The collapse of the AI model: How the abuse of LLMs contributes to the Dead Internet Theory.

The quality of online content is plummeting and spreading rapidly. Furthermore, artificial intelligence models are trained on this poor-quality input, resulting in even worse outcomes that contribute to the ever-increasing volume of low-quality content… but there is good news (I promise).

Believe me, I understand: it's very tempting to tell ChatGPT to write your next report, your next financial analysis, or even ask them to help you write an opinion essay, an expert article, or even the entire deliverable of your consulting project.

In the following lines I will explain why:

Large language models (LLMs) have amazed us and even generated a false sense of credibility through their use of statistical probability to predict error-free grammatical constructions. And the speed with which they provide answers to our questions, often quite accurate ones, has led many users to confuse "artificial intelligence" with "truth."

We must also accept that many of us have fallen into the trap of surrendering our creativity and authorship responsibility, based on ease and speed.

There's no reason to be ashamed; all these AI qualities are attainable, and it would be arrogant to think that humans have the mechanical and mental capacity to process and produce texts in response to a prompt with the eloquence and speed of an LLM agent that consumes an average of 0.24 watt-hours and 0.26 liters of water to agree with you.

Dumbification of the AI models.

The problem is that as long as we continue down this path, there's a high risk that, collectively, we'll be charting a very unpromising course for artificial intelligence models themselves . In other words, the indiscriminate use of these tools to generate content will lead to a spiral of their dumbing down , a technical process that many experts call AI model collapse . And why should you care? Because if you understand this premise and recognize the likelihood of heading in this direction, many possibilities open up for you, your brand, and your project.

Let me explain, let's go step by step:

The generosity of the term “ artificial intelligence ”

“Artificial intelligence told me so…” is a phrase I’ve been hearing more and more over the past few months. It reflects how, just like when your grandpa swore something was true because “they sent it to me on WhatsApp,” more and more users are blindly trusting the responses they get from Gemini, ChatGPT, Claude, and (holy shit) even Grok! There’s a tremendous humility in trusting the medium because of our own limitations.

As a student, I had the privilege of being trained in critical thinking. Teachers whom I hold in the highest esteem taught me how to evaluate, validate, and question sources of information. Later, I learned about echo chambers and confirmation bias… and finally, out of simple professional curiosity, and despite not being an expert in computer science or computer systems, I set out to understand HOW an LLM model is trained and what processes it uses to generate an eloquent response .

“Intelligence”, as defined by the RAE (Royal Spanish Academy), is the ability to understand, comprehend, and solve problems by adapting effectively to the environment through cognitive skills such as Learning, memory, and reasoning . And it is under this definition that I want to challenge our understanding and confidence in the outputs of artificial intelligence, acknowledging that human intelligence also has multiple flaws, but seeking to make us understand the fundamental difference.

While LLMs do utilize what we might call learning and can DEFINITELY boast greater memory capabilities than humans, to date their inability to reason and the non-clonable human quality of creativity , This makes them far more prone to error than expert opinions . LLMs have the capacity to generate outputs, but they haven't progressed in terms of being able to question their own validity or certainty. They assume that what they've generated is correct … and if you question them, they immediately abandon that assumption and offer a different answer. I don't want to sound like a broken record, so I'll just say to my article in which I talk about the content diminished by error and hallucination to which we are exposed today with LLMs.

Credit: Level Up Coding.

LLMs are trained to process a user prompt and, based on a series of carefully designed instructions and steps, utilize immense amounts of data they have received from various digital channels. Using mathematical models and statistical probability , they aim to generate a coherently constructed response and resolution. To arrive at their answer, they use an impressive volume of content, breaking it down, structuring it, and presenting it. And often they are right… but have you ever challenged the answer they give you? Have you ever told ChatGPT, “No, your answer is wrong, and you are being extremely conservative in your calculation”? As I mentioned in the previous paragraph, the behavior this will trigger is that the LLM will immediately recalibrate its response, admitting it was wrong and generating new values. Because it lacks the capacity to reason and trust its own reasoning.

My point in bringing up the limitations of LLMs again this time is to make sure you understand why, when someone says "Artificial intelligence told me so," they are over-relying on predictive models rather than truth-generating ones.

Next point:

Dead internet or dumber internet? The collapse of the AI model

We cannot assign full blame for the concept of the “dead internet” to LLMs exclusively… but we can blame their accelerated growth rate.

If you're not familiar with the concept of the "dead internet ," it's a term that began to gain traction in the mid-2010s and generally refers to three key concepts:

1. The proliferation of bots makes it seem as though most interactions on digital channels and traffic to websites are not generated by real humans.

2. Algorithmic and AI-generated content, primarily through images and video, deepfakes, and automated articles.

3. The erosion of human connections due to a lack of authenticity in digital media.

Follower farms existed long before LLMs were introduced to society, and that's why I say they're not the only perpetrators in the murder of the internet. BUT below I'll tell you why I insist that AI is indeed responsible for accelerating the pace of its impending demise.

“Thought leadership” articles that claim expertise—but are actually generated by AI.

Evidence #1: Opinion and expertise articles requested from LLMs

Every time someone decides to ask their preferred artificial intelligence agent to write their report, blog post, essay, or similar document, the volume of generic content across the internet grows. Consequently, the proportion of authentic, differentiated, valuable content that offers unique perspectives decreases. According to a A recent study by data scientists from Graphite, just 12 months after ChatGPT launched, found that articles generated by artificial intelligence already represented 39% of articles published online, and by the beginning of last year, they had surpassed 50%... and the upward trend has continued.

