Menschliches Gehirn und KI-gestütztes Gehirn im Vergleich.

The cognitive chasm: When AI overtakes our thinking

2 April 2026 Heinz Süess

New research · March 2026

One researcher has measured what many have suspected: AI systems now process information more than a thousand times more efficiently than humans – and the gap is growing rapidly. Worse still: We are helping to contribute to this.

3.906×
Growth of AI context windows since 2017
−87%
Decrease in human capacity to absorb information since 2004
1.111×
Raw ratio of AI to humans in 2026

The problem with the short sentence

Imagine it's 2026, you open your AI assistant and type: „Write me a cancellation for this meeting, two sentences.» Thirty seconds of uninterrupted thought – that would have cost this task earlier. Today you delegate it without thinking. It has become so natural that it feels wrong to even mention it.

This is exactly where the analysis by Netanel Eliav from the Machine Human Intelligence Lab begins. He calls this phenomenon the Cognitive divergence: the measurable gap that is growing between what AI systems can process and what humans are still willing to think for themselves.

Two curves that intersect

The basis of the paper are two well-documented trends that have never been directly compared to each other before.

AI page: The so-called context window – the information space that an AI model can process in a single thought step – has grown from 512 tokens in 2017 to two million tokens in 2026. That corresponds to a factor of just under 4,000 in nine years. For comparison: Moore’s Law describes a doubling every two years – the AI context windows double in size on average every 14 months.

Human side: Gloria Mark, a researcher at UC Irvine, measured from 2003 to 2020 how long knowledge workers remained focused on a screen for an uninterrupted period of time. The result: from an average of 150 seconds per year in 2004 to 47 seconds per year in 2020. Eliav translates these data into tokens – the same unit as the AI – and obtains the Effective Context Span (ECS): from around 16,000 tokens in 2004 to an estimated 1,800 tokens in 2026.

  • 2017
    AI: 512 tokens · Human: ~13,500 tokens — Humans are still clearly the advantage
  • 2021
    Codex (OpenAI): 4,096 tokens · Human: ~7,500 tokens — The curves are approaching
  • November 2022 — Turning point
    ChatGPT launches. AI overtakes humans for the first time. The gap begins to narrow.
  • 2023
    Claude 2: 100,000 tokens · Human: ~4,500 tokens — Factor 22
  • 2026
    Grok 4.20: 2,000,000 tokens · Human: ~1,800 tokens — Factor 1,111
„The lines have crossed. What began in 2022 was not a turning point in the history of AI alone – it was a turning point in the relationship between humans and machines.» Netanel Eliav, MHIL 2026

Why the distance is real – even after subtractions

To be fair, Eliav admits: The AI context window is a theoretical maximum, not a practical achievement. Research shows that language models process information worse at the middle of long input – the so-called „Lost in the Middle»-Phenomenon. Adjusted for quality, the effective AI context is closer to 100,000 to 200,000 tokens.

The ratio thus shrinks from 1,111-fold to 56–111-fold. A factor of up to 111 remains. The gap remains enormous.

The brain under constant digital bombardment

Why is human capacity for processing information actually declining? Eliav draws on eight neuroimaging studies. The picture is consistent:

Social media platforms train the brain on short-term rewards. Variable notifications – sometimes many likes, sometimes few – activate the same dopamine mechanism as gambling. The prefrontal cortex, responsible for concentration and impulse control, shows measurably reduced activity in multi-users. EEG data show altered brainwave patterns. And: Short videos (TikTok, Reels) are associated with a decrease in attention from r = −0.38 – a substantial effect across 98,000 subjects.

The consolation: It is reversible. Those who limit their social media use to 30 minutes a day show measurable recovery after three weeks.

The actual theory: The delegation feedback loop

The centerpiece of the paper is not the measurement itself, but the hypothesis about what will happen next. Eliav calls it the Delegation Feedback Loop.

The delegation feedback loop – four steps

  1. AI is becoming better and more accessible → The threshold at which humans delegate tasks is decreasing. Tasks that previously required their own thought are becoming routine tasks to delegate.
  2. Delegation below one’s own capacity → The cognitive work no longer takes place. No thinking = no training.
  3. Less training → Real atrophy. Attention, writing ability, critical thinking require regular practice. Without it, they deteriorate measurably.
  4. Lower capacity → Even more tasks fall under the reduced threshold value → The AI takes on even more tasks → The circle closes in.

Empirical evidence exists for every step of the cycle: A study with 666 participants shows that more intensive AI use reduces critical thinking. An MIT EEG study proves that AI-assisted writing leads to less cognitive engagement – and those who then write without AI perform worse than someone who never had help. The researchers call this „Cognitive Debt».

In short: We don't get worse because AI gets smarter. We get worse because we stopped practicing.

What that means for us

🎓
Education

AI learning systems that solve tasks for students instead of guiding them prevent the very cognitive learning process they are meant to promote.

🔍
Verifiability

When AI processes 100x more context than a human can review, classic testing workflows lose their foundation.

🏢
Profession

Anthropics own data show risks of deskilling: Those who permanently delegate complex thinking tasks lose them in the long run.

🧠
Public Health

Cognitive capacity is not a purely individual problem. A decline in populations is a structural risk – analogous to physical fitness.

What the paper does not claim

Eliav is careful enough to draw three clear boundaries:

The use of AI is not inherently harmful. Anyone who uses AI for tasks that really exceed their capabilities – such as analyzing a 200,000-token document – will not lose any skills they would have had anyway.

AI did not cause the decline in attention span. The curve has been declining since 2004 – long before ChatGPT. Smartphones and social media were the main drivers until 2020. The AI thesis concerns only what comes after that.

It is reversible. Neuroplasticity works in both directions. The cycle is self-reinforcing under current conditions – not inevitable.

The paper ends with a sentence that lingers:

„The professional who typed ‚send me a cancellation for this meeting‘ will not stop. The model will not get smaller. The question is what happens to the cognitive abilities that are no longer used.»

No study has measured this so far. Not yet. Eliava’s research agenda calls for exactly that: a long-term study over several years that tracks what is really happening to us while we are being made to think.

We do not yet know the answer. The question is late.

Source: Eliav, N. (2026). The Cognitive Divergence. arXiv:2603.26707. Preprint, not peer-reviewed.

Leave a comment

GDPR Cookie Consent with Real Cookie Banner