For several years, I have been watching a change. Social media, short video and algorithmic feeds have become ever more developed. People encounter far more information each day than they once did, yet many have less patience with a difficult problem.

This is not only a matter of attention being occupied. More troublingly, our cognitive habits may be changing. We are increasingly accustomed to stimuli that arrive in short bursts and increasingly expect immediate feedback. If something takes a long time to produce a result, or has no result for now, it easily feels boring.

I do not want to call this brain damage in the medical sense. That would require different evidence. I would rather begin by treating it as a functional change. Intelligence has not vanished, but it is becoming harder to keep it running for long and harder to consolidate it into something stable.

Intelligence has power and duration as well as height

We usually speak of intelligence as a relatively static value. But in real life, whether someone is intelligent and whether they can keep calling upon that intelligence are two different questions.

To describe the difference, I want to introduce the concept of ‘intellectual power.’ My first name for it was ‘intellectual power density,’ but if the subject is output per unit of time, ‘intellectual power’ is more exact. It approaches a power density only after division by attention, energy or some other resource.

Let \(W(t)\) be the effective intellectual work a person has produced by time \(t\). Their intellectual power is:

\[ P(t)=\frac{\mathrm{d}W(t)}{\mathrm{d}t}. \]

‘Work’ here is not the number of thoughts that crossed the mind or pieces of information viewed. It may be understanding a problem, building a model or making a judgement. It may also be an essay, a piece of code, or a method that remains useful later.

Total intellectual output over a period is the area under the power curve:

\[ W(T)=\int_0^T P(t)\,\mathrm{d}t. \]

The distinction matters. Some people reach very high instantaneous peaks. They look quick and active, but the peak is brief and the next thing soon interrupts it. Others may start slowly but continue working on the same problem. As the context becomes more complete, thought moves deeper.

The power curve of an extended thinker does not rise without limit. A session has a warm-up, entry into focus, the formation of structure, a plateau and finally fatigue. What can rise over years is average power, the upper limit of that power, and the ability to return to a deep state.

Fragmented thought looks more like a sequence of spikes. Each stimulus calls up some attention, but the next item arrives just as thought is preparing to descend. By the end of the day, a person may feel they have seen and thought a great deal. Integrate the whole curve, however, and little effective area remains.

Why the cognitive peaks keep falling

Fragmentation has another easily missed feature. A brief stimulus does not produce the same effect each time. Its marginal effect diminishes.

Let the intensity of stimulus \(n\) be \(S_n\), and a person’s sensitivity to this kind of stimulus be \(h_n\). The subjective reward is:

\[ R_n=h_nS_n. \]

As similar stimuli repeat, sensitivity may fall:

\[ h_{n+1}<h_n. \]

In a simple decay form, this can be written as:

\[ h_n=h_0e^{-\lambda n},\qquad R_n=h_0S_ne^{-\lambda n}. \]

This is a conceptual model, not a claim that real neural processes follow this exact exponential curve. It expresses a simpler idea: as the same stimulus repeats, the response it produces becomes smaller.

To maintain excitement, a platform must then increase the speed, novelty and emotional intensity of its content. The user swipes faster as well. At first, an item may hold attention for tens of seconds. Later, a few seconds are enough to decide that it is dull and seek the next one.

A strange separation then appears. External stimulation becomes stronger while less enters the cognitive structure.

Effective cognitive input can be written as:

\[ P_{\mathrm{in,eff}}(t)=P_{\mathrm{stim}}(t)A(t)M(t)\eta(t). \]

Here, \(P_{\mathrm{stim}}(t)\) is the power of external stimulation, \(A(t)\) is the attention actually invested, \(M(t)\) is the degree to which information forms meaning and structure, and \(\eta(t)\) is the efficiency of understanding and consolidation.

A fragmented environment may produce:

\[ P_{\mathrm{stim}}(t)\uparrow\qquad P_{\mathrm{in,eff}}(t)\downarrow. \]

Stimulation rises while absorption falls. A person continuously receives and reacts, with little time to place what just happened inside a larger structure.

