Total information saturation. Saturation is the condition, the filter is the power

I spent half my life believing the problem was finding. That it was enough to search well, read more, get there first. And suddenly the problem turned itself inside out without warning: now there's too much of everything, and the only thing scarce is someone to choose, for me, what deserves my time. Any topic drags along a thousand analyses, any question ten thousand answers, and the social intelligence that once served to generate content has moved wholesale into the trade of discarding it. Those who exercise that discarding today —the search ranking, the feed recommendation, the autocomplete, the inbox priority, the assistant that answers— are artificial intelligence systems trained on criteria someone decided and almost no one audits. Foucault would have read it at a glance: whoever governs the selection governs what gets to be sayable.

The day everything became too much

It's worth fixing an order of magnitude first, because without it the word "saturation" sounds like a café complaint.

IDC has been measuring for years what it calls the datasphere (the total volume of digital data generated, captured or replicated in the world). In its report Data Age 2025, sponsored by Seagate and published in 2018, it put that mass at 33 zettabytes in 2018 and projected some 175 for 2025. Other, more recent estimates push it towards the region of 180. The curve grows at a pace the report put at around 61% compound annual; the human capacity to read, meanwhile, is still that of a primate with a twenty-four-hour day who wants to sleep through part of it.

That asymmetry between what's produced and what can be consumed wasn't invented by the internet. It already existed with the printed book, and before that with the manuscript. What's gone awry is its nature. When a decent library held a thousand volumes, a tenacious reader could aspire to cover a fraction that meant something, form a picture, sustain a conversation. To YouTube today are uploaded on the order of five hundred hours of video per minute, according to the platform itself. That's seven hundred and twenty thousand hours a day, eighty-two years of footage every twenty-four hours. Nobody covers that. It gets sifted. And sifting, at that scale, is no longer done by a person.

The intelligence moved house without our noticing

What's striking isn't the excess. The excess is background noise, it's been with us for decades. What's worth looking at is where the intelligence that once put order into the little there was has gone to live.

Take a national newspaper editor in 1960. They received a manageable flow of wire copy and reports, and from that flow they chose by hand the scant thirty news items that would fill the next day's front page. Their job was to select, and they signed it with their surname. If they erred, or if they kept something quiet on purpose, the first outraged person who crossed them at the bar across the street could call them out. In 2026 that same function is performed by an algorithmic curator that sifts millions of items per hour and composes a different feed on every screen on the planet. The word naming the task hasn't changed. The rest —the scale, the speed and, above all, the opacity— looks nothing alike.

The same displacement repeats at every point where there used to be a human between the question and the answer. The researcher who in 1970 walked to the card catalogue today interrogates a search engine, and between their doubt and the sources there's no longer a cataloguer with a name but an ordering function nobody explains to them. Whoever needed a datum and called an expert got a fan of leads, nuances, "it depends"; now they ask a conversational assistant that hands back a closed, finished text with the air of a single answer. The gesture of filtering hasn't evaporated in any of these cases. It changed hands. And the new hands don't pick up the phone.

The layers we cross without seeing them

Let's reconstruct any digital day, because the sensation of free access rests precisely on not counting the layers that interpose themselves.

You start by searching for something and you see the first ten results from Google, Bing or Perplexity. Page three, for practical purposes, doesn't exist, and what the ranking buries isn't merely seen little: it stops existing operationally, which is another thing. Then you open a social network, and what appears in the scroll of X, Instagram, TikTok or LinkedIn was chosen from among millions of candidates by systems tuned to a single obsession, keeping you one more minute. You go into your email, where Gmail or Outlook already split your messages into important, promotion and noise before you read a single subject line. You write a reply and autocomplete suggests how to finish the sentence, conditioning along the way the very phrasing of what you meant to say. You ask for music or a film, and Spotify, Netflix or Amazon propose; what they don't propose, you rarely think to search for. And when you finally ask an assistant something, you get an answer and not a menu, with the filter already applied at the factory, without seeing anything discarded.

These layers stack one on top of another. In an ordinary interaction, a person crosses two or three before bumping into what they naively call "content". The fantasy of the one-on-one with what there is, of direct access to information, is at this point a sentimental relic best retired.

