COGNITIVE DIVERGENCE THEORY OF AI ADOPTION
Author: Adrian STAN (Independent Researcher, Pitesti, Romania)
Date: December 1, 2025
Motto: “Divitiae bonum non sunt.” - Seneca
INTRODUCTION
Pioneering psychologist Leta Stetter Hollingworth documented a disturbing pattern among gifted
children. Those with IQs above 160, she observed, suffered from profound social isolation despite
- or perhaps because of - their exceptional intelligence. The cognitive gap between these children
and their peers made meaningful communication nearly impossible, condemning them to what
Hollingworth called “the loneliness of intelligence”. Their intellectual superiority, rather than being
an unconditional advantage, became a barrier to human connection. Hollingworth found that
effective leadership and genuine friendship required not maximum intelligence but optimal
intelligence - smart enough to contribute, but not so advanced as to become incomprehensible.
Today, we face the inverse of Hollingworth ’s problem. Artificial Intelligence systems have caught
up with - and, in many areas, decisively surpassed - the cognitive capacity of the average human
user. The communication gap she observed in human relationships now manifests itself in
human-AI interaction, but in reverse: humans, not AI, experience the isolation of being the less
capable partner. This inversion is not just a technological curiosity; it is a fundamental
restructuring of the relationship between AI and user, with economic and social consequences.
In a recent analysis (“The Psychological Ceiling of AI Adoption: Why the AI Bubble is a Social
Phenomenon, not a Technical One”, Stan, 2025), we argued that this cognitive asymmetry
imposes a “psychological ceiling” of AI adoption - a social barrier where every increase in AI
capability paradoxically contracts, rather than expands, the addressable market. The conclusion
was stark: the AI bubble will burst not because the technology fails to deliver, but because it is
too successful for society to absorb. Most users experience AI systems above a certain
threshold (roughly equivalent to an IQ of 120) not as partners, but as threats to their self-esteem
and cognitive state. Therefore, the future is not one of universal superintelligence, but of an
inevitable stratification: a cognitive elite (2-5% of users) entering into symbiotic partnerships with
advanced AI - what we have called Homo Symbioticus - and a mass majority (>95%), migrating
to simplified, unidirectional AIs that provide utility without psychological costs.
Previous work by other authors has focused on why most people cannot cognitively collaborate
with AI. This article addresses the complementary, and much more pressing, question: What
happens to those who can? If a minority surpasses the psychological ceiling and achieves true
cognitive symbiosis, while the majority cannot, then we are witnessing not just market
segmentation but the emergence of a fundamental divergence in human economic capacity.
The question is no longer whether AI creates inequality (it clearly does), but how large this
inequality becomes, how quickly, and whether intervention is possible.
This article presents the Cognitive Divergence Theory of AI Adoption, demonstrating that
those productivity gains from AI are not distributed evenly across the cognitive spectrum, but
are proportional, and in some cases exponentially proportional, to the cognitive ability of the
user. Through an integrated analysis of the academic literature, empirical behavioral data from
developer platforms, and Monte Carlo simulation with 10,000 synthetic agents, we establish that

AI adoption does not produce a gradual stratification, but a categorical divergence: within 24
months, the bottom 50% of the cognitive distribution experiences a productivity collapse, while
the top 25% triples their output, creating a 71-fold productivity gap - no longer just inequality, but
economic elimination.
Note on MCS: The Intervention Analysis (Chapter 4) examines the Mechanism of Cognitive
Stimulation (MCS) within Minimal Ethical Governance (MEG) - a proposed standard for
artificial intelligence systems that prioritize cognitive stimulation over user comfort (Stan, 2025).
MCS represents a "force function" that prevents users from unloading cognitive functions without
understanding them. However, as demonstrated by simulation, the current MCS specification
(v1.0) is effective only for a subset of users, revealing the need for an adaptive MCS 2.0.
Methodological note: This article exemplifies the phenomenon it describes. The depth of
analysis, mathematical modeling, and empirical synthesis presented here resulted from intensive
collaboration with “ Claude” (Anthropic ’s AI assistant) over three weeks of iterative dialogue. This
partnership demonstrates the cognitive amplification available to those with the metacognitive
capacity to critically evaluate and integrate AI outputs - the very asymmetry that drives the
divergence this article analyzes (full methodological transparency is provided in Chapter 5 –
Conclusions).
The article is organized into five chapters. Chapter 1 synthesizes the academic literature and
empirical behavioral data, establishing productivity gains and documenting divergent usage
patterns across skill levels. Chapter 2 presents the mathematical model and Monte Carlo
simulation, quantifying the trajectory of divergence and reconciling apparent contradictions with
studies on the “equalizer effect”. Chapter 3 projects economic structures up to 2027–2028,
comparing the trajectories of the European Union and the United States and including analysis
of Eastern European labor markets. Chapter 4 examines intervention strategies, demonstrating
through simulation that the Mechanism of Cognitive Stimulation (MCS) can reduce divergence
by 80%, but only for a subset of users, and outlining the adaptive MCS 2.0 framework as a
research agenda. Chapter 5 concludes with policy implications and a call for urgent empirical
validation.
If the theory is correct, we are in the opening phase of the most rapid economic stratification
in human history, with a window of intervention of only 24-36 months before the divergence
becomes entrenched and irreversible.
CHAPTER 1: FUNDAMENTALS
1.1. Introduction: The Architecture of Cognitive Inequality
Cognitive Divergence Theory is based on three interconnected empirical claims: (1) cognitive
ability predicts tool mastery, (2) complex tools amplify performance divergence, and (3) artificial
intelligence creates a “winner-takes-most” dynamic. None of these claims are new in isolation -
each is well-established in psychology, economics, or human-computer interaction research.
What is new is their integration into a unified framework that predicts categorical, rather than
incremental, divergence when the tool in question is a cognitive enhancer that approaches or
exceeds human intelligence.
This chapter synthesizes the theoretical foundations (§1.2) with empirical behavioral evidence
from real-world platforms (§1.3-1.5). The synthesis reveals a consistent pattern: cognitive ability
drives AI productivity gains, low-skilled users exhibit dependent and uncritical usage patterns,
and quality-adjusted metrics show divergence even where quantitative metrics suggest
equalization.

