. Large models are trained on vast scrapes of the internet—news, forums, social posts, blogs. Because falsehood is abundant online, it enters the training set. The model learns the statistical shape of that content, false claims included, and reproduces it in generated text. That text gets published, indexed by search engines, and eventually scraped into the training data for the next generation of models. Researchers call this the misinformation feedback loop, and its defining feature is that each pass makes the falsehood more fluent and more widely distributed—so a claim that began as a fringe forum post can come back around as a confident, well-formatted paragraph that reads like settled fact (Bulletin of the Atomic Scientists, 2025).

graph LR
  A[False claim online] --> B[Scraped into training data]
  B --> C[Reproduced in AI output]
  C --> D[Published and indexed]
  D --> E[Amplified on platforms]
  E --> B
  D --> F[Re-scraped for next model]
  F --> B

The loop runs on a second track, too—through humans. When an AI-seeded story starts trending, a surviving newsroom faces a choice: publish fast to catch the traffic it needs to stay solvent, or verify carefully and lose the moment. The economics reward speed, speed produces error, and the error feeds the machine track. Verification tools have improved—Agence France-Presse and partners built systems like Vera.ai and WeVerify to help fact-checkers—but the volume of fabricated material continues to outrun human capacity to check it (AFP).

The synthetic outrage machine—and how much it actually moves

AI's fabrication capacity reaches past filling news deserts into manufacturing sentiment. In August 2025, analysts examining the online furor over restaurant chain Cracker Barrel's logo redesign concluded that a large share of the apparent outrage was synthetic—bot accounts and AI personas inflating a minor design change into a national controversy (Wired, 2025). The infrastructure to counterfeit public opinion at scale is operational. Foreign influence operations have used AI to seed divisive narratives tailored to local idioms, and the same tooling can flood a competitive district with content that imitates independent local reporting while carrying a single campaign's frame.

Here honesty requires a firm distinction the alarming headlines tend to erase. That these operations exist, and that their volume is enormous, is well documented. Whether they change outcomes—votes cast, minds actually altered—is not. This is one of the genuine open questions of the field. A synthetic controversy can trend without a single wavering voter changing a ballot; measured effects of online persuasion campaigns on real political behavior have generally been small, and researchers have not yet isolated a case where an AI-driven operation demonstrably swung an election. The capability is real and the reach is vast. The proven downstream effect on beliefs and behavior remains modest and hard to measure. Both halves of that sentence are true, and a serious account has to hold them together rather than collapse into either the panic that assumes influence follows automatically from volume or the complacency that dismisses a working capability because its body count is not yet visible. The likeliest near-term harm may be indirect: not that any given fake persuades you, but that the knowledge that such personas are everywhere corrodes your willingness to trust anything.

The trust gap, and the survey's blind spots

The 2025 Reuters Institute Digital News Report offers a revealing pair of numbers. Asked how they would check a suspect claim, 38 percent of respondents said they would turn to a news source they trust—the top answer. AI chatbots came last, at 9 percent (Reuters Institute, 2025). People still rank human institutions above machines when they consciously reach for the truth.

The reassurance and the alarm live in the same finding. Reassuring, because institutional credibility still counts for something. Alarming, because the trusted institutions are the ones disappearing—the local papers, the regional desks—while the least-trusted source, the chatbot, is increasingly what a search actually returns. The gap between where people say they would turn and what is in fact in front of them is widening.

It is worth being skeptical of the instrument itself, because the book's honesty demands it. Self-reported trust is not behavior. A survey captures what people say they do when they stop to reflect; it cannot easily capture the scrolling, half-attentive, algorithmically-fed reality of how information actually reaches them. Someone can tell a pollster, sincerely, that they distrust chatbots and trust their local paper, then spend the evening absorbing AI-summarized search snippets and platform video without registering either the source or the contradiction. The Reuters report is the best recurring instrument we have, and it is indispensable for tracking trends over time. But it measures stated attitudes, and the degradation of an information ecosystem happens largely below the level of stated attitude, in the ambient defaults of what gets served and what gets seen.

Why journalism's economics broke, and why the fixes don't scale

The collapse predates generative AI, which matters for assigning blame honestly. Print advertising—the subsidy that once paid for the city hall reporter—migrated online two decades ago and then fragmented across platforms that pay content producers a pittance. Subscriptions rescued a handful of national brands with the reach to sell them: the New York Times, the Washington Post, the Guardian. They have never worked for a paper in a town of 47,000, because the addressable audience is too small and the willingness to pay too thin. Philanthropy and nonprofit newsrooms have plugged specific gaps in specific markets, but they depend on donors and grants that do not renew reliably and have never approached the scale of the commercial industry they are meant to replace. The arithmetic is unforgiving: covering a community adequately costs roughly what it always did, and none of the replacement revenue streams produce that sum at the number of communities that need it.

