The emptiness that speaks dense

Posted on: 10 May 2026

You open the article because the headline promised something, you read the first two hundred words, you realise there is nothing there, you close it. Thirty seconds of your life that will not return, and a small additional dose of generalised distrust towards the system. Not towards that author, towards the system, because that small gesture, repeated for years by anyone with competence in any field, has silently eroded confidence in digital writing to the point that today, when you open any page, you start from the presumption of emptiness and look for counter-evidence, rather than from the presumption of substance and look for confirmation.

The emptiness is legible only to those who know, and that is precisely the reason why those who produce it do not see it and continue to produce it. The thought leader publishing the thousandth "5 strategies to optimise your leadership" thinks they are writing for someone who needs to read those five strategies, when in reality they are writing for someone who has already read them in five hundred other variants, because the audience that reads content on LinkedIn is, statistically, the same audience that produces it. People write empty things to the same people who write empty things, in a closed loop of mutual confirmation of superficiality, where the like is the only metric and the like is free because it costs no attention, only a second of the thumb. It is a click economy in which the product is the click itself, the content is merely the pretext, and the competent reader is a negative externality who self-removes from the system without the producer noticing.

There is a technical word for this phenomenon and it is called reverse selection. By writing for fools, you guarantee yourself fools. The thought leader who produces emptiness believes they are politely lowering the entry barrier, when in reality they are declaring to the competent reader: "you are not my client, please leave." And the competent reader thanks them, leaves, and does not return. Then the thought leader complains that their leads do not convert, that their audience is "engaged but unqualified", that their funnel is broken. The funnel is not broken. It is functioning exactly as designed. It is filtering perfectly for those who cannot tell signal from noise, because that is the audience the content selects. Those who write empty attract those who consume empty. This is social network mathematics, not sociology.

LinkedIn deserves a paragraph of its own because it has algorithmised this mechanism to the level of mandatory grammar. Open any feed, read the first twenty posts, and you will see the same structure every time: three lines of hook with a rhetorical question or a false paradox, ten to fifteen lines of self-absolving personal confession with a fake-vulnerable detail, twelve to eighteen lines of universal lesson extracted from that confession as if it were a law of nature, two lines of call to action disguised as a question to the audience. That is the format. Those who do not follow it are punished by the algorithm, those who follow it are rewarded by the algorithm, and so everyone follows it, and LinkedIn has become a single-script theatre in which thousands of people replicate the same screenplay with minor variations of costume. The result is that LinkedIn today is fishing ground for fools, not in a pejorative sense towards the audience but in a structural sense towards the system. The platform selects producers and consumers who mirror each other in the same performative logic, and the intellectual reader either leaves or, if they stay, they do so with flippers, the way one snorkels in murky water, knowing that ninety percent of what passes before their eyes is plankton.

So far, none of this is new and anyone with functioning neurons could arrive at it; this is the baseline of the diagnosis. The more uncomfortable point, and the more technically recent one, comes next, when one considers that all this empty content, produced at industrial pace over the last fifteen years, did not disappear when it was read and forgotten. It remained on the web, indexed, archived, and scraped, and then went on to become part of the training corpus of the generative artificial intelligence systems that today billions of people use to obtain answers. When someone asks ChatGPT, Gemini, Claude or any of their cousins how to handle a difficult negotiation, part of what they receive comes from Harvard Business Review, part from academic papers, part from Reddit threads with competent users, and part, regrettably not a small part, from the five hundred thousand LinkedIn articles titled "5 strategies to optimise your leadership". The statistical model does not distinguish quality; it distinguishes frequency. And the frequency of emptiness, on the web of 2010-2024, is extremely high. The British and American press has by now given the resulting AI output a name, "AI slop", and the term has stuck because it captures something true: the fluent, plausible, uniformly mediocre prose that emerges when a system trained on noise is asked to produce signal.

Ilia Shumailov and his co-authors published a paper in Nature in 2024 that gave a technical name to what happens when LLMs are trained on corpora contaminated by their own output: model collapse. The mechanism is elegant and devastating in equal measure. Models, when trained recursively on content generated by other models, progressively lose the capacity to represent the statistical tails of the original distribution. In less technical language: they lose the signal of quality, the rare voices, the anomalous cases, the dense reasoning, and they flatten towards the statistical centre, that is, the average of the corpus, that is, the noise. The degradation is not cyclical but progressive, and each iteration amplifies the distortions of the previous one.

The implication is uncomfortable. The thought leader who complains that ChatGPT "produces superficial answers" is not criticising ChatGPT. They are looking at themselves in the mirror, because the superficial content that the model returns is precisely the superficial content they and thousands of their colleagues have produced for fifteen years, statistically ingested, fluently reformulated. The system is returning to them their own image, processed. It is a form of poetic justice that is structurally impeccable, and probably the only form of justice that the content marketing of the last fifteen years deserved.

There is an honest note to make at this point, because otherwise this piece would be dishonest. When I write, I work in tandem with a methodologically constrained AI system, and I know perfectly well that without constant methodological pressure, that system would slide towards the statistical average register of the corpus on which it was trained, that is, towards noise. The pressure I exercise is Popperian, systematic, daily: pre-emptive falsification, mandatory fact-checking, refusal of the AI register, explicit demand for density, control of repetitive patterns. It is active labour, not a passive property of the system. The average reader does not know they ought to exercise this pressure, and therefore consumes outputs that are corpus-averages, that is, garbage-averages, presented with the linguistic fluency that makes it impossible to distinguish substance from appearance at first sight. This is not an AI problem; it is a problem of how AI is used. But it is a real problem, and this piece would not be honest if it did not say so.

The full picture, then, is the following. Fifteen years of empty content marketing have eroded the trust of the competent reader, have saturated the digital space to the point of burying those who produce substance, have selected producers and consumers who mirror each other in performative logic, and have contaminated the corpus on which the entire next generation of information-synthesis systems rests. The damage is permanent, irreversible, and industrial in scale. There is no technical solution because the damage has already been done and the contaminated corpora are already in circulation. There is only one way forward, and it is not a solution but a personal choice: write dense, knowing that the audience will be reduced, accepting the reduced audience as the price of quality, and letting the noise talk to itself in its own echo chamber until it eventually tires.

If this resonates, it resonates. If not, you have just spent ten minutes reading something that was not written for you. That is the point.