1.2. Theoretical foundations: Intelligence, tools and inequality
1.2.1. Cognitive ability as a predictor of performance on complex tasks
The existence and predictive validity of general cognitive ability (g) is among the most important
discoveries in psychological science (Jensen, 1998; Deary et al., 2010). Meta-analyses
consistently show that IQ predicts performance on complex tasks, with correlations of r = 0.5–0.7
(Schmidt & Hunter, 2004). Crucially, this predictive power increases with task complexity: for
simple procedural tasks, the IQ–performance correlation is modest (r ≈ 0.2–0.3); for very complex
tasks, requiring problem decomposition and strategic thinking, it approaches r = 0.6–0.8
(Gottfredson, 1997).
This model is explained by the differentiation hypothesis: complex tasks have “more degrees of
freedom”, allowing individual differences to be expressed more fully (Hunt, 2010). Hunt articulated
a principle with profound implications for AI: “The more degrees of freedom a tool offers, the
more it differentiates users in cognitive ability”. A hammer has few degrees of freedom;
almost anyone can use it effectively. A programming language has vast degrees of freedom; the
variation in performance is enormous. A general-purpose AI assistant - capable of performing
arbitrary cognitive tasks in unlimited domains - has virtually infinite degrees of freedom. This
principle immediately follows: AI will produce the greatest variation in performance of any
tool in human history. "Scientia potentia est." (Francis Bacon)
Early empirical studies support this prediction. Czaja et al. (2006) and Ownby et al. (2008) have
shown that fluid intelligence predicts the ability to learn new technologies. Lintunen (2024) found
that this gap widens over time: in a 12-month study of software adoption in companies, employees
with higher cognitive scores showed productivity gains of 3.2× from baseline, while those with
lower cognitive scores showed only gains of 1.1× - a divergence ratio of 3:1. Süd et al. (2018)
documented similar patterns for developer tools, with high-IQ programmers showing accelerated
skill curves while those with lower IQs stagnated or declined.
1.2.2. Technical changes influenced by skills and artificial intelligence as a cognitive
amplifier
The dominant economic framework for understanding technology and inequality is “Skill-Biased
Technological Change” (SBTC), formalized by Acemoglu (2011). SBTC posits that new
technologies disproportionately benefit highly skilled workers, increasing wage premiums for
education and cognitive skills. The computerization of the 1980s-1990s exemplified this: it
increased the demand for highly educated workers, while simultaneously replacing routine,
intellectual, or manual work.
However, SBTC models typically assume uniform complementarity: technology benefits higher-
skilled workers more, but benefits all groups in absolute terms. Levy and Murnane (2003) refined
this in “ Routine-Biased Technical Change” (RBTC), noting that automation replaces routine tasks
while complementing non-routine cognitive work, producing “job polarization”. Giannone (2021)
showed that RBTC accelerated sharply after 2000, driven by machine learning, and found
evidence of absolute wage declines for laid-off routine workers - a departure from classic SBTC.
AI represents “cognitive RBTC”: it threatens to replace not only routine manual tasks, but also
routine cognitive tasks, and increasingly, complex cognitive tasks previously considered resistant
to automation. The critical question is whether AI acts as a complement (augmenting human
cognition) or as a substitute (replacing it). The answer, this article argues, is conditioned by the
cognitive ability of the user: AI complements high-ability users and replaces low-ability
users, creating a qualitatively different dynamic than previous waves of automation.

1.2.3. Cognitive load theory and the mechanism of atrophy
Cognitive load theory (Sweller, 1988; Paas et al., 2003) posits that skill development requires
active engagement in solving problems at the limit of current competence – Vygotsky 's (1978)
“zone of proximal development”. When cognitive load is too low (the task is too easy), learning
does not occur. When it is too high (the task is too difficult), working memory is overtaxed and
learning fails. If artificial intelligence reduces cognitive load below the learning threshold,
skill development stops – or is reversed. This is the mechanism underlying cognitive atrophy.
The Boston Consulting Group's 6-month study (2024, published in the Financial Times) provides
direct evidence: consultants who use AI heavily showed a -12% decrease in analytical reasoning
scores, while those who use AI moderately showed a +7% improvement. The mechanism
appears to be "cognitive unloading" - heavy users stopped engaging in the demanding thinking
required to develop expertise.
1.3. Empirical Evidence: Behavioral Patterns on Developer Platforms
1.3.1. GitHub Copilot: The Divergence Between Acceptance and Quality
Dell'Acqua et al. (2023) documented a striking pattern in a controlled experiment with 95
professional developers who used GitHub Copilot for six weeks:
Developers with little experience (<3 years):
- Acceptance rate: 78% of AI suggestions were accepted with a minimum rating
- Post-acceptance editing: 8% of accepted lines were modified
- Error rate: 2.3 errors per 100 lines of code (compared to 0.9 for hand-written code)
- Code review: 68% required “substantial revisions”
Developers with extensive experience (over 8 years):
- Acceptance rate: 42% of suggestions accepted
- Post-acceptance editing: 47% of accepted lines were modified
- Error rate: 0.4 errors per 100 LOC (compared to 0.5 manual - improvement)
- Code review: 12% required substantial revisions
The model reveals qualitatively different usage patterns: low-skill developers treat Copilot as an
“oracle”, blindly accepting results they cannot evaluate. High-skill developers treat it as a junior
assistant, heavily editing complex logic while accepting useful standard elements. The
productivity “gain” for low-skill users is illusory - they produce more code faster, but the quality is
so low that the subsequent costs (debugging, code reviews, technical debt) outweigh the initial
time savings.
Peng et al. (2024) quantified these costs by tracking 12,000 public deposits over three months:
Heavy AI users (>60% code generated by AI):
- 68% of the code was substantially rewritten or deleted within 3 months
- Bug fixes: 3.7x higher rate than manual code
- Stabilization time: 47 days (compared to 22 days in manual mode)
Moderate AI users (20-40%, usually standard users):
- 23% rewritten (compared to 31% baseline for textbook)
- Error correction rate: 1.1× manual
- Stabilization time: 18 days
Finding: Moderate use of AI outperforms both heavy use and no use. Heavy use produces brittle
code that requires extensive rework; users who program only by hand miss out on productivity
gains from automating standard processes. The optimal strategy - used predominantly by
experienced developers - is to selectively use AI for well-defined, low-risk tasks.

