We have never produced so much text and, at the same time, we may be entering an age in which writing and reading are becoming increasingly unnecessary activities.
The contradiction is only apparent. Every day, we entrust artificial intelligence with a growing number of small linguistic acts: we ask it to reply to an email, make a thought sound more elegant, turn a few notes into a document, translate a conversation, prepare a report, revise a letter. At the other end of the same process, something perfectly symmetrical is taking place: we ask artificial intelligence to summarize the emails we receive, extract the essential points from a document, tell us about an article we do not have time to read, turn fifty pages into five minutes of audio.
The algorithmic machine is entering both ends of communication at the same time.
This is the condition I propose to call the AI Mirror Loop: a circuit in which a human being entrusts an artificial intelligence with formulating what they wish to communicate, while at the other end another human being entrusts another artificial intelligence with reading, interpreting, and summarizing that same message.
The pattern is very simple:
human → AI → text → AI → human
The text continues to exist, but it takes on a curiously new function. It is no longer necessarily something written by one human being for another human being to read. It can become a kind of intermediate territory, produced by one machine mainly so that another machine can pass through it—a sort of interchange format.
Imagine an entirely ordinary situation. I need to explain that an exhibition scheduled for Saturday has to be postponed because one of the works will not be ready. I could simply write that, but instead I ask a generative system to turn the thought into a professional, elegant, sufficiently diplomatic message that will not upset anyone. One sentence becomes four hundred words. The recipient receives the message, sees that it is long, and asks their assistant to summarize it. A few seconds later they read: Saturday's exhibition has been postponed because the work will not be ready.
We have arrived exactly where we started.
In between, we have produced hundreds of words that nobody really needed to write and nobody really intended to read.
It is difficult not to see something comic in this gigantic machine for expanding and subsequently compressing language, a process not unlike the compression and decompression of software transferred over the internet—the familiar zipped file. One artificial intelligence takes twenty words and turns them into four hundred; a second takes those four hundred words and reduces them back to twenty. If this behaviour became systematic, we might witness an apparently paradoxical phenomenon: the greatest textual inflation in history would coincide with the gradual disappearance of the two activities that historically justified the existence of texts.
The paradox becomes even clearer if we move only a few years forward into a hypothetical, and quite plausible, future.
In the morning we find one hundred and fifty messages in our inbox, but we have no reason to read them. Our assistant has already examined them during the night and informs us that only four require a decision. We can accept the first; the second proposal is too expensive; the third does not interest us; for the fourth, we can suggest Thursday afternoon.
We write nothing. We give four instructions, perhaps simply by speaking.
Our assistant produces four perfectly worded messages and sends them to four different people. But those people may not read them either. Their assistants analyse the messages and report back simply: Claudio accepts; Claudio thinks the price is too high; Claudio is not interested; Claudio suggests Thursday.
Thousands of words may have been generated between the two sides.
Not a single human being has read them.
At this point an obvious question arises: why continue producing them at all?
If the sender's artificial intelligence already knows the original intention, while the recipient's artificial intelligence merely has to reconstruct that same intention, the text becomes a redundant intermediary. The two systems could exchange structured information directly, negotiate an appointment, compare terms, check availability, and involve their respective owners only when an actual decision is required.
The first element to disappear from the AI Mirror Loop could therefore be precisely the thing that for millennia we have regarded as indispensable to communication at a distance: the text itself.
human → AI ⇄ AI → human
The more interesting point, however, does not concern only what happens between two people. All the material produced in the meantime remains in the world. Emails are archived. Articles are published. Reviews are indexed. Press releases are copied. Descriptions of artworks end up in catalogues and on museum websites. Posts are reproduced in other posts. Texts are translated, summarized, quoted, and recombined.
AI-generated material thus begins slowly to settle and accumulate within the very culture from which artificial intelligence originally learned.
And this is where the Mirror Loop ceases to be merely a communication circuit and begins to resemble a cultural one.
Human culture provided the material with which we trained the machines. The machines produce new material that flows back into human culture. That material is read by humans, processed by other machines, and contributes, directly or indirectly, to the production of further texts, images, and interpretations.