AI enters this wreckage as both tool and solvent. Used transparently—transcribing interviews, drafting routine recaps, sifting a dataset while a human keeps editorial control—it is genuinely useful, and outlets that disclosed such use in 2025 largely kept their credibility (Poynter, 2025). But the financial logic does not stay in its lane. If a model can draft the story, the case for paying a reporter to draft it weakens; if a model can analyze the spreadsheet, the data journalist's job gets harder to defend in a budget meeting. The tension between AI-as-tool and AI-as-replacement is not settled by good intentions. It is settled by economic pressure, and economic pressure reliably prefers the cheaper option.

Algorithmic gatekeepers

Most people in wealthy countries now meet the news through intermediaries—Google Search, Meta's feeds, TikTok, YouTube, X—that rank content not by editorial judgment but by predicted engagement. Clicks, shares, dwell time, emotional reaction: these are the signals, and AI content is engineered to hit them. It is formulaic, fast, and tuned to provoke the reactions the metrics reward. Good journalism performs worse against precisely these measures, and the reason is not incidental—it is a property of the thing itself. Careful reporting is often nuanced where the algorithm rewards certainty, contextual where it rewards immediacy, measured where it rewards outrage, and slow where it rewards volume. The qualities that make journalism trustworthy are the same qualities that make it underperform in an auction for attention. Platforms have tried modest corrections—boosting "authoritative sources" or "original reporting"—but these sit atop an incentive structure in which ad revenue scales with engagement, and that structure is untouched. As synthetic content grows better at imitating the emotional signature of real reporting, the visible layer of the information environment drifts further from the reliable one.

What people do inside the wreckage

People do not passively absorb this; they adapt. The trouble is that the three common adaptations are each, in their own way, corrosive.

The first is partisan enclosure. When institutional credibility fails, alignment becomes a shortcut for reliability: what my side says is treated as true, what the other side says is discounted regardless of its accuracy. The result is not a population uniformly deceived but one that believes different, incompatible things by group, which makes shared deliberation nearly impossible.

The second is radical skepticism—not the healthy demand for evidence but a reflexive assumption that every institutional source is manipulated. In this frame the absence of official confirmation reads as proof of suppression rather than proof of falsehood, which is exactly the soil in which conspiracy grows. What was once a fringe posture becomes almost reasonable when it really is true that a large share of what you encounter is fake.

The third is withdrawal. News avoidance has risen across countries and demographics, driven by exhaustion, distrust, and the emotional weight of a feed dominated by conflict and manufactured fury. The avoidant are not safe—they still swim in algorithmic content—but they are less likely to seek corrections, follow civic processes, or participate in the informed deliberation that self-government assumes.

Each response makes democratic accountability harder. And the specific losses in a news desert are concrete, not abstract: no one attends the zoning board, so development deals go unexamined; no one reads the audit, so the misspent funds stay misspent; local turnout and split-ticket voting decline while municipal borrowing costs rise, because lenders price in the corruption that no reporter is watching for. Research on communities that lost their papers has documented exactly these effects. The vacuum does not stay empty. It fills with the content mill, the partisan Facebook group, and the campaign's own messaging—sources with every incentive except the one that matters.

Is this new? A sense of proportion

Every information technology has arrived trailing a moral panic, and most panics were partly wrong. The penny press and yellow journalism of the late nineteenth century were sensational, dishonest, and blamed for stampeding a country toward war; the republic survived and the profession eventually professionalized. Radio in the 1930s carried demagogues into every kitchen and was thought to hold uncanny hypnotic power over listeners; that power was real but smaller than feared. Early social media was hailed as a democratizing force and then indicted for polarization and manufactured consent. History counsels humility: we have overestimated the persuasive potency of new media before.

But two features of the present disruption do not have clean historical parallels, and pretending otherwise would be its own dishonesty. The first is that earlier disruptions were about distribution—new ways to spread content that humans still had to produce. Generative AI attacks production itself, collapsing the cost of manufacturing plausible content to near zero, which removes the natural volume ceiling that even yellow journalism obeyed. The second is the closed loop: no prior technology fed its own output back into the mechanism that generates the next round of output. Tabloids did not train on tabloids. The reassurance from history is that societies have absorbed information shocks before. The warning is that the two things making this shock distinctive are precisely the two things earlier absorptions never had to cope with.

The threshold nobody can mark exactly

There is a recovery problem lurking here, and it has the structure of a ratchet. If AI content comes to dominate search results and social amplification, and if that dominance starves the surviving newsrooms of the audience and revenue they need to exist, then the human reporting required to correct the machine layer disappears—and with it the fresh, verified, human-generated material that models need to stay tethered to reality. Past some threshold, the loop closes on itself and quality journalism cannot recover on its own, because the economic base that would fund a recovery has been consumed. No one can name that threshold with a number; the honest position is that it is a real risk with an unknown trigger point, and that the cost of crossing it—an information supply that recycles itself indefinitely with no external check—is high enough that waiting for certainty is its own decision.

What isolating AI's role actually requires

A hard epistemic caveat runs under this whole chapter. Misinformation, trust collapse, and polarization all predate ChatGPT. Local newspapers were dying in 2010. Trust in media was falling in 2015. Partisan enclosure was well underway before any of this generation's models existed. So how much of the present degradation is caused by AI, as opposed to accelerated, decorated, or merely coincident with it?