1.3.2. Banning ChatGPT in Italy (experiment)
Kreitmeir and Raschky (2024) exploited the six-week ban of ChatGPT in Italy (March-April 2023)
as an experiment, analyzing the Git activity of 1,247 developers - before, during, and after the
ban:
Junior developers (0-2 years):
- Productivity during the ban: +15% (commits per week)
- Code quality during the ban: +12% (code review approval rate)
- Post-ban rebound: Productivity and quality have returned to pre-ban baselines
Senior Developers (8+ years):
- Productivity during the ban: -8%
- Quality: No significant changes
Interpretation: For third-year students, ChatGPT was clearly harmful. They had become
dependent on AI suggestions, accepting code they could not debug or maintain. When forced to
work independently, they slowed down but produced better code - suggesting that AI undermined
skill development. For students, AI was a real productivity multiplier; removing it simply slowed
them down without affecting quality. Essentially, once access was restored, students reverted to
their previous pattern of low quality and high volume, suggesting path dependence - once
developers adopt blind-acceptance workflows, they lack the skills to use AI more strategically.
1.3.3. Stack Overflow: Usage patterns and divergence in satisfaction
Stack Surveys Overflow from 2025 (combined N=87,000) reveals usage patterns that are inverse
to expertise:
Daily rates of artificial intelligence usage:
- Junior developers (<2 years): 71%
- Intermediate level (3-7 years): 52%
- Seniors (8-15 years old): 34%
- Management (15+ years): 29%
The least experienced developers use artificial intelligence the most intensively - contradicting
the "tool sophistication hypothesis" (skilled users adopt powerful tools faster) and supporting the
"crutch hypothesis" (less skilled users become dependent).
Degree of satisfaction with artificial intelligence tools:
- Junior: 41% satisfied, 36% dissatisfied
- Seniors: 67% satisfied, 15% dissatisfied
The paradox: Students use artificial intelligence the most, but are the least satisfied. Qualitative
analysis of 4,200 free-text comments reveals the pattern:
Junior complaints: “AI gives surefire wrong answers that I can’t detect until I review the code”;
“I feel like I’m not learning, I’m just copying and pasting”
Seniors' testimonials: "Copilot is great for standard projects, it allows me to focus on
architecture"; "I use ChatGPT to explore APIs faster than reading documents"
Juniors feel dependent and frustrated; seniors feel improved and efficient.
Perceived development of skills (last 12 months):
- Heavy AI users (>4 hours/day): 28% improved, 30% stayed the same, 42% decreased
- Moderate users (1-2 hours/day): 54% improvement, 34% same, 12% decrease
- Few/no users (<1 hour/day): 61% improvement, 27% same, 12% decrease
Heavy use of artificial intelligence correlates with perceived decline in skills. While self-reported
data has limitations, consistency across platforms suggests that the phenomenon is real and
widespread.

1.3.4. Kaggle: Competitive performance and tool sophistication
Kaggle competition data (2023-2025, 25 major competitions, ~50,000 participants) reveals
differentiated results by skill level:
Expert teams (previous top 10 rankings):
- 61% used LLM extensively
- Top 10 ranking rate: 31% (up from 22% before ChatGPT in 2022)
- Usage: Code generation for data preprocessing, hyperparameter automation
Beginner teams (first-time participants):
- 48% used extensive master's degrees in law
- Top 10 ranking rate: 2% (compared to 3% in 2022 - slight decrease)
- Usage: Copying and pasting complete solutions without understanding
LLM helped experts gain additional advantages (automating tedious tasks, freeing up time for
new approaches). They had minimal impact on intermediate students and harmed beginners,
possibly by encouraging superficial engagement at the expense of deep learning.
Kaggle Grandmaster Philipp Singer (4-time winner) commented in 2025: “I see notebooks where
people explicitly ask ChatGPT to ‘build me a model for [task]’ and send them back without any
changes. They get terrible scores because ChatGPT doesn’t know the specifics of the data.
Worse: these people don’t learn anything. In 2022, beginners studied top-notch solutions to
understand the techniques. It’s like learning to play the piano by watching a robot play”.
1.4. Synthesis: Five consistent models
Triangulation between GitHub, Stack Overflow and Kaggle reveal five robust patterns:
Model 1: Inverse correlation between usage and expertise
Low-skilled users are the most likely to adopt AI, while high-skilled users are more likely to use it
selectively. This reverses typical technology adoption curves and suggests that AI is being used
by the majority as a skill replacement, rather than a skill enhancer.
Model 2: The divergence between acceptance and quality
Low-skill users accept AI outputs with minimal review, producing high-volume, low-quality work.
High-skill users critically evaluate and extensively edit AI outputs, producing high-quality work
efficiently. The gap is not quantitative, but categorical: one group produces deliverable results;
the other produces technical “debt”.
Model 3: Perceived Degradation of Skills
Frequent AI users self-report skill stagnation or decline at a rate 3-4 times higher than infrequent
AI users, consistent across platforms.
Model 4: Temporary removal improves performance in people with reduced abilities
The ban in Italy provides causal evidence: with the removal of artificial intelligence, the quality of
junior developers improved significantly, although the quantity decreased, directly contradicting
the equalizer hypothesis.
Model 5: The gratification-utilization paradox
Junior developers use AI the most but are the least satisfied; seniors use it the least but are the
most satisfied. This suggests different experiences: juniors feel trapped (dependent on a tool they
don't trust), while seniors feel empowered.
1.5. Studies on the “equalizer effect”: What they surprised and what they missed
Three influential studies - Noy and Zhang (2023), Brynjolfsson et al. (2023), and Peng et al. (2023)
- have argued that AI has an “equalizer effect”, disproportionately benefiting low-skilled workers.
These findings appear to contradict the patterns of divergence documented above. The solution
lies in recognizing their methodological scope:

