Meta quietly updated its commercial terms to make broad web scraping for AI training official policy, and I keep staring at the secondary effects. We are building an information engine designed to digest reality, but we are cutting off the supply of reality itself. If an AI consumes the news, summarizes it, drives human traffic away from the source, and then trains on the resulting echo, what does ground truth look like five years from now?

I want to understand the exact mechanics of this feedback loop because it feels less like corporate greed and more like an accidental ecological disaster.

The Disappearing Footprint of Primary Sources

Consider how an actual fact enters human civilization. A reporter sits in a municipal zoning hearing in Ohio for six hours. Someone sifts through 900 pages of bankruptcy filings from an energy company. An investigative team spends $50,000 and eight months tracing supply chains across three continents. That is raw, primary-source labor. It is expensive, inefficient, and deeply human.

Now look at the modern distribution pipe. Meta’s family of apps sent roughly 30% to 40% less referral traffic to major news sites over the course of 2023 alone, following deliberate platform tweaks to demote external links. When an AI crawler absorbs the reporting without sending back the human reader, the economic model funding that original six-hour zoning board attendance simply vanishes.

What replaces it? Cheap, high-volume secondary summaries. An algorithm rewrites the local scoop in three seconds. A content farm publishes four variations to capture search intent. If Meta's systems crawl the web tomorrow, they aren't just ingesting the original investigation anymore. They are ingesting hundreds of slightly degraded, synthetic paraphrases of that investigation.

The Math Behind Model Autophagy

Researchers at Rice and Stanford coined a term for what happens when generative models train on the output of other generative models: Model Autophagy Disorder, or MAD. It turns out that without enough fresh, real-world data, synthetic models experience cognitive collapse. The variance shrinks. Nuance flattens into an uncanny, beige consensus.

Think of it like making a photocopy of a photocopy. The first pass loses the subtle grain of the paper. By the fifth pass, you are left with illegible black smears and phantom artifacts. If Meta scrapes an internet increasingly dominated by automated aggregators, the training data becomes a hall of mirrors.

  • The original event happens in physical space.
  • A diminishing pool of humans documents it imperfectly.
  • Thousands of bot-operated blogs rewrite the human account to capture algorithmic yield.
  • Large foundational models ingest both the human account and the thousands of echo-chamber iterations.
  • The next model update weighs the majority narrative (the echoes) over the solitary source.

I wonder whether we will even notice when it happens. The grammar will still be flawless. The tone will remain authoritative. But the underlying connection to what actually happened on the pavement will have snapped completely.

The Weird Economics of Digital Strip-Mining

There is an undeniable irony in treating journalism as a raw mineral deposit. If an oil company extracts petroleum from a field, it knows the field is finite. Tech giants seem to view human-generated cultural output as an infinite spring that will keep bubbling up regardless of whether anyone pays for the plumbing.

When Meta shifts its legal posture to unapologetically crawl news organizations, it treats the reporting as ambient data—like ambient temperature or background radiation. But news is not a natural phenomenon. It does not occur spontaneously. If you don't pay the person reading the court docket, the court docket does not get read. The ambient data ceases to exist.

So what happens when the well runs dry? Does Meta eventually have to fund proprietary newsrooms just to generate organic tokens for its next model iteration? That sounds absurd, yet OpenAI has already signed multimillion-dollar licensing deals with publishers like Axel Springer and the Associated Press. Meta appears to be gambling that public web scraping and sheer computational scale will bypass the need to pay the piper. It makes you wonder which side of that bet understands information ecology better.

What This Actually Means

The real hazard here isn't merely that legacy newspapers go out of business. It is that our collective record of current events gets replaced by consensus fiction. When recursive models eat zombie news, they don't produce obvious errors like "the moon is made of cheese." Instead, they produce subtly sanitized, averaged-out versions of history where anomalous truths get smoothed away because they don't fit the statistical pattern.

We are running an unprecedented cultural experiment: systematically starving the apparatus that documents reality, while building the most sophisticated narrative-generation engines in human history. The incentives push entirely toward generation and entirely away from verification.

If the primary sources go quiet, we won't get silence. We will get an endless, deafening murmur of models quoting models that quoted models that once heard a rumor about an event nobody was paid to witness.

Quick Answers

Did Meta explicitly say they don't care about journalists?
No, Meta frames their crawler policies around standard open-web indexing and building helpful consumer utilities. The systemic squeeze on original reporting is a structural byproduct of how web scraping intersects with platform economics, not a declared vendetta.

Can news sites just block Meta’s crawlers?
Publishers can use robots.txt to tell Meta's user-agents to stay out, but it is an imperfect defense. It does not stop third-party scrapers from copying the work and funneling it back into the training pool under different domains.

Why can't AI just verify facts against the original source?
LLMs do not evaluate external reality; they evaluate probabilistic relationships between words. When thousands of automated mirrors repeat a slight distortion of a fact, the model's mathematics interpret that sheer frequency as truth.