The great disappearing internet—and what it could mean for your LLM

The last time I published a Waters Cooler, our weekly round-up of industry announcements and news stories, I was tagged in a post promoting the newsletter on Twitter X. According to metrics, the post received 410 views and seven likes since August 8. It also received a few responses.

From @sangpham2005988, or CoinTitan.btc, we got: “count me in @RebNatale the waters are warm for AI moves, what partnerships and launches are you tracking this week at WatersCooler.” From, presumably, CoinTitan’s brother, @sangpham200513 a.k.a. HODLMaestro.btc, we got: “@RebNatale on the WatersCooler – ready for partnerships, acquisitions and AI launches.” Their cousin, @sangpham2005986: “WatersCooler time: partnerships, acquisitions, AI launches.” And from his fraternal twin, @sangpham@2005983: “@RebNatale at WatersCooler: the week’s AI launches and big partnerships on deck.”

This is what it’s like to use the web now. Words are written and posted not by people but by lines of code—or clankers, if you’ve noticed recently that we now have our first slur for robots—which are then engaged by other lines of code. Though the individual sentences often make sense structurally, sometimes even grammatically, they typically amount to nothing. They say nothing, they mean nothing, and they contain only a loose association to the sentence or concept that prompted it.

Caught in the middle of it, there’s all of us. To borrow a relic of the Covid era, it’s the new normal, hardly worth mentioning, and it should be driving us all insane. With a non-zero number of people now experiencing AI psychosis, maybe it is.

What happens when that corpus of democratized data that made the world go ‘round—which underpins your models, workflows, research abilities, writing skills, your own cognition—is a cold, dead corpse?

Oh, let me stop because I know what you’re thinking. Reb, you’re biased. You hate AI. You just want everyone to get a humanities degree. Nobody cares for your contrarian technology takes in a technology publication of all places. But I am a cheerleader for no one (except Taylor Swift—we all have our vices) and nothing. 

As the internet bottoms out, as it is molded in the image of clankers by those who would replace—and are replacing—you to save themselves a nickel to put toward their superyachts, what will become of the panacea you’ve made of large language models? What happens when that corpus of democratized data that made the world go ‘round—which underpins your models, workflows, research abilities, writing skills, your own cognition—is a cold, dead corpse?

If I can make the assumption that you’re not a bot and that you are a technologist, then I’ll further assume you’ve encountered the dead-internet theory, which posits that the internet has been taken over by artificially intelligent bots and has been since 2016—the same year, you’ll note, a red-hot US presidential election brought us the term “fake news” and the Oxford English Dictionary’s word of the year, “post-truth.” According to a study by researchers from Oxford University, Corvinus University, and University of Washington, political bot activity reached an all-time high in 2016, and the use of automated accounts was deliberate and strategic throughout the election.

Imperva, a cybersecurity company that has been tracking online bot activity since 2012, found in its 2025 Bad Bot Report that automated traffic last year surpassed human activity, accounting for 51% of all web traffic, driven largely by the rapid adoption of AI and large language models. In June, internet and cloud company Fastly published lower estimates, finding that more than a third of all observed internet traffic is from automated traffic or bots, with 89% of that bot traffic classified as unwanted.

The Imperva study, which broke up bot activity into good bots and bad bots, also found that bad bot activity rose for the sixth consecutive year, with bad bots accounting for 37%, up from 32%, of all internet traffic. Bad bots are software applications that perform automated tasks with malicious intent, used in scalping, DDoS attacks, and fraud.

While you might think these bots are merely an annoyance, often identifiable and therefore avoidable, I think they’re more like an annoying part of a larger, and more insidious, story of what’s happening to the web.

The images we conjure when we think of museums, libraries, and national archives—for now, at least—seem almost antithetical to the web as it functions today. In any of these institutions, all information contained therein is equally discoverable to those who enter it. On a search engine, discoverability has a range, and it’s dictated not by you or your interests but private companies and what their algorithms and website cookies have determined to be your interests.

These technologies don’t just distort information accessibility; the domino next falls on availability. When users’ interests are formed and then re-enforced by what most easily and directly reaches them, it leaves creators with little incentive to produce material outside those parameters.

All the qualities that once made the internet interesting and useful, and that once made the world feel large yet intimately connected, are being shrunken, broken, and buried under a deluge of such “content.”