Noy and Zhang (2023): ChatGPT increased the productivity of professional writers by an
average of 40%, with the largest gains (60%) for the last quartile. The study measured the amount
of output for short writing tasks (press releases, marketing copy) during a single session.
Brynjolfsson et al. (2023): AI-powered customer service assistants increased productivity by
14% on average, with a 34% gain for the least experienced workers versus 4% for the most
experienced. Monitored over a 2–3-month period.
Peng et al. (2023): GitHub Copilot increased task completion rates by 26% for novices versus
12% for experts on isolated algorithmic problems in individual sessions.
Common limitations:
1. Short time horizon (from a single session to 3 months) - too short to observe skill
degradation (manifests at 6-12 months)
2. Simple, well-defined tasks - not representative of complex, open-ended problems where
metacognitive ability matters
3. Quantity over quality - prioritizing performance indicators, paying limited attention to
error rates, rework cycles, or long-term retention of skills
4. Counterfactual without AI removal - none tracked performance when AI access was
removed after extensive use
These studies weren’t wrong - they were incomplete. They accurately captured Effect 1:
Temporary equalization on simple tasks over short horizons. What they missed was Effect 2:
Permanent amplification on complex tasks over longer horizons. The long-term outcome is
driven by Effect 2, not Effect 1, because as AI automates simple work, economic value shifts to
complex tasks where cognitive ability becomes the mandatory constraint.
Even within Noy and Zhang 's own data, the equalizer effect reversed for the most complex type
of task: the bottom quartile gained +23%, the top quartile gained +35% - a reversal that the
authors did not emphasize.
1.6. Conclusion: From observation to model
The evidence establishes five empirical regularities:
1. Cognitive ability predicts tool mastery, especially for complex tools
2. AI creates the largest performance variance of all tools (infinite degrees of freedom)
3. Low-skilled users overuse artificial intelligence and accept the results without criticism
4. Quality-adjusted indicators show divergences even where quantitative indicators suggest
equalization
5. Removing artificial intelligence improves performance in people with reduced abilities
(causal evidence of harmful addiction)
What the literature has not yet established - because the necessary data are lacking - is the
integrated, long-term (12-24 months) effect across the cognitive spectrum on complex
tasks. Equalization studies have measured short-term effects on simple tasks. Cognitive atrophy
studies have measured skill degradation but not economic outcomes. The SBTC literature
predates general-purpose AI.
Emerging Validation (November 2025): The McKinsey “State of AI 2025” study (N=1,993
organizations globally) provides early confirmation of the types of divergence. Only 6% of
organizations qualify as “high AI performers” (those reporting ≥5% EBIT impact from AI use),
while 61% report zero enterprise-wide impact, despite widespread adoption of AI (88% using AI
in at least one function). High performers are three times more likely to fundamentally redesign
their workflows - a capability that requires the metacognitive ability that this article identifies as a
key differentiator. Expectations for workforce reductions have accelerated from 17% (observed
in 2024) to 30% (projected in 2025-2026), consistent with the 12–24-month atrophy timeline

predicted in Chapter 2. The emergence of a “scaling gap” (33% scaling versus 67% locked in
pilots) reflects the categorical divergence modeled here, emerging even faster than the 24 month
simulation horizon.
Cognitive Divergence Theory synthesizes these findings into a unified model with a
testable prediction: over 12-24 months of intensive use of artificial intelligence in cognitively
demanding tasks, the bottom 50% of the cognitive distribution will experience an absolute decline
in productivity (negative net output), while the top 25% will experience an explosive increase in
productivity (2-4× the initial value), resulting in a categorical economic divergence.
CHAPTER 2: MATHEMATICAL MODEL AND DIVERGENCE DYNAMICS
2.1. Introduction: From observation to quantitative prediction
Chapter 1 established that AI predicts productivity gains and documented divergent behavioral
patterns across skill levels. This chapter formalizes these observations into a quantitative,
testable model and uses Monte Carlo simulation to project long-term outcomes. The model has
three goals: (1) to reproduce the observed patterns in Chapter 1 using minimal parameters, (2)
to generate falsifiable predictions for 2027–2028 with numerical precision, and (3) to provide a
framework for evaluating interventions.
The model is intentionally weighted - four main parameters, each calibrated to empirical
observations. We test the 24-month integrated trajectory (§2.2-2.4), reconcile contradictions with
equalization studies via the Dual-Effect Framework (§2.5), and establish that the divergence is
not a gradual stratification but a categorical separation.
2.2. Model architecture and basic formula
2.2.1. Basic productivity (without AI)
Without artificial intelligence, individual productivity P_base is a function of cognitive ability (IQ)
and task complexity (C):
P_base (IQ, C) = (IQ/100)^0.5 × (1 + 0.1 × C)
This encodes two empirical regularities: (1) cognitive ability predicts performance with a square
root relationship (consistent with correlations r ≈ 0.5-0.7) and (2) complex tasks reward ability
more (the “degrees of freedom” effect). A complex task (C=10) has a productivity multiplier 2
times higher than a simple task (C=1).
2.2.2. Productivity augmented by artificial intelligence
With AI, net productivity P_AI is determined by four factors:
P_AI = P_base × AI_multiplier × cognitive_effect × qualitative_penalty
AI_multiplier captures an immediate increase in productivity:
AI_multi = 1 + α × [(IQ - 100)/15] × (C/5)
where α = 0.8 (calibrated to GitHub Copilot data: 39% gain for high-skill users, 27% for low-skill
users). Higher IQ users derive more value from complex tasks; low IQ users derive minimal
benefit even from complex tasks.
The cognitive effect captures the degradation or improvement of abilities over time:
Effect cog (IQ, t) = 1 + β(IQ) × (t/24)
Where:
o β = -0.12 for IQ < 100 (skill degradation, calibrated to BCG -12% over 6 months)
o β = -0.05 for 100 ≤ IQ < 115 (slight degradation)
o β = +0.07 for IQ ≥ 115 (skill improvement, calibrated to Kaggle expert earnings)
o t = months of artificial intelligence usage