Wix recently integrated an AI agent into its web hosting services that, without any specific prompts and based solely on the industry in which your site operates, can automate the publication of relevant articles on your blog. In other words, it now takes just a couple of clicks for someone who succumbs to temptation to automatically become partly responsible for contributing generic, low-value content to the vastness of the internet.

Evidence #2: Parallelism of self-translated content

In January 2024, a A study published at Cornell University revealed the long-term dangers of paralleling automatically translated content with the use of AI .

I highly recommend reading the full study, but here I'll share the main findings and what they mean sequentially for my premise of intelligence loss:

1. In order to reach wider audiences, and thanks to the convenience of these tools, we are currently experiencing exponential growth in digital content being automatically replicated into different languages . From entire websites to even video reels on YouTube and Instagram, we are asking automated agents to create clones of our original content in various languages.

2. These translations are useful, BUT they are definitely far from perfect . They are riddled with contextual errors, literal translations, and sometimes, due to incorrect input encoding, even generate nonsensical constructions in other languages . The study reveals that the use of auto-generated translations for the sake of multiparallelism, despite advancements, produces erroneous content with elements of "hallucination" and poor grammatical constructions. The next point is key: for the sake of convenience, we have decided to accept these imperfections.

3. The result is that instead of having a single point of reference for this content, we now have the original and its flawed clones , which add up as additional points of reference for the same piece. In other words, the same message is repeated multiple times in different languages, but now with lower quality.

4. Data QUALITY is crucial for effectively training an LLM model . When the first version of ChatGPT was launched, it was trained on a HUGE but finite volume of data (remember that it couldn't refer to current information because its databases only went up to a certain cutoff date?). However, as competitors and new versions emerged, the models have been fed and trained in real time on ALL available content, indiscriminately.

A spiral of low-value content that ultimately leads to the collapse of the AI model.

These pieces of evidence, combined, lead me to the following conclusion:

If there is an increasing volume of low-value content due to auto-generated content, and this volume is automatically replicated indiscriminately into different languages, resulting in imperfect copies, LLMs end up with MORE poor-quality content feeding their training processes and reference databases. This generates the phenomenon known as "AI model collapse," resulting in a worrying spiral of lower-value content... and the cycle repeats itself endlessly.

Now for the good news for you , your brand, and your voice: You can use it to your advantage.

Companies at the forefront of artificial intelligence expansion know EVERYTHING I just mentioned and are internally taking steps to, ironically, become their own worst enemy. Let me explain with a couple of examples:

• While Google invests resources in encouraging you to use its Gemini AI more and more, the results it returns after a prompt favor content NOT generated by artificial intelligence , as I explain in This article in which I talk about honesty in the age of artificial intelligence or This other one explains why you shouldn't use ChatGPT to generate your blog content.

• In an effort to capture that authentic, human content, a recent study by SEMRUSH revealed that the two domains most cited by ChatGPT, Google AI, and Perplexity were Reddit and LinkedIn, characterized as spaces where much more content is generated by humans.

The E-E-A-T model for evaluating content quality.

While the EEAT content rating model remains valid (and it should be because it responds to a logic of better quality expected by the end consumer), the proliferation of "dumber internet" opens up enormous doors of possibilities for those who have the real (not artificial) intelligence to identify and even catalyze them .

If you ensure your content marketing strategy avoids the trap of self-generated content and instead prioritizes valuable, authentic content, and you also use artificial intelligence intelligently, your outputs will be registered by the same tools as more appealing to the end user. Not only that, but your messages themselves will be of greater value to your audiences and potential customers. In other words, you'll have better content that converts more and ranks higher in search engines and other traffic sources.

This, especially when handled by experts, will also mean you have better data to fuel your inbound marketing efforts, such as a Google Ads Search campaign or a continuous SEO and GEO optimization process for your website.

Additionally, as the proportion of authentic, real, and human content continues to decline, scarcity will further enhance the value of your strategy that prioritizes the usefulness and value of your output. And I'm not just saying this as an unproven hypothesis.

At Werko Marketing Solutions, we've seen the rollout of this strategy firsthand. We use artificial intelligence in MANY of our services and deliverables:

• We leverage the benefits of Machine Learning to optimize performance and content scheduling on social media;

• We have SEO audit engines that utilize LLMs to build reports of real metrics presented in a very attractive way, and of course we have used Google in AI mode or LLMs as search engines…

• We have also used Adobe AI on more than one occasion to perform photo correction or retouching.

• And the AI tools in our email marketing platform are largely responsible for our high open rates, which far exceed the market average.

BUT artificial intelligence is not the author of 1% of our content in articles, social media copy, or website content .

Our website receives daily visits from AI bots that reference our content because it meets high E-E-A-T standards and is not generated by LLMs.

Furthermore, we've ensured that our website and backend design are geared towards helping LLMs like ChatGPT, Gemini, Claude, Perplexity, and others visit our site and use our content as a source for their responses to user prompts. And this strategy is yielding impressive results . Month after month, we're seeing (because we constantly monitor and track it) double-digit percentage increases in traffic from AI agents to the website, BUT also in human traffic. Our rankings with the various LLMs are rising, and the metrics from Google Search Console show that we're increasingly appearing as a landing page for real searches by real people seeking real services.

Similarly, this strategy is helping us attract new potential clients. Last month, two different accounts told me they came to Werko Marketing Solutions by "asking ChatGPT," and while I see our competitors (because we're constantly keeping an eye on them too) succumbing to the temptation of filling their channels with AI slop, I celebrate our promising future ... and I'd love to celebrate yours. We can discuss it sometime.