And that’s a good word for it. We are contented by it but not informed or moved by it. It’s not news, and it’s not art. It’s not an accurate account of history, and it’s hardly a medium for conversation or good-faith debate (maybe excluding Reddit?). It’s just content—empty calories, a banal approximation of humanness that’s off by miles but measured in inches, soma pills for a brave new world. We accept it not because it’s acceptable, but because no one’s selling us anything else.

That’s not melodrama. Google, the largest search engine, doesn’t allow you to turn off its AI Overview. Under every viral post on Twitter X is a canyon of replies, echoing a tinny “@grok is this real” all the way down. Good luck finding a photograph taken with a physical camera on Pinterest or even Amazon, where you’re supposed to fork over real money for items that only vaguely resemble their product listing. Your Netflix recommendations are curated by algorithms, and the shows seem to have been written by aliens that may have studied humans but have missed something fundamental about the experience.

My editor, Anthony, recently wrote one of his bi-weekly Waters Wraps, the title of which pulled a quote from BNY’s chief executive, Robin Vince: “We invested heavily early on in the psychology of it in the company so that we have AI for everyone, everywhere, with everything. And that’s really how we think about AI.”

That’s how, it seems, everybody’s thinking about AI, but I find the sentiment sinister, if not outright cultish. This stuff is great! But just to be sure, cut off all exit points immediately.

If artificial, low-trust content is what now populates the internet, dwarfing humans and their own contributions—which, by the way, are increasingly informed by this same content—then how is what’s left not some kind of meta wasteland where LLMs are just feasting on their own excrement? What happens when the internet eats itself?

There is a counterargument here, which is that I haven’t been talking about the same kinds of LLMs that banks and asset managers are using in their work. Heavily regulated institutions are building their own models, or they’re paying for purpose-built, enterprise-grade applications from trusted vendors, and the pool of data from which these LLMs draw their responses is vetted and secure. You’re thinking, We’re safe here in our fortress.

When the host dies, it’s bad news for the parasite

But I don’t think you are. Your LLM may not be scraping the web, and your vendors’ LLMs may not be scraping the web, but these applications are nonetheless inextricable from the web. The very blueprint for your safe, proprietary model—whether that’s ChatGPT, Claude, Gemini, Copilot, DeepSeek, or something else—doesn’t exist without the big, wide web. Those Big Tech companies, to borrow a line from Broadway darling “Hamilton,” are in The Room Where It Happens.

It’s why I think efforts to stop GenAI from hallucinating will ultimately fall short. You know the adage, garbage in, garbage out. Have we considered there’s simply too much garbage now? Though I can’t say whether Imperva or Fastly is more accurate in its estimation of bot traffic, it’s either more than half of all internet traffic or approaching half and rising. Most people and companies don’t have the will to clean up litter lying on their street or floating in their city’s rivers; I don’t expect them to de-pollute the web.

If the internet itself is decaying, and the LLM is the internet now, then it’s not a leap to say that LLMs, which are trained on the whole of the internet, are in trouble. When the host dies, it’s bad news for the parasite.

NSFW

The threats don’t stop with bots and hollow, half-true content.

Global indexes that track censorship and press freedoms are showing worrisome indicators. The 2025 RSF World Press Freedom Index shows that the economic indicator, an underappreciated factor in media weakening, now stands at an “unprecedented, critical low.” Where government interference and physical attacks on press are more visible affronts, the economic indicator illustrates how ownership concentration, pressure from advertisers and financial backers, and decreasing public funds are impeding a free press.

You might be inclined to think that the index operator, Reporters Without Borders, clearly has a dog in this fight, so maybe it’s liable to play up an existential threat to itself. Fair. Here’s UNESCO’s 2021/2022 report, which found that 85% of the world’s population had experienced a decline in press freedom since—you guessed it—2016.

And thanks to Google’s latest innovations, it’s harder to access the fruits of the free press that still exists. Earlier this month, the Wall Street Journal published a story, detailing how news sites are being crushed under the weight of the search engine’s new AI tools.

Nicholas Thompson, chief executive of The Atlantic, was quoted as saying: “Google is shifting from being a search engine to an answer engine.”

It doesn’t even seem to matter that, so much of the time, the answer is wrong, or at least, not right.