Boredom here does not merely mean that content is bad. It may show that the reference point for stimulation has shifted. The same book or ordinary conversation has not suddenly become more boring. It supplies stimulation more slowly than the person has learned to expect.

Information can grow while intellectual capital declines

Intellectual power describes present processing. I care more about what remains after years.

I call the long-term stock of knowledge, skills, models, methods, experience and work intellectual capital, \(K(t)\). It is not the volume of information encountered. Information becomes something a person can continue using only after understanding, reconstruction and repeated recall.

The change in intellectual capital can be written as:

\[ \frac{\mathrm{d}K(t)}{\mathrm{d}t}=\eta(t)P_{\mathrm{eff}}(t)-[\delta+\phi F(t)]K(t). \]

Here, \(\eta(t)\) is the rate at which work is consolidated, \(P_{\mathrm{eff}}(t)\) is effective intellectual power, \(\delta\) represents forgetting and depreciation of ability, \(F(t)\) is the frequency of attention switching, and \(\phi\) represents the additional dissipation that switching imposes on cognitive structure.

The model contains two opposing forces. One consolidates present thought into long-term capital. The other keeps making established structures vague, loose or difficult to retrieve.

When:

\[ \eta(t)P_{\mathrm{eff}}(t)<[\delta+\phi F(t)]K(t), \]

then:

\[ \frac{\mathrm{d}K(t)}{\mathrm{d}t}<0. \]

A person can encounter great quantities of information every day while their intellectual capital shrinks.

There is nothing mysterious about it. Much of the information passes before the eyes without entering memory. Chains of reasoning are repeatedly broken halfway through, while deep reading and extended reasoning remain unpractised. The final loss is not limited to the hours spent on video. Even after putting the phone down, a person may need longer to enter a complex problem again.

This is also why extended thought is becoming more valuable. Its value is not the same as length. Two hours of empty video can remain shallow, while three minutes of speech may condense years of thought. What is scarce is a willingness to keep one problem in mind long enough for different materials to meet, to let temporary contradictions remain, and to wait while a stable model slowly forms.

The ability has direct economic value as well. Total intellectual output is not strictly proportional to income. A person must also choose worthwhile problems, turn thought into products, works, decisions or methods, and use technology, organisations and distribution as leverage. Other things being roughly equal, however, someone who can think continuously and accumulate intellectual capital usually holds more cognitive resources that remain useful over time.

Long-term cognitive advantage can be represented roughly as:

\[ G(T)\approx\overline{P}_{\mathrm{eff}}\cdot T\cdot\overline{\eta}\cdot\Gamma(T). \]

Here, \(\overline{P}_{\mathrm{eff}}\) is long-term average effective intellectual power, \(T\) is the duration of sustained investment, \(\overline{\eta}\) is the average consolidation rate, and \(\Gamma(T)\) is the compound effect produced by knowledge, work, technology and reputation.

AI supplies the answer, and may remove the process of forming it

AI adds another layer to the problem.

It can certainly reinforce cognition. People need not remember every source or complete every search, organisation and calculation alone. AI can supply external memory, connect fields and help reveal relations that were previously invisible.

But the reinforcement easily creates an illusion. The human-AI system has clearly become stronger, so the user naturally attributes the system’s ability to themselves.

Let a person’s endogenous intellectual power be \(P_H(t)\), the external intellectual power supplied by AI be \(P_{\mathrm{AI}}(t)\), and the total power displayed by the human-AI system be:

\[ P_{\mathrm{sys}}(t)=P_H(t)+\kappa(t)P_{\mathrm{AI}}(t). \]

AI readily raises \(P_{\mathrm{sys}}(t)\). Reports are written faster, answers arrive faster and the expression is more complete. But better system performance does not prove that the human’s intellectual capital \(K_H(t)\) has grown.