Foucault, forty years early

Michel Foucault formulated in L'ordre du discours (Gallimard, 1971) an idea that back then sounded like seminar metaphysics and today reads as an instruction manual. Discourse, he was saying, isn't so much what is said as the set of procedures of selection, exclusion and rarefaction that determine what can be said and what can't. There are things we keep quiet because they're forbidden, others because the context doesn't admit them as sayable, others because the prevailing order doesn't even recognise them as true.

Carry that over to the filters and the metaphor becomes a literal description, without effort. What the feed doesn't show isn't said. What the search doesn't return isn't cited. What the assistant omits from its answer isn't considered, not even to refute it. The algorithms execute at an industrial scale the same operation of selection and exclusion Foucault attributed to the order of discourse, stripped of the human slowness that once made it disputable and, therefore, attackable.

The question he left hanging —who rules that order— admits in 2026 a more concrete answer than he could allow himself in 1971. The product teams of the four or six companies that dominate the digital mediation layer rule it. You don't need to imagine a political design behind it; it's enough that their technical and commercial decisions have an effect on the sayable. And they do, whether they seek it or not.

The human fingerprint beneath the weights

A point that shouldn't get diluted dilutes easily: the filter has an author before it has weights.

A system learns to order results because it optimises an objective function an engineer wrote one afternoon. A feed prioritises retention because in some meeting it was decided retention was the metric that paid the salaries. An assistant answers as it answers because a team chose its data, applied its reinforcement tuning and drew its red lines. The supposed "technical neutrality" of the algorithm is a comfortable alibi for whoever builds it, who thus presents as a finding of mathematics what was a decision of theirs.

Frank Pasquale dismantled it in The Black Box Society (Harvard University Press, 2015): the opacity of the algorithm lets its authors hide perfectly human choices behind the façade of some sums that, they say, hold no opinion. More than ten years have passed and the argument, far from ageing, has become more accurate as the boxes closed up entirely.

Producing stopped guaranteeing being read

Here surfaces an asymmetry worth naming without make-up. Whoever produces —the journalist, the academic, the creator, the neighbour who posts a thread— exercises their capacity to produce, and that's as far as their power goes. It's the filter that decides whether anyone reads them. Two functions that for centuries travelled glued together have come apart with unprecedented clarity.

Before, the chain had links you could point at, and each answered for its own: you produced, you published in an outlet with a masthead, you distributed through a channel with a known owner, you reached a reader. With the algorithmic filter the chain contracts until it almost disappears: produce, algorithm, reader. The algorithm swallows in one bite the publication, the distribution and the curation, and it swallows them without accounting for any of them. Whoever publishes works blind, not knowing what slice of the public will receive them. Whoever reads works just as blind, not knowing what was hidden from them so they'd receive exactly that and nothing else.

The bubble and its small print

There's an honest nuance worth incorporating, even if it spoils the easy sermon about bubbles.

Eytan Bakshy, Solomon Messing and Lada Adamic, in "Exposure to ideologically diverse news and opinion on Facebook" (Science 348, 2015), measured over ten million users and found that exposure to ideologically diverse information does shrink, but that the factor weighing most in that shrinkage was the user's own choice —who they follow, what they click— above Facebook's algorithm considered on its own. The filter cut the diversity; the user's will, according to them, cut more.

That nuance is true and doesn't absolve the filter. Filter and will don't compete in opposing camps, they feed each other in a tight loop. The system proposes what the user already chose, the user chooses within what's proposed, the system learns from that choice and tunes the next round. Apportioning percentages of blame between one and the other is a legitimate debate for academia, but it doesn't graze the political consequence, which is dry and stated without circumlocution: citizens' exposure to diverse information is mediated by systems whose weights nobody audits in public.

The sayable as contested territory

The field of what can be said and, above all, be heard, was always contested. It was shared out by traditions, churches, parties, newspapers, whole schools of thought, and all of them carried a useful discomfort: they could be answered back.

The novelty of this century is that onto that old division has clambered an unprecedented player, the filter, which exercises a selection of substantive effect on the result and that, unlike all who preceded it, doesn't debate, doesn't publish a creed to rebut, doesn't let itself be challenged at the bar of any pub. Where there used to be a censor with a face, now there's a parameter.

