What does AI-generated content do to media economics and trust — the attention economy's K-shape
Why AI Is Splitting the News World in Two — and Why Fixing It Is Harder Than It Looks
Based on analysis of a 113-node, 412-edge knowledge graph mapping the economics, trust dynamics, and feedback loops of AI-generated content in media.
What Are We Actually Talking About?
Imagine the media world is a swimming pool. For a long time, most people swam in roughly the same water. Some reporters worked for big newspapers, some for small local ones, but the pool was shared.
Now picture someone dumping an enormous bag of sand into that pool. The sand is cheap, AI-generated content — articles, videos, posts — produced by the millions every day at almost no cost. Some of the water stays clean at the deep end, where expensive lifeguards (editors, fact-checkers, trusted brand names) keep the sand out. But at the shallow end, the water gets murky fast.
This is what researchers call the “K-shape” — one arm going up, one going down. Prestigious outlets get more subscribers. Local newspapers close. The pool splits into two separate pools.
The graph we analyzed maps exactly how this happens and why it is so hard to stop.
The K-Shape Is an Effect, Not a Cause
Here is the first non-obvious finding: the K-shape itself does not cause anything. It is the result of at least nine separate forces all pushing in the same direction at the same time.
Think of it like a traffic jam. If you ask “why is there a traffic jam?”, the answer is not “because the cars aren’t moving.” That describes the jam, it doesn’t explain it. The real answers are: there’s an accident, plus a lane closure, plus it’s rush hour, plus the on-ramp merges badly.
The graph shows nine separate “accidents” all hitting the media highway at once:
- AI content is now almost free to produce, so there is vastly more of it
- Search engines now answer questions directly, so fewer people click through to read the original article
- Most advertising money has drained from news websites to Google and Meta
- It is increasingly hard to tell what is real online, so people trust less
- Local newspapers have closed en masse, leaving civic information deserts
Each of these would cause problems on its own. Together, they produce the K-shape. Calling it “the K-shape problem” is a bit like calling the traffic jam “the cars problem” — technically accurate, but it points you at the symptom rather than the causes.
The Machine That Feeds Itself
Several of these forces do not just add up — they loop back and make each other worse.
Here is a simple one: cheaper AI generation means more AI content gets made. More AI content means more AI content ends up in the training data used to build the next generation of AI. AI trained on AI-generated content produces lower-quality outputs. Those lower-quality outputs get published online. They end up in the next round of training data. The circle tightens.
This is like a photocopier that copies its own copies. The first copy of a document looks fine. The copy of the copy is a little blurrier. By the twentieth generation, you can barely read it.
Another loop involves search engines and advertising. When Google started answering questions directly in search results, fewer people clicked through to news sites. News sites lost the traffic, which meant they lost the advertising money that paid their reporters. Less revenue meant less original reporting. Meanwhile, the advertisers who left news sites moved to Google and Meta, which used that money to build better AI search — which answers even more questions directly, which drives even fewer people to click through. The two effects fund each other.
The Remedy That Defeats Itself
The graph contains an internal contradiction that is worth slowing down to understand.
There is a formal standard called C2PA — think of it as a kind of nutrition label for digital content. It would let you verify whether a piece of content was made by a human or generated by an AI, and what AI produced it. The analysis shows this standard has been formally identified as the solution to the problem of not being able to tell real from fake.
But the same graph shows why the solution probably will not work: open-source AI.
When a company like a major tech firm releases a version of its AI as free, open-source software, anyone can download it, modify it, and use it without restrictions. If a provenance standard requires AI companies to tag their outputs, open-source models can simply be modified to remove the tag. You cannot regulate a piece of software that anyone can copy and run on their own computer.
The graph encodes this as a direct structural defeat: the same open-source economics that made AI generation cheap enough to flood the internet also make any content-labeling standard effectively optional for anyone who wants to opt out.
So the identified fix and the identified problem share the same root cause. The cost curve that drives the flood also defeats the levy designed to hold it back.
The Hidden Path from AI Content to Your Social Security
One of the less obvious paths the graph traces goes somewhere unexpected: from AI-generated content to the financial solvency of Social Security.
Here is how the chain works, step by step.
AI is very good at doing the tasks that used to be entry-level creative jobs. Fact-checking, copy editing, basic graphic design, writing product descriptions, captioning images — these were the jobs that young people took to get started in media and communications. AI now does most of them at near-zero cost. Those jobs are disappearing.