Culture teaches the machine; the machine gives back post-culture; and that post-culture becomes the landscape in which humans and machines learn all over again.
It is at this point that a phenomenon studied in the field of artificial intelligence—model collapse—takes on a significance that goes beyond the purely technical terms in which it has so far been studied and theorized.
The term describes a specific problem that can arise when successive generations of models are increasingly fed synthetic data produced by previous generations. During this process, the original distribution can become distorted, and some of the first information to be sacrificed lies at its extremes, in the so-called tails: rare, underrepresented, improbable cases.
Transferred from statistics to culture, the image is striking.
The risk is not necessarily that a culture increasingly mediated by artificial intelligence will suddenly become stupid or mediocre. Something far subtler could happen. It could become sufficiently beautiful, sufficiently intelligent, sufficiently original, sufficiently elegant.
A vast quantity of local differences within an increasing overall similarity.
The statistical centre of culture could become immensely populated while its peripheries become progressively harder to reach. What disappears would not necessarily be what is wrong. It might be what is improbable: an absurd linguistic construction, an incomprehensible taste, an idiosyncrasy, a private word, an error stubbornly preserved, a solution no manual would recommend, an artwork that initially bears no resemblance to what we are accustomed to recognizing as art.
We might call this possible consequence of the AI Mirror Loop the collapse of exception.
A considerable part of cultural history has been produced precisely in the tails of the distribution. The avant-garde is, almost by definition, statistically marginal at the moment of its appearance. What twenty years later will be recognized, imitated, and taught often begins as a mistake, a provocation, an idiosyncrasy, or simply something that very few people find interesting.
A machine today can imitate Van Gogh, Cage, Duchamp, or Artaud. It can perfectly recognize the anomalies they introduced into culture because, once historicized, those anomalies became information. They entered the corpus.
Far more interesting is the anomaly that does not yet have a name: the one that has not yet become a style, the one we cannot ask for in a prompt because we do not yet know that it exists, the one that, viewed from the centre of the distribution, simply looks wrong.
Perhaps it is also from this much more ordinary and less spectacular perspective that we should reconsider the idea of the post-human, which has circulated through philosophy, art, and contemporary culture for decades.
When we think of the post-human, we tend to imagine rather conspicuous scenarios: augmented bodies, intelligent prostheses, biological hybridization, consciousness transferred into machines. But the post-human produced by the AI Mirror Loop could be considerably less cinematic. It may require no transformation of our bodies and no appearance of some imaginary omnipotent artificial intelligence. It may consist simply in our progressive marginalization within processes that continue, formally, to be built around us.
We remain the reason the system communicates, but we no longer necessarily formulate the communication. We remain the recipients of information we do not read. We produce books that other machines summarize, images that other machines classify, music selected by algorithms, documents that artificial agents compare with other documents.
The human being does not disappear. It remains at the ends of the process, at the margins of the loop.
And perhaps this is a more interesting form of the post-human precisely because it requires no technological apocalypse. It is not the machine theatrically replacing the human being. It is the human being who, one small act of delegation after another, withdraws from the centre of their own symbolic universe.
But together with effort, slowness, and inefficiency, we may also delegate something we had not anticipated: noise.
Because the human being is a formidable machine for producing noise. We forget, misunderstand, contradict ourselves, change our minds, develop obsessions, misinterpret instructions, misuse words, follow intuitions for no apparent reason and sometimes persist precisely when every rational analysis suggests that we should stop.
From the point of view of efficiency, these are flaws. From the point of view of culture, it is far from certain that they are.
If this hypothesis is correct, the radical gesture of the artist of the future might take a curiously simple form.
Not to produce more, but to deviate.
To preserve an error when correcting it would be effortless. To pursue an obsession when no data suggests that anyone might care. To choose an inefficient solution. To write a sentence that a model would suggest rewriting. To continue working on something that does not yet have an audience, a category, or a justification.
Ultimately, to claim the right to be statistically irrelevant.
Because if the AI Mirror Loop tends continuously to pull culture back towards what culture already knows, the place where we should look for something genuinely new may be exactly where it has always been.
Far from the centre. In the tail.
This text may have been written by an artificial intelligence... or perhaps not.