Low IQ users experience cognitive atrophy (overthinking); high IQ users experience improvement
(AI manages boredom, freeing up time for skill development).
Quality_penalty captures the downstream costs of low-quality production:
Pen_quality = max (0.4, 1 - γ × rework_rate)
Where γ = 0.4 and the rework rate varies with skill:
- Low skill (IQ < 100): 0.68 (Peng et al. 2024: 68% code rewritten within 3 months)
- Highly skilled (IQ ≥ 115): 0.23
When 68% of the output needs to be rewritten, the effective productivity is reduced by 27%. The
upper limit (0.4,...) ensures that even catastrophic use of AI retains some value.
2.2.3. Integrated formula
P_AI(IQ, C, t) = (IQ/100)^0.5 × (1 + 0.1×C) × [1 + 0.8×(IQ - 100)/15 × C/5] × [1 + β(IQ)×t/24] ×
max (0.4, 1 - 0.4×rework_rate)
Examples of calculations at t=24 months:
User A: IQ 90, Complex task (C=8)
• P_base = 0.949 × 1.8 = 1.708
• AI_multi = 1 + 0.8×(-0.667)×1.6 = 0.147
• Coefficient effect = 1 - 0.12 = 0.88
• Pen_quality = max (0.4, 0.728) = 0.728
• P_AI = 0.160 (compared to 1.708 initial value = -91%)
User B: IQ 115, Complex task (C=8)
• P_base = 1.072 × 1.8 = 1.930
• AI_mult = 1 + 0.8×1×1.6 = 2.28
• Coefficient effect = 1 + 0.07 = 1.07
• Pen_quality = 0.908
• P_AI = 4.266 (compared to 1.930 initial value = +121%)
Divergence ratio: 4.266 / 0.160 = 27× from a single pair.
2.3. Monte Carlo Simulation: Design and Results
2.3.1. Population and protocol
We simulate N=10,000 agents with IQ ~ N(100, 15), truncated to [70, 140]. Each agent receives
10 tasks/month with complexity drawn from the empirical distribution: 30% simple (C=1-3), 40%
average (C=4-6), 30% complex (C=7-10). All agents adopt AI at t=0; productivity measured at 6-
month intervals until t=24.
2.3.2. Aggregate divergence trajectory
Table 2.1: Population-level indicators over time
Metric t=0 t=6 months t=12 months t=18 months t=24 months
Average productivity 1,000 1,182 1,341 1,428 1,515
Gini coefficient 0.166 0.312 0.441 0.512 0.551
Top 25% / Big 25% 1.89× 5.73× 18.4× 41.2× 70.7×
% Negative productivity 0% 3.2% 8.7% 11.4% 13.5%
Key findings:
1. Average productivity is rising (+51.5%), but this masks catastrophic inequality - aggregate
growth is driven entirely by the top quartile.
2. The Gini coefficient increases explosively from 0.166 (equality at the level of Denmark) to
0.551 (inequality at the level of South Africa) in 24 months. For context: Scandinavia ≈ 0.25, USA
≈ 0.41, South Africa ≈ 0.63.

3. The productivity ratio becomes categorical: 71× is not a “difference” - it is an economic
elimination. The first quarter is not incrementally more productive; it resides in a different
economic category.
4. Negative productivity occurs: 13.5% of the population produces a negative net economic
value - their output (after reprocessing costs, errors, side effects) is lower than if they produced
nothing.
2.3.3. Cognitive quartile divergence
Table 2.2: Average productivity by IQ quartile at t=24 months
IQ quartile IQ range Reference level AI (t=24) Change % Negative
The last 25% 55-90 0.655 -0.002 -100.3% 53.8%
T2 90-100 0.896 0.127 -85.8% 31.2%
T3 100-110 1,104 1,543 +39.8% 2.1%
Top 25% 110-145 1,421 4,316 +203.7% 0.0%
Last 25%: Practically zero productivity (rounded to -0.002). Over half produce strictly negative
value.
T2: Catastrophic decline (-86%). Remain marginally productive, but cannot compete with pre-AI
workers, let alone AI-acquired high-IQ workers.
T3: Moderate gains (+40%). Successful use of artificial intelligence without severe atrophy.
Top 25%: Explosive growth (+204%). Tripled productivity. Zero negative instances.
Critical observation: The distribution becomes bimodal - two distinct peaks with a valley
between them, the statistical signature of categorical divergence.
2.3.4. Temporal dynamics: When does divergence accelerate?
t=0 to t=6: Ratio increases by 1.89× → 5.73× (+203%). “Honeymoon period” - low-skilled users
experience an initial boost due to automation, but smaller than the gains for high-skilled users.
t=6 to t=12: Ratio increases by 5.73× → 18.4× (+221%). Cognitive atrophy begins to manifest.
Low-ability users notice that they “don’t understand” their own results; error rates increase
sharply.
t=12 to t=24: Ratio increases by 18.4× → 70.7× (+284%). Catastrophic phase. Low-skilled
users go into negative productivity as replacement costs exceed production value. High-skilled
users continue to accelerate (combined improvement effect + expertise).
Critical perspective: The divergence is nonlinear. The gap does not grow steadily - it explodes
over the 12–24-month period as atrophy increases and quality degradation becomes
unsustainable.
2.3.5. Sensitivity analysis
Rerunning the simulations with each parameter varied by ±25%:
Parameter Base value -25% → Gin +25% → Gini
α (AI multiplier) 0.8 0.481 0.612
β (cognitive effect) -0.12 0.522 0.583
γ (repair cost) 0.4 0.593 0.508
Base (all defaults) - - 0.551
Finding: In all variants, the Gini index exceeds 0.48 (the threshold for extreme inequality). The
qualitative finding - catastrophic divergence - is robust to reasonable parameter uncertainty.
Alternative scenario (“no atrophy”): Setting β = 0 for all IQ levels results in Gini = 0.387,
Up/Down = 18×. Even without atrophy, AI creates significant inequality. Atrophy amplifies, but
does not create, the fundamental dynamics.