What comes next is my opinion, though it can be documented. To discuss freedom of expression without discussing freedom of filter is to discuss it half-way. Freedom of expression, written in so many constitutions, shields the right to emit; it doesn't shield the right to be received, which is a different kettle of fish. If the filter decides what's received and the filter is subject neither to that protection nor to a comparable audit, the formal guarantee remains there, intact on paper, while it empties out inside. Nobody breaks the law. The law has simply stopped covering the spot where the match is really played.

The regulation arrives late and looks askance

Something is moving, slowly and trailing behind.

The European Digital Services Act —Regulation (EU) 2022/2065— imposes on the big platforms obligations of transparency about the main parameters of their recommender systems, systemic risk assessment and, in its article 38, the obligation to offer at least one version of the recommender not based on profiling. It applies gradually from 2023 and its real effectiveness is still to be proven. Regulation (EU) 2024/1689, passed on 13 June 2024 and known as the AI Act, raises a framework for artificial intelligence systems with a staggered timetable that reaches its various provisions in the following years. Beneath both texts circulate proposals for independent algorithmic auditing that no country has yet turned into a general obligation, and the idea of a right to choose an alternative filter —a chronological feed, a transparent one— that some platform already offers as an option buried in the settings, almost none as an enforceable duty.

What still doesn't exist is oversight of the criteria that train the filters comparable to what weighs on commercial practices or on the traditional media. The layer that decides what reaches each screen operates with considerably less external scrutiny than any of the mediation layers that preceded it. Of the monastery copying by hand we knew, at least, what it decided not to copy.

Definitions

Information saturation. The state of the digital environment in which the supply of content comfortably overflows the human capacity to consume, so that accessing anything useful obliges filtering.

Algorithmic filter. A technical system that selects, orders or prioritises information before presenting it to the user, according to criteria defined by its designers and learned from the data.

Order of discourse. In Foucault, the set of procedures of selection, exclusion and rarefaction that determine what can be said and be heard in a given context.

Filter bubble. The informational isolation produced by algorithmic personalisation, which reduces exposure to diverse information.

Datasphere. A term used by IDC to designate the total volume of digital data generated, captured or replicated in the world.

Algorithmic gatekeeping. The function of control over what information reaches the receiver, exercised by algorithmic systems, the digital parallel of the old human editorial control.

RLHF (reinforcement learning from human feedback). A training technique in which a model's responses are tuned according to preferences marked by people.

References

Michel Foucault, L'ordre du discours (Gallimard, 1971). Origin of the notion of the order of discourse as a system of selection and exclusion, the central framework of the article.

Frank Pasquale, The Black Box Society (Harvard University Press, 2015). Argues that the technical opacity of the algorithm conceals human decisions behind an appearance of neutrality.

Eytan Bakshy, Solomon Messing and Lada Adamic, "Exposure to ideologically diverse news and opinion on Facebook", Science 348(6239), 2015, pp. 1130-1132. Qualifies the weight of the algorithm versus the user's choice in the reduction of diverse exposure, over a sample of some ten million users.

Eli Pariser, The Filter Bubble. What the Internet Is Hiding from You (Penguin Press, 2011). The coining and spread of the filter-bubble concept.

Regulation (EU) 2022/2065 (Digital Services Act). Obligations of transparency about recommenders, systemic-risk mitigation and the option of a recommender without profiling (art. 38) for the big platforms; cited in the section on regulation.

Regulation (EU) 2024/1689 (AI Act), passed on 13 June 2024. The European framework for artificial intelligence systems, with staggered application; cited in the same section.

IDC, Data Age 2025 — Worldwide Global DataSphere Forecast (report sponsored by Seagate, 2018 edition and later updates). Source of the data-volume figures: 33 zettabytes in 2018 and a projection of some 175-180 for 2025.

You may also be interested in

• AI as a layer over reality. What did you see unfiltered today?
• The AI that orders
• Chronic cognitive saturation
• Mental overstimulation

Elsewhere

• Automating Society — Report on automated decision-making in Europe (AlgorithmWatch)
• Ethics of Artificial Intelligence and Robotics — section on manipulation (Stanford Encyclopedia of Philosophy)