Description
Massive, superficial AI-generated content isn't just a problem, it's also a symptom. Technology amplifies a consumption model that rewards fluidity and drains our attention span.
We listen to interviews, podcasts and audios of our family at 2x. We watch videos cut into highlights, and we base decisions and criteria on articles and reports that we have only read summarized with AI. We consume information in ultra-fast mode, but at a cognitive level we give it the same validity as when we consumed it more slowly, and we even apply it in decision-making. What is affected by this process is not the basic memory of contents, which seems to be maintained according to controlled studies, but the ability to connect that knowledge with what we already had and to elaborate our own ideas with it. More than superficiality, it is worrying that this new way of thinking is sufficient in so many contexts.
What's new and what's not?
We may think that generative AI has only intensified an old dynamic in which content production is infinite, but our attention spans are the same. We cannot fool ourselves either, because since the Internet has existed, infinity is not new. If we were to say that the problem is that there is too much content, we would be complaining about a situation in which we have been living for more than twenty years. Nor is the crisis of authority of official information or the difficulty in distinguishing reliable sources from those that are not.
However, the AI slop, which is the flood of AI-generated digital content on the Internet, brings its own logic and new considerations, such as breaking the link between effort and content, or that all that is generated is a statistical average of what already existed. This uniform and uncontrolled flow has consequences: behind the mass-generated content there may be an orchestrated intention of manipulation, an algorithmic bias, voluntary or not, that harms certain groups or slows down social advances, and also a random and unpredictable distortion of reality.
But how much of what I read is AI?
By 2025, it has been estimated that a large portion of online content incorporates synthetic text: an Ahrefs analysis of nearly one million web pages published in the first half of the year found that 74.2% of new pages contained signals of AI-generated content. Graphite research from the same year cites that, during the first year of ChatGPT alone, 39% of all online content was already generated with AI. Since November 2024, that figure has remained stable at around 52%, meaning that since then AI content outnumbers human content.
However, there are two questions we should ask ourselves when we come across estimates of this type:
1. Is there a reliable mechanism to distinguish a written text from a generated text? If the answer is no, no matter how striking and coherent the conclusions are, we cannot give them value, because they could be true or not. It is a valuable quantitative data, but one that does not yet exist.
With the information we currently have, we can say that "AI-generated text" detectors fail as often as a random model would, so we cannot attribute reliability to them. In a recent study cited by The Guardian, detectors were correct about whether the text was generated with AI or not in less than 40% of cases. On the other hand, in the first paragraph of Don Quixote, certain detectors have also returned an 86% probability that the text was created by AI.
2. What does it mean that a text is generated with AI? On the other hand, the process is not always completely automatic (what we call copying and pasting) but there are many grays in the scale: AI inspires, organizes, assists, rewrites or expands ideas, and denying, delegitimizing or penalizing this writing would be ignoring an installed reality.
The two nuances above do not cancel out the fact that the AI slop exists, but this does not have to be an inevitable fate. There are ways to mitigate its effects on our abilities.
What are the antidotes?
We may not contribute to the production of synthetic content, but we cannot slow down what is happening, so the challenge is to review the criteria and habits of mind with which we approach both reading and writing content.
1. Prioritize what clicks: one of the few reliable signals we have left is that clicking sensation at the moment when something connects with a previous knowledge, an intuition that we had diffused or an experience of our own, and reorganizes it or makes it clear. We also often say that it "resonates". If something clicks, it's worth following, confirming, researching, and briefly elaborating on a personal level.
2. Look for friction with data: anchoring content in open data and verifiable sources introduces healthy friction against the AI slop. It reduces, above all, arbitrariness and the feeling of interchangeable content, because the data force us to interpret and put it in context. It is a way of putting stones in the excessively fluid river that is the generation of language, and it works when we read and when we write.
3. Who is responsible? The text exists easily now, the question is why it exists or what it wants to achieve, and who is ultimately responsible for that goal. It seeks the signature of people or organizations, not so much for authorship but for responsibility. He is wary of collective signatures, also in translations and adaptations.
4. Change the focus of merit: evaluate your inertia when reading, because perhaps one day you learned to give merit to texts that sounded convincing, used certain structures or went up to a specific register. It shifts value to non-generatable elements such as finding a good story, knowing how to formulate a vague idea or daring to give a point of view in a controversial context.
On the other side of the coin, it is also a fact that content created with AI enters with an advantage in the flow, but with a disadvantage in credibility. This means that the real risk now is that AI can create high-value content, but people have lost the ability to concentrate on valuing it. To this we must add the installed prejudice that, if it is with AI, it is not valid content. Protecting our cognitive abilities and learning to differentiate between compressible and non-compressible content is therefore not a nostalgic gesture, but a skill that in the long run can improve the quality of public debate and the substrate of common knowledge.
Content created by Carmen Torrijos, expert in AI applied to language and communication. The content and views expressed in this publication are the sole responsibility of the author.