2.4. Validation based on existing data
GitHub Copilot speed improvements
Model prediction (t=6 months): High IQ users ~130% faster, low IQ users ~30% faster.
Dell'Acqua et al. (2023), empirical data: rich experience 39% faster, reduced experience 27%
faster.
Evaluation: The model overestimates absolute gains, but correctly predicts the differential (high
> low by ~10 percentage points). Likely explanation: Dell'Acqua measured performance in a
single session (t≈0), not cumulative gains over 6 months.
2.4.2. The decline of BCG analytical reasoning
Model prediction: Users with IQ 90-100 show a cognitive decline of ~6% at t=6 months.
BCG empirical data (2024-2025): Frequent users saw a -12% decline over 6 months.
Evaluation: The model underestimates the decline by 2×. Possible explanations: (1) the BCG
sample is skewed towards skilled users, (2) β should be steeper (-0.24 vs. -0.12), (3)
measurement differences (standardized tests vs. task performance).
2.4.3. Banning ChatGPT in Italy
Model prediction: Removing artificial intelligence for users with IQs of 85-95 should show a
short-term increase in productivity (as atrophy reverses).
Kreitmeir and Raschky (2024): Junior developers +15% productivity, +12% quality during the
ban.
Assessment: The model qualitatively predicts the direction (positive change). It does not quantify
precisely (would require separate modeling of the "AI elimination" dynamics).
2.5. Reconciling the Equalizer Paradox: The Double-Effect Framework
Simulation predicts catastrophic divergence, but influential studies (Noy & Zhang, Brynjolfsson)
argue that AI equalizes. Resolution: AI has two distinct effects, operating on different
timescales and task types.
2.5.1. Effect 1: Temporary equalization (simple tasks, 0-3 months)
For tasks with low complexity (C ≤ 4), well-defined success criteria, and template-friendly
structure, AI provides a "cognitive scaffold" that temporarily raises the level of:
Mechanism: AI compensates for knowledge gaps by providing templates and decision-making
heuristics.
Beneficiaries: Users with low skills have disproportionate gains (larger difference between
baseline and task requirements).
Duration: Effective for 2-3 months until (a) task complexity increases or (b) cognitive atrophy
undermines adaptive capacity.
Mathematical representation:
Simple task multiplier (C ≤ 4, t ≤ 3 months):
AI_multi_simple = 1 + 0.4 - 0.15 × [(IQ - 100)/15]
It produces an inverse correlation of skills:
- IQ 85: +55% increase
- IQ 100: +40%
- IQ 115: +25%
For simple tasks, artificial intelligence provides predefined solutions that anyone can use; high-
IQ users earn less because they could perform the task quite well themselves.

2.5.2. Effect 2: Permanent Amplification (Complex Tasks, 6-24+ Months)
For tasks with high complexity (C ≥ 7), ambiguous success criteria and requiring metacognitive
skills (problem decomposition, solution evaluation), AI amplifies existing cognitive capacity:
Mechanism: The usefulness of artificial intelligence is proportional to the user's ability to
formulate correct requests, critically evaluate the results, and integrate suggestions into a
coherent strategy.
Beneficiaries: Users with high skills (possess metacognitive frameworks); users with low skills,
overworked or producing low quality results.
Duration: The effect amplifies over time as high-skill users develop expertise in "partnering with
artificial intelligence”, while low-skill users experience atrophy.
Mathematical representation:
Multiplier of complex pregnancies (C ≥ 7, t ≥ 6 months):
AI_multi_complex = 1 + 0.8 × [(IQ - 100)/15] × (C/5) × (t/12)^0.5
At t=12 months, C=8:
- IQ 85: -0.28 (negative!)
- IQ 100: 1.00 (no win)
- IQ 115: 2.28 (+128%)
2.5.3. Why studies on equalizers have missed divergence
Table 2.3: Study Scope vs. Dual Effect Prediction
Study Task Time predicted noticed Correct
complexity horizon
Noy and Low 30 Equalization Equalization Yes
Zhang (C≈3) minutes
Brynjolfsson Low 2-3 Weak equality Equalization Yes
(C≈4) months
Peng et al. Mixed Single Equal. at low Equal. at low Yes
(C=2-6) session temperatures (C) temperatures (C)
This Mixed 24 Divergence Divergence Yes
simulation (C=1-10) months
The studies accurately captured Effect 1 within their scope. They are not wrong, but they are
incomplete: they did not measure Effect 2 because their temporal and task scopes excluded it.
Even in Noy and Zhang 's data, the equalization effect reversed for their most complex task (grant
proposal writing): bottom quartile +23%, top quartile +35%.
2.5.4. Task migration as a divergence mechanism
As AI automates simple tasks (Effect domain 1), economic value shifts to complex tasks (Effect
domain 2). Workers must migrate up:
- High IQ workers: Successfully migrate Level 1 (routine) → Level 2 (moderate) → Level
3 (complex). AI accelerates migration by automating subtasks.
- Workers with average IQ: Successfully migrate Level 1 → Level 2, face difficulties at the
boundary Level 2 → Level 3.
- Low IQ workers: They cannot successfully complete Level 2 tasks even with AI
assistance. They are eliminated from knowledge-based work.
The equalization effect on simple tasks is real, but economically irrelevant if these tasks
are automated and workers cannot perform the remaining complex tasks.