The following can occur:

\[ \Delta P_{\mathrm{sys}}>0\qquad\Delta K_H<0. \]

System output rises while the person’s endogenous ability falls. The pattern is easy to miss because everything looks better when judged by the final result.

I understand this as an attribution error involving externally attached intelligence. Obtaining an answer quickly and independently generating, understanding and testing it are different abilities. AI rapidly expands the first and may hide deterioration in the second.

Short video and AI do not create exactly the same problem. Short video breaks a long chain of thought into pieces. AI need not break the chain. Sometimes it simply removes the middle.

Without AI, a person often passes through not knowing, becoming curious, searching for material, forming a hypothesis, meeting a contradiction and revising the model before arriving at a provisional answer. AI can compress all of that into question and response. The final prose may be better, while the problem never truly unfolds inside the person.

Why the unknown drives thought

Much extended thought is driven by the unknown. A person has no answer, and the resulting tension keeps them searching. Let uncertainty be \(U\), and the drive to explore be \(D(U)\). An inverted-U conceptual model can represent their relation:

\[ D(U)=aUe^{-bU}. \]

With too little uncertainty, there is no reason to continue. With too much, a person may have no point of entry. A moderate unknown is most likely to sustain thought.

If AI supplies a complete and fluent answer too soon, it may rapidly reduce the unknown:

\[ U\rightarrow0\qquad D(U)\rightarrow0. \]

Sometimes the answer has not solved the problem. It has only made the problem look solved. Curiosity is soothed before a model forms, and extended thought stops.

AI’s long-term effect on a person depends on what it replaces and what it helps them practise. This can be written as:

\[ \frac{\mathrm{d}K_H(t)}{\mathrm{d}t}=\alpha R(t)+\beta V(t)-\mu O(t)-\delta K_H(t). \]

Here, \(R(t)\) represents understanding and reconstruction personally performed after using AI, \(V(t)\) represents verification, challenge and testing in reality, and \(O(t)\) represents cognitive activity outsourced to AI without understanding.

If AI helps someone reconstruct a problem, find counterexamples and compare explanations, \(R(t)\) and \(V(t)\) rise. AI becomes part of learning and thought. If memory, derivation, judgement and expression are all handed over, \(O(t)\) rises. The immediate output may still be excellent while endogenous intellectual capital slowly declines.

I therefore do not think the issue reduces to whether AI is a good or bad tool. Its default mode favours immediate answers, and immediate answers easily close thought too early. People must deliberately change their relationship with it.

Three forms of resistance

For now, I group resistance to cognitive dissipation into three parts: metacognition, multidimensional databases and timeline thinking. They are not isolated tricks so much as different parts of one cognitive system.

Metacognition

Metacognition first requires me to step outside the behaviour for a moment and observe myself.

Do I truly understand, or do I merely possess an answer? Why did I open AI at this moment? Am I removing mechanical work, or avoiding a piece of thinking that I should have done myself? If I close AI, can I still explain the logic?

This does not happen automatically. The more smoothly AI answers, the more easily I confuse familiarity with understanding. I therefore need to record my own judgement before asking, even if it is immature. After receiving the answer, I can close AI and try to explain, derive and challenge it again.

If I cannot reconstruct it, the answer remains inside AI. It has not entered my intellectual capital.

Multidimensional databases

A biological brain clearly has limits. The answer should not be to reject external tools, but to place AI inside a larger cognitive structure.

In that structure, the human mind is the central node. Personal notes, primary material, lived experience and AIs with different functions form the other nodes. One AI retrieves, another finds counterexamples, another makes connections across fields, and another can simulate a different position.

A multidimensional database cannot store conclusions alone. It must retain sources, strength of evidence, degrees of uncertainty, conflicting explanations, and the conditions under which a view was formed. Otherwise, a larger database may create more confusion.

AI can be one node or a cluster of nodes, but it should not become the final judge of the whole system. A person must still choose the questions, accept or reject evidence and decide when to revise a judgement.

Used this way, external intelligence does not simply cover human thought. It extends cognition into more fields and connects knowledge that was once separate. New ideas often appear at those connections.

Timeline thinking

Static analysis easily creates an impression of completeness. A model may explain past events beautifully without showing how events will change next.

Timeline thinking requires me to keep asking: how did this state form? Which variables changed first, and which appeared later? What outcome follows from the current model? Did earlier judgements come true?

The dynamic process can be written as:

\[ X_{t+1}=F(X_t,U_t,\varepsilon_t). \]

Here, \(X_t\) is the current state, \(U_t\) is human action and external input, and \(\varepsilon_t\) is an event that cannot yet be controlled.

A view that exists only in static prose can look correct forever. Put it on a timeline and it must carry a date, conditions and a prediction. Returning later gradually reveals which parts of the model worked, which failed and which causal relations were explanations invented after the event.

A timeline also resists the shallowness AI can bring. AI may provide an apparently complete answer at one moment, but I need not treat it as the endpoint. I can save it as a provisional version and let later facts continue to test it.

Cognitive sovereignty cannot be outsourced

Metacognition, multidimensional databases and timeline thinking all lead to the same question. Once AI enters my cognitive system, who controls that system?

I can give AI retrieval, organisation, calculation and even part of the reasoning. Human minds have always used paper, books, databases and other tools. There is no virtue in doing everything alone. But AI differs from earlier tools. It does not only preserve information. It organises evidence, explains causes and generates judgements. Without noticing, a user can move from thinker to receiver of results.

We cannot decide whether AI has augmented someone by output speed and quality alone. I must observe two variables at once: the total power of the human-AI system, \(P_{\mathrm{sys}}\), and the person’s endogenous intellectual capital, \(K_H\).

The dangerous state is:

\[ \Delta P_{\mathrm{sys}}>0\qquad\Delta K_H<0. \]

The system grows stronger while the person becomes more dependent on it. Without AI, they cannot reconstruct the answer, identify its assumptions or notice where it may be wrong.

I want to reach a different state:

\[ \Delta P_{\mathrm{sys}}>0\qquad\Delta K_H>0. \]

AI raises the processing ability of the system, while collaboration with it leaves the person with more knowledge, models and methods of judgement they can call upon independently. Only then does AI augment the person rather than replace one.

The two states are difficult to distinguish in the short term. Both produce polished prose, rapid answers and apparently complete analysis. The difference appears later. Can I derive the result again without the original answer? Can I explain why I believe it? When I meet a counterexample, can I revise the model, or can I only ask AI for another answer?

That is why timeline thinking is indispensable. Cognitive sovereignty cannot be proved by one moment of confidence. It depends on whether a person’s own \(K_H(t)\) grows or shrinks over long use of AI.

AI can be a node or a powerful cluster of nodes in a multidimensional database, but it cannot quietly replace the human control layer. A person must still decide which questions deserve continued attention, what evidence is enough to change a judgement, and when a conclusion should be retained or abandoned. Otherwise, the larger the database becomes, the easier it is to lose direction.

I do not object to obtaining answers from AI. I object when the answer arrives so quickly that the question has not formed inside me; when the expression is complete but I have not gone through the process of forming a judgement; and when the system has thought in my place but I mistake its ability for my own.

Some unknowns should therefore remain for a while. The confusion and discomfort they produce are not entirely inefficient. Sometimes an absent answer is precisely what keeps someone searching for material, finding contradictions, revising assumptions and slowly connecting things that seemed unrelated. Fill the unknown too early and thought may stop growing with it.

Cognitive sovereignty does not require me to perform every calculation personally or reject all external intelligence. It requires me to retain the ability to ask questions, reconstruct answers and revise models after borrowing AI’s power.

I decide why a question deserves thought. I examine whether an answer deserves trust. I also bear the consequences of the judgement.

If those powers remain in my hands, AI is an extension of my cognition.

If I have handed them away, then no matter how beautiful the answer on the screen may be, only the system has grown stronger. I have not.