You are not imagining it. Something changed, and it changed fast.
You consume more information than five years ago — more news, more messages, more updates, more content of every kind. And yet you feel less oriented, not more. You finish a day of reading and deciding and you are not sure what you actually decided, or why. You have the sensation of having processed a lot and understood very little.
The differences between things are collapsing. A shop is interchangeable with the next one for a cent less. An opinion only exists if it is shouted, or endorsed by someone with a large audience. Products, services, articles, people — everything seems to occupy the same undifferentiated space. You know they are different, but you cannot feel the difference. The signal is there, but it does not cross.
None of this is weakness. None of it is a failure of attention or intelligence. It is what happens to any system when the input exceeds the throughput capacity — and the input keeps growing while the capacity stays fixed.
What follows is not an interpretation of this experience. It is a structural model of the system producing it — with numbers, a formal framework verified across 14 independent domains, and a deterministic simulator. The experience is the entry point. The argument is what comes after.
The question is: does this have a structure? Can it be measured? Or is it just a feeling?
One cause is easy to see, and almost no one is talking about it correctly.
When you use AI to save time — to write faster, research faster, draft faster, respond faster — you gain something real. You get minutes back. But what happens to those minutes? They become more content. More messages. More output, released into a shared space where other people have to read it, filter it, evaluate it, and respond to it.
You saved one hour. You cost dozens of people fractions of hours. At scale, across millions of users, the net effect is not a gain in collective time. It is an acceleration of the input that everyone else must process.
Every person who adopts AI does not only accelerate themselves. They accelerate the consumption of attention for the entire population. It is an externality on other people's time — invisible, unintended, and structural.
This is compounded by a second mechanism: the more content there is in a space, the less each individual piece can be distinguished from the others. The vocabulary of recognisably different things — arguments, styles, positions, products — has a ceiling. When production exceeds that ceiling, you are not getting more variety. You are getting repetition that looks like variety.
The result is the sensation you already know: you process more and understand less. Because more and more of what you process is structurally identical to what you have already seen, even if it looks different on the surface.
Until recently, the honest answer was: not precisely. The phenomenon was observable — everyone feels it — but there was no formal structure that connected the individual experience to a measurable system-level property.
The question needs to be reframed. The right measurement is not "how much information is being produced" — that number grows and is meaningless on its own. The right measurement is: how much distinguishable information crosses the boundary per unit of time? That is, how much of what is produced actually registers as something new, actionable, and different?
This is a harder quantity to define. But it turns out to be well-defined — and calculable — once you identify the structural conditions that govern it.
Those conditions are three. They are satisfied by the planetary information system the same way they are satisfied by a dozen other apparently unrelated systems — from prime number sequences to the periodic table, from Western music to the structure of fairy tales. When a system satisfies these three conditions, four properties follow inevitably. One of those properties is the efficiency paradox: the locally optimal strategy accelerates the consumption of shared space, rather than preserving it.
The framework that identifies and formalises these conditions is called Sub-Limit Dynamics. What follows are its findings applied to the information system we all inhabit.
The findings are not opinion. They are verifiable, falsifiable, and open to challenge. The reader is invited to contest the mapping — not the conclusions. If the mapping holds, the conclusions follow by necessity.
The key datum: screen time has been essentially flat for a decade. +31 minutes in 11 years. The total attention available to humanity has not grown. It cannot grow. But information production, driven by AI, has exploded. The gap between the two curves is what you feel every day.
If the system were absorbing more, you would expect productivity to be rising with adoption. Here is what is actually happening:
Adoption is at maximum. Results are near zero. Trust collapses as use rises. This is not a failure of the technology. It is the efficiency paradox operating at planetary scale: the strategy that looks optimal for each individual organisation is fragmenting the residual space for everyone.
There is no need to measure how much information is produced. What matters is how much distinguishable information crosses the boundary per unit of time. This quantity is called D(t).
The critical ratio is λ(t) = ε(t) / ρ(t): the ratio between sterile expansion (scrolling, noise, maintenance) and productive recovery. When λ grows, the system consumes time without crossing the boundary. The D(t) curve does not rise indefinitely. It rises, reaches a peak, then falls — because the vocabulary of distinguishable forms is exhausted well before the physical limit of time.
With any reasonable set of parameters, the peak of D(t) falls between 2022 and 2024. Not ahead of us. Behind. Production continues to rise. Distinguishability is falling.
D(t) is an average. But collapse is not uniform. S(t) measures the fraction of the system that still crosses the boundary — how many organisations, individuals, and processes still produce a real, distinguishable outcome.
The April 2026 data provide a direct anchor: 80% of organisations report no EBIT impact from AI. This means S(t) ≈ 20%. Twenty percent of the system still crosses the boundary. The rest produces, adopts, invests — but does not cross.
The coherence threshold for complex systems is approximately 30%. Below this, residual boundaries become incompatible with one another. The system loses global coordination. We are already below it.
The analytical model indicates the direction. But there is a more powerful instrument: the Ω-field simulator, a deterministic equation on a 220×220 grid with four operators governing diffusion, compression, rotation, and damping. We translated the planetary information system into its parameters.
The simulator ran for 8,400 steps. Verdict: Structural. The system converges to a state the analytical model had not identified with precision. Here is what that state looks like:
85% is volume without boundary. The vast majority of content, initiatives, products, and communications does not cross. It exists, it occupies time, it generates maintenance, but it produces no real difference. Not because it is "bad." Because the system no longer has the capacity to distinguish it.
The islands do not communicate. The high-intensity clusters — the 20% that still produce real outcomes — are concentrated, isolated, and do not transfer value to others. Each island lives its own reality. Shared reality dissolves.
Expansion self-extinguishes. No crisis or regulation is needed. Under strong selection, production shuts down on its own because there is no longer space to expand profitably. But it shuts down after consuming nearly all the remaining margin.
The efficiency paradox is already visible. Organisations that use more AI do not obtain more results. 95% of pilots produce no P&L. Trust collapses while use rises. Those who optimise most fragment the residual space fastest. This is not a bug in AI. It is a structural property of any system under irreversible constraint.
We have tested every parameter. Adoption speed, filtering capacity, decision threshold, environmental multiplier. None inverts the curve. The process is autocatalytic: even with constant parameters, the system continues to converge on the attractor.
The only variable that changes the system's fate is B: cultural resistance to subtraction.
B is not a technical parameter. It is the way 8 billion people conceive of loss. Subtracting is failure. Giving up is weakness. Growth is good. Less is bad. No CEO has ever presented a quarterly report saying "we removed 40% and now it works."
There is no education in subtraction. Adding is demonstrable, visible, measurable. Subtracting seems to do nothing. And in a world that measures results rather than processes, doing nothing is social death.
And yet it is the only operation that restores time. The only one that reopens the boundary. The only one the system structurally requires.
In the model: when B drops below 0.5 — when half the culture learns to subtract — S(t) stabilises. The islands regain boundary. The system resumes crossing. B = 0.5 does not yet exist anywhere on the planet.
Everything above rests on a formal structure. This section makes it explicit. Not to intimidate, but because the claim is falsifiable — and falsification requires knowing exactly where the argument stands.
Sub-Limit Dynamics identifies three necessary conditions. If a system satisfies them, four structural properties follow inevitably. The planetary information system satisfies all three.
The set of available actions is finite. For the information system: human time is finite. 16 waking hours, 6 billion connected people, approximately 40 billion person-hours per day. It does not grow.
Every action permanently reduces the available space. Every minute spent scrolling, reading, responding, processing is a minute that does not return. The consumption of attention-time is irreversible.
The classification of every action depends on the interaction between what has been accumulated and what remains. Content is "new" only relative to what you have already seen and the time you have left. Not in absolute terms.
AI adoption data (McKinsey, Stanford HAI, U.S. Federal Reserve, ITU) and screen time data (DataReportal, GWI) are public, replicated measurements. The +31 minutes in 11 years is an observed datum. The 80% with no EBIT impact is survey data (McKinsey Q1 2026, Writer Enterprise Survey 2026, Gartner). The trust collapse (−18%) is ManpowerGroup 2026. These are not models. They are measurements.
The SDL framework (three axioms, four properties) is a formal system, verified across 14 independent domains (C1–C9, M1, S1–S2, B1–B3). The D(t) and S(t) curves are structural consequences of the axioms applied to the information domain, calibrated on the data. Not measurements.
The claim that the planetary information system is a T3/R2-type system, that the peak of D(t) has already passed, and that the attractor is Q35, are inferences. They rest on a mapping. If the mapping is adequate, the conclusions follow by structural necessity. If the mapping is inadequate, the conclusions fall. Contest the mapping.
All papers are published on Zenodo with DOI. The Ω-field simulator is available. The framework is open to falsification: anyone can verify the axioms, contest the mapping, or propose a domain that satisfies them without exhibiting the four properties. To date, across 14 domains, none has done so.