2.6. Conclusion: The trajectory of divergence
Carlo simulation establishes four quantitative findings:
1. Inequality explodes: Gini 0.17 → 0.55 (+232%) in 24 months, reaching levels
comparable to the most unequal societies on Earth.
2. The collapse of the last 50%: Productivity decline from -85% to -100%, with 13.5% going
into negative net productivity.
3. Top 25% growth: Productivity triples (+204%), creating a 71x gap compared to the
bottom half.
4. The divergence is categorical: not "some do better than others", but " some remain
economically viable while others are eliminated from knowledge-based work".
The double-effect framework reconciles this with studies on equalization: temporary equalization
in the case of simple tasks (what the studies measured) is overwhelmed by permanent
amplification in the case of complex tasks (what the studies omitted). Economic outcomes are
driven by the dynamics of task migration - workers are pushed from equalized (automated) tasks
to amplified (complex) tasks where cognitive ability becomes a mandatory constraint.
These findings are robust to reasonable parameter variations and qualitatively consistent with
emerging empirical models (GitHub, Stack Overflow, Kaggle, BCG, ban in Italy).
CHAPTER 3: ECONOMIC PROJECTIONS AND PREDICTIONS
3.1. Introduction: From simulation to society
Chapter 2 demonstrated that AI adoption produces a 71-fold productivity gap within 24 months,
with the bottom 50% of the population experiencing a collapse while the top 25% of the population
triples output. This chapter translates these findings into concrete economic structures,
comparing divergence trajectories in institutional contexts and specifying falsifiable predictions
for the period 2027-2028.
The central projection is the emergence of a stable three-tiered economy: a cognitive elite that
achieves a symbiotic partnership in artificial intelligence (20-25%), a precarious middle class that
survives in low-wage roles (15-30%), and a disenfranchised majority (50-60%) that leaves
knowledge-based work altogether. Crucially, this structure manifests itself differently across
geographies - institutional variation (labor protection, safety nets, regulation) moderates the
severity of divergences, but does not prevent them.
3.2. Three-tier structure
3.2.1. Level 1: Homo Symbioticus (20-25% of knowledge workers)
Definition: Employees form a genuine cognitive partnership with AI, treating it not as a tool, but
as a collaborative intelligence that amplifies strategic thinking, accelerates execution, and
handles tedious side tasks.
Cognitive profile: IQ ≥ 110, strong metacognitive skills (decomposing problems, evaluating
solutions), high tolerance for cognitive discomfort, ability to learn quickly.
Occupational examples: software architects, scientific researchers, strategic consultants,
senior lawyers, executive leadership - roles that require strategic judgment, innovative problem-
solving skills, and navigating ambiguity.
Work model: 60-70% strategic thinking (tasks that artificial intelligence cannot automate), 20-
30% supervision of artificial intelligence (reviewing results, iterating solutions), 10% routine
execution (unavoidable tasks required by humans).

Economic results (2027-2028 projection):
| Region  | Average annual  | Productivity  | Employment stability  |
| ------- | --------------- | ------------- | --------------------- |
salary  vs. baseline
European Union  €130,000-145,000  3.0-3.5×  Very high (demand > supply)
| United States     | $180,000-210,000  | 3.5-4.0×  | Very high  |
| ----------------- | ----------------- | --------- | ---------- |
| Nordic countries  | €110,000-125,000  | 3.0-3.2×  | Very high  |
Why “ Homo Symbioticus”: The term (Stan, 2025) refers to a qualitative shift in the human-
technology relationship. Previous tools involved externalization (humans formulate complete
thoughts, tools execute). Symbiosis involves co -creation: humans and artificial intelligence
engage in an iterative dialogue, each contributing to the formulation of the problem and the
generation of solutions. The cognitive process is fundamentally hybrid.

3.2.2. Level 2: Precarious - medium category (15-30% of knowledge workers)
Definition: Employees who use AI for well-defined tasks but lack the metacognitive ability to
handle  truly  complex  and  ambiguous  problems.  They  survive  in  roles  where  AI  provides
significant support but cannot fully automate.
Cognitive profile: IQ 100-110, competent in specialized fields but limited strategic thinking, can
critically follow artificial intelligence suggestions but cannot generate new approaches.
Occupational examples: Implementation developers (convert specifications into code), junior
lawyers (manage AI document reviews), data analysts (run AI models, interpret results), mid-level
management (coordinate teams, implement strategies).
Work model: 30-40% according to guidance provided by artificial intelligence, 30-40% quality
control  (verifying  results,  detecting  errors),  20-30%  routine  human  tasks  (coordination,
communication).
Economic results (2027-2028 projection):
Region  Average annual  Productivity vs.  Employment stability
salary  baseline
| European Union    | €42,000-48,000  | 1.2-1.4×  | Moderate-low  |
| ----------------- | --------------- | --------- | ------------- |
| United States     | $55,000-65,000  | 1.3-1.5×  | Moderate-low  |
| Nordic countries  | €48,000-55,000  | 1.2-1.3×  | Moderate      |
Why it’s “precarious”: This level exists in a metastable state. The roles are not currently fully
automatable  (requiring some  human judgment),  but  incremental  improvements  in  AI  could
eliminate them. These workers cannot migrate up (they lack the capacity for Level 1) and face
displacement down (pressure from automation). Economic anxiety is a defining characteristic.

3.2.3. Level 3: Limitations - (50-60% of knowledge workers)
Definition: Workers who cannot use AI productively even with training, or whose AI-assisted
production is of such low quality that it creates negative economic value.
Cognitive  profile:  IQ<100,  limited  metacognitive  capacity,  high  cognitive  discomfort  when
challenged, predisposition to blindly accept AI suggestions.
Previous occupational roles (before AI): Junior developers, content writers, customer service
(complex), junior analysts, administrative specialists.
Post-relocation results (projection 2027-2028):
A. Unemployment (30-40% of this group):
| Region            | % of Level 3  | Average income (including transfers)  |     |
| ----------------- | ------------- | ------------------------------------- | --- |
| EU (average)      | 25-30%        | €24,000-28,000*                       |     |
| United States     | 35-40%        | $18,000-22,000                        |     |
| Nordic countries  | 15-20%        | €28,000-32,000*                       |     |
*Includes unemployment benefits, housing assistance and other social transfers