Researchers cannot yet cleanly separate the signal. The trends are confounded: the same decade that produced generative AI also produced platform consolidation, a pandemic, and acute political shocks, and no controlled experiment can isolate one strand at societal scale. The defensible claim is narrow and worth stating precisely. AI did not create the crisis of information; it arrived into a crisis already underway and supplied it with new capabilities—near-zero-cost production, the self-training loop, synthetic personas at scale—that plausibly worsen it and certainly change its shape. Anyone who tells you they have measured AI's exact contribution to the degradation of the public sphere is overselling. The mechanisms are visible and real. The magnitude, honestly, is not yet known.

What would actually help

Because the disease predates AI, so do the remedies, and they cluster in three places.

Public investment as civic infrastructure. The strongest evidence that funding journalism works comes from comparison: countries with well-funded, independent public broadcasters tend to show higher levels of news trust, more shared factual common ground, and lower vulnerability to disinformation than countries without them. That points toward subsidy models—public broadcasting insulated from political control, tax credits for employing local reporters, nonprofit news endowments—justified not as a cultural amenity but as the same kind of public good as a functioning court or an honest census. The argument is civic: a community without journalism cannot govern itself, and the market has demonstrably declined to supply it.

Platform obligations. The intermediaries that now distribute most news are the point of maximum leverage. The candidate obligations are concrete—algorithmic changes that stop systematically burying original reporting, transparency about how ranking works, and revenue-sharing arrangements that return some of the value platforms extract from news to the people who produce it. Australia's bargaining code and similar European measures are early, contested experiments in exactly this; their results are mixed and their design flawed, but they establish the principle that the relationship between engagement optimization and informational quality is a matter for policy, not just corporate discretion.

Disclosure—and the question of who owes it. Should AI-generated content be labeled? The case for yes is strong, but the interesting question is who bears the duty. The AI company can watermark or tag outputs at the source; the platform can detect and flag synthetic material as it circulates; the publisher can disclose when it presents machine-written work as its own. The most durable answer is that the obligation is layered rather than singular—the producer marks, the distributor surfaces the mark, and the publisher owns the disclosure to the reader—because any single point of labeling can be stripped by the next actor in the chain. A watermark the platform ignores is useless; a platform label on content the publisher has laundered as human reporting is a half-measure. Disclosure that survives contact with the real supply chain has to be redundant.

None of these is speculative. They are understood in principle and, in places, tested in practice. What is missing is not the design but the will to deploy them at the scale the problem now demands—before the ratchet clicks past the point where the reporting needed to sustain a recovery no longer exists to fund it.

Key Takeaways

1. The line is structural, not stylistic. More than 1,200 AI-generated news sites were operating in 16 languages by mid-2025. They can restate the public record fluently; they cannot generate a record that did not exist—cannot cultivate a source, attend the meeting, or be held accountable when wrong. That is the permanent difference between content and journalism.

2. The feedback loop is real and mechanical. False content is scraped into training data, reproduced in output, published, and re-scraped—each pass more fluent and harder to debunk. No prior information technology fed its own output back into its own production.

3. Volume is proven; persuasion is not. Synthetic influence operations exist and operate at scale (the Cracker Barrel outrage was substantially fake). Whether they move votes and beliefs remains genuinely unmeasured. Hold both facts at once.

4. Trust surveys reveal a widening gap—and have blind spots. People rank trusted news sources far above AI chatbots (38% vs. 9%) for checking claims, yet the trusted institutions are the ones vanishing. Self-reported trust also fails to capture ambient, algorithmically-fed behavior.

5. The economics broke before AI, and the fixes don't scale. Subscriptions save a few national brands; philanthropy patches specific markets; neither funds coverage of ordinary communities. AI enters as both a useful tool and a solvent that dissolves the case for paying reporters.

6. Engagement algorithms disadvantage good journalism by design, because nuance, context, and slowness—the sources of its reliability—are exactly what the attention auction penalizes.

7. AI accelerated a crisis it did not create. Its precise contribution cannot be cleanly isolated from platform consolidation, polarization, and the two-decade decline of local news. The mechanisms are visible; the magnitude is not yet known.

8. The remedies are known, layered, and unfunded: public-broadcasting-style investment as civic infrastructure, platform obligations on ranking and revenue-sharing, and redundant AI-disclosure duties spread across producer, distributor, and publisher. What is lacking is the will to deploy them before the loop closes.

Sources

• AI Tracking Center — NewsGuard
• 2025 Reuters Institute Digital News Report
• AI Is Polluting Truth in Journalism — Bulletin of the Atomic Scientists
• How Will AI Reshape the News in 2026? — Reuters Institute
• News Outlets That Got AI Right in 2025 — Poynter
• Hoodline's AI-Generated News Experiment — Nieman Lab
• Synthetic Social Media and the Cracker Barrel Controversy — Wired
• AFP Verification Tools: Vera.ai and WeVerify
• AI Misinformation and the Value of Trusted News — CEPR
• Chaos and Credibility: How AI Is Impacting Press Freedom — GIJN

Last updated: 2026-08-03