This is relevant because, like everything else, journalism went digital. But if we no longer trust the digital realm, that means even the sources we regard as safe and trustworthy—and which likely are important data sources for your LLM, if not socials—are in peril.

Since I’m conscious I’ll be accused of navel-gazing (truly, it is endemic to the profession), let me move away from the press. Regardless of where you fall along the political expanse, there is evidence that governments are growing increasingly hostile to certain ideas, whether or not the ideas are true (vaccines, anyone?).

The most glaring example I can think of in recent weeks is in the US, where President Donald Trump signed an executive order, directing the removal of “improper ideology” from the Smithsonian’s exhibitions. Then the museum bent the knee. The Smithsonian National Museum of American History removed a placard related to Trump from its impeachment exhibit, though mentions of his two impeachments were restored after public backlash.

Perhaps worse, support for censorship is rising. Book bans in American public schools went up by 200% last year, and many support the removal of offensive or harmful material from social media, but what constitutes “offensive” and “harmful” is not universal. They might also support the removal of disinformation and propaganda, perpetuated by automated bots and human grifters alike. But what, exactly, is disinformation and propaganda? In the paradox of tolerance, it depends who you ask.

Censorship is so prevalent on TikTok—which has north of 1.5 billion daily active users—that it’s led to a new quasi-language, where existing words are substituted with others to outsmart the oppressive algorithm, and it’s set down roots far beyond the border of TikTok’s app. Which means, I, who has never used TikTok a day in my life, must now be frequently assaulted by the phrase “unalive” because ByteDance doesn’t like “kill.”

Your LLM is the culmination of the Information Age, which has been characterized by the rapid development and democratized use of information technology. But the lights have gone out in Camelot.

Maybe it’s always been like this, but it seems like everywhere you go, things are defined by the fact that no one can agree on their definition. Is TikTok a Chinese propaganda machine or is it a fun way to access and share diverse stories? Is the New York Times a liberal echo chamber of fake news, or is it one of the last men standing in a free press? Is political correctness a hollow virtue signal, modern-day Newspeak, or an attempt to preserve inclusion through language if we can’t get there through policy?

You’re asking, How did we get here? How is this related to my LLM? You’ve heard the doomsday chatter that superintelligent robots will get us back one day. And you’ve heard the rational counter, which is that AI is merely a lesser version of ourselves, made by humans, containing our same faults and subject to the same limitations. It can’t beat us at our own game. (Nevermind that the web was also created by humans but has nonetheless grown into a problem that we cannot fix.) But whether the logicians or the doomsday preppers are right, I think we still lose. If we are such a mess, then it follows that so must be our technology, which does not act without our action.

Your LLM is the culmination of the Information Age, which has been characterized by the rapid development and democratized use of information technology. But the lights have gone out in Camelot.

Finance, like so many others, positioned itself not as a niche vertical but as an extension of Big Tech that just happens to cater to a specific subset (financial) of technology and data users. It has adopted, or tried to adopt, the “move fast and break things” ethos. Well, this is what’s happening in the world of technology, what’s happening to the Internet of Things. Finance wanted to be part of it, so I think it should be part of it. I don’t think it will do itself any favors to find some sand and make like an ostrich.

You might be familiar with the newly minted term, “enshittification.” It has a Wikipedia page and a Merriam-Webster entry. You may instead know it as “platform decay.” Whatever you call it, it describes the ongoing degradation of quality in online services and platforms.

It’s in the feeling you have when you call a customer service line and have to repeat the words “human” or “representative” over and over. It’s in the fast fashion items that fall apart after three wears. The misspelled Amazon listings. The LinkedIn influencers. Airbnbs that cost more than hotels. The dating apps you delete not because you found someone but because there’s no amount of time, effort, or money that makes them work. The fact that Microsoft Outlook’s search function will never find that email you’re looking for.

These are not necessarily AI’s fault, but they are its problem. The internet made the world better; it seems like it’s now making it worse. If we want AI to be the remedy we say it can be—able to detect cancer before it grows, potentially even cure it, do our laundry and our dishes, solve metaphysics’ mysteries—then we should stop letting it make us, and by extension our technology and our society, so sick.

Can you answer my questions? Drop me a line at rebecca.natale@infopro-digital.com