When you work, you pay a tax called FICA, which funds Social Security. If an entire generation of young workers in certain industries loses their foothold on the career ladder, they make less money over their lifetime, which means they contribute less in FICA taxes. Less FICA tax means the Social Security trust fund gets depleted faster than official projections expect.
But the chain continues. Economic precarity — not having enough money, not having stable work — makes people more vulnerable to misinformation. When you are stressed and financially insecure, you have less time to verify what you read, and you may be more drawn to explanations that confirm what you are already feeling. The graph labels this the “epistemic poverty trap”: being economically poor also tends to make you informationally poorer, because quality journalism costs money to access and time to evaluate.
So: AI content reduces entry-level media jobs, which reduces lifetime earnings for a generation, which reduces Social Security funding, which deepens economic precarity, which expands the audience most vulnerable to AI disinformation. A technology shaping how news is made ends up connecting to how retirement is funded.
The Clean-Water Escape and Its Ceiling
The obvious hope in this picture is the “top arm” of the K-shape. Some outlets are actually doing better. The New York Times has millions of digital subscribers. Substack writers make real incomes. Podcasters with loyal audiences can fund themselves through listener support.
This is real, and the graph acknowledges it. But it also traces two problems with counting on it.
First, there is a ceiling. Human beings only have so much money to spend on subscriptions. When you are already paying for streaming video, music, a news outlet, a podcast network, and maybe two or three Substack writers you like, there is a limit to how much more you will add. The graph calls this “subscription fatigue.” The top arm can grow, but it cannot grow without bound.
Second — and this is the less obvious problem — the top arm makes the bottom arm worse in a specific way. When high-quality information is only available to people who can pay for it, the people who cannot afford it end up in a lower-quality information environment. The graph encodes this directly: the mechanism that rescues some journalism from the flood simultaneously deepens the information gap for people who cannot pay. The solution to the problem at the top creates a new version of the problem at the bottom.
The Nodes That Matter More Than They Look
The graph has a structural quirk worth mentioning: three of the most connected points in the entire map have been assigned the minimum importance score.
Imagine a road map where the city that has the most highways running through it is labeled as a minor rest stop. The “Open Web Value Extraction Loop” — the overall dynamic by which platforms extract value from content without compensating the people who made it — has 26 connections to other concepts in the graph, but a weight of 1 out of 10. Same with “Liar’s Dividend Epistemic Trap” (the idea that widespread deepfakes make all real evidence deniable, even by people caught on camera doing things) and “Narrative Economics Viral Contagion” (the economic structure that makes false stories spread faster than true ones).
These appear to be older concepts — ideas that were identified before AI entered the picture — that were included in the map but never had their importance scores updated as the analysis developed. Structurally, they sit at the center of the map. By their assigned scores, they look peripheral.
This matters for anyone trying to figure out where to intervene. If these nodes are actually as important as their connectivity suggests, then targeting the pre-AI dynamics they represent — platform value extraction, deepfake denialism, viral story economics — might produce larger effects than the current scoring would predict.
The Bottom Line
The graph’s structural findings, taken together, point at several things that are not obvious from reading news coverage of AI:
The K-shape is downstream. It is a result of many separate mechanisms, not a cause. Calling something a K-shape problem does not tell you which mechanism to fix.
The feedback loops are real and closed. Several of the individual forces are locked into self-reinforcing cycles. Cheaper AI generates more slop, which contaminates training data, which generates cheaper slop. Platform advertising finances the search behavior that drains publisher revenue, which drives more traffic to platforms.
The main proposed technical fix is structurally defeated by the same economics that drive the problem. Content provenance standards require compliance from AI developers. Open-source models can bypass compliance by design.
The corrective mechanisms exist but are constrained. Subscription journalism, patronage models, and content provenance efforts are all present in the graph. Each is bounded: by subscription fatigue, by the class filter it imposes, or by open-source circumvention.
The node that would reduce the epistemic poverty trap does not exist in the graph. The analysis encodes multiple mechanisms that worsen it — economic displacement, the class filter from subscription costs, fiscal stress. No node is encoded that reduces it. The graph does not show a path out of the poverty trap, only paths in.
Pre-AI dynamics may be doing more structural work than they appear. The three lowest-weighted hub nodes suggest the current disruption is running through older, underappreciated infrastructure — value extraction dynamics, deepfake denialism, viral narrative economics — that predates AI but amplifies its effects.
The graph does not predict a particular outcome. It maps which forces are pushing which directions and where they connect. What it shows, structurally, is a system where the problem-generating mechanisms are well-connected and self-reinforcing, and the corrective mechanisms are present but topologically peripheral — real, but not yet strong enough to redirect the main flows.