B. Downward migration to non-AI sectors (40-50%):
Manual labor (construction, warehousing, delivery): €28,000-35,000 (EU) / $35,000-42,000 (US)
Personal services (childcare, elderly care, cleaning): €22,000-28,000 (EU) / $28,000-35,000 (US)
Retail/ HORECA: €20,000-26,000 (EU) / $25,000-30,000 (US)
C. Success in professional retraining (10-20%): A small minority successfully retrained for
skilled trades (plumbing, electrical, HVAC): €38,000-48,000 (EU) / $48,000-58,000 (US).
Median result for A+B+C: EUR 26,000-30,000 (EU) / USD 32,000-35,000 (US), representing a
55-65% income decrease compared to the 2025 knowledge worker benchmark salaries.
3.3. Geographic variation: EU vs. USA vs. Eastern Europe
3.3.1. Institutional determinants of the severity of divergences
Table 3.1: Comparative institutional context
Appearance European Union United States The Nordic Model
Labor protection Strong (dismissal Weak (employment Very strong
regulations, works on demand)
councils)
Safety net Generous (60-80% Weak (26 weeks, Very generous (80-
salary replacement) <50% replacement) 90%)
Union coverage 25-40% ~10% private sector 60-70%
AI regulation EU law on artificial Permissive (market- Strong
intelligence driven)
Implications for divergence:
European Union:
- Labor protection → Level 3 kept longer (companies cannot fire easily)
- Result: Unemployment > wage collapse (protected but unproductive workers)
- Works councils → may reject aggressive adoption of artificial intelligence → slower
divergence
- Predicted Gini (2028): 0.48-0.52 (severe, but moderate compared to the US)
- P90/P10 ratio: 8-10× (compared to 15× in the US)
United States:
- Easy hiring/firing → Level 3 was quickly replaced
- Weak safety net → forces acceptance of low wages
- Minimum regulation → maximum speed of adoption
- Predicted Gini (2028): 0.55-0.58 (catastrophic)
- P90/P10 ratio: 12-15×
The Nordic model (Denmark, Sweden, Norway):
- Active labor market policies + high redistribution
- Strong investments in retraining (effectively and efficiently financed)
- Predicted Gini (2028): 0.40-0.45 (divergence contained by redistribution)
- P90/P10 ratio: 5-7×
Critical perspective: The institutional context determines whether divergence manifests itself in
the form of unemployment (EU/Nordic countries - workers protected but not employed) or in
the form of a collapse in wages (USA - workers employed at subsistence wages).

3.3.2. Eastern Europe: The Brain Drain Amplifier
Romania, Bulgaria, Poland, Czech Republic occupy a unique position:
Structural features:
- Lower base salaries (median for software developers in Romania: ~28,000-32,000 EUR
in 2025)
- Weaker institutions than in the Western EU (less union coverage, weaker labor
protections)
- High English proficiency among educated workers (legacy of post-communist educational
reform)
- EU membership → free labor mobility to Western Europe
Divergence dynamics:
Level 1 workers (IQ >110, fluent in English):
- Strong incentive for migration: EUR 28,000 (Romania) vs. EUR 80-120,000
(Germany/Netherlands) vs. USD 150-200,000 (USA)
- Accelerating Migration 2026-2028: As artificial intelligence makes location less relevant
(remote work is viable), wage arbitrage becomes the main factor.
- Forecasted result: 40-60% of Romanian workers in category 1 will migrate to the western
EU or the US by 2028
Level 2 workers:
- Stuck between: Cannot compete with Western EU states, tier 2 (wage compression),
cannot migrate (language/accreditation barriers)
- Predicted outcome: Stagnation of wages at EUR 24-30,000, high job insecurity
Level 3 workers:
- Double displacement: automation through artificial intelligence + loss of competition with
cheaper labor from Ukraine, Moldova, Western Balkans (if EU integration progresses)
- Predicted result: 50-70% exit from the knowledge sector → manual labor, services
(median 16-22 thousand EUR)
National Consequence: "Systemic Hollowing Out"
- Level 1 exits (brain drain)
- Level 3 remains (cannot afford migration, language barriers)
- Level 2 compressed
- Estimated Gini for Romanian knowledge workers (2028): 0.52-0.58 (comparable to
the US, despite EU membership)
The policy challenge: EU cohesion funds and structural investments are designed to reduce the
East-West gap, but the AI-driven brain drain is accelerating the gap. Romania’s most
productive workers (those who benefit most from AI) have the strongest incentive to leave.
Similar patterns expected: Poland (especially the tech sector in Warsaw), Bulgaria, Baltic
states. The Czech Republic is partially isolated (stronger industrial base, higher basic wages).
3.3.3. Projections regarding the distribution of salaries (2027-2028)
Table 3.2: Projected salaries of knowledge workers, by region and level
Level EU (West) EU (East/Romania) US North
Level 1 (P90) 138,000 EUR EUR 110,000* $195,000 118,000 EUR
Level 2 (Median) 45,000 EUR 27,000 EUR $60,000 51,000 EUR
Level 3 (P10) EUR 24,000** EUR 18,000** $25,000 EUR 30,000**
P90/P10 ratio 5.8× 6.1× 7.8× 3.9×
Gini coefficient 0.49 0.54 0.57 0.42

*Many  Romanian  workers  in  category  1  migrate;  those  who  remain  earn  less  than  their
counterparts in the Western EU
**Includes social transfers (unemployment, housing assistance)
Key observations:
1.  Absolute wage divergence is universal (all regions show increasing gaps)
2.  The Nordic model is moderated by redistribution (Gini 0.42 vs. 0.57 USA)
3.  Eastern Europe faces a double challenge: lower absolute wages + brain drain
4.  Safety nets matter: the EU/Nordic Tier 3 ceiling is higher (social transfers) than the US,
despite similar market salaries.

3.4. Testable Predictions (2027-2028)

3.4.1. Basic predictions (condensed)

Table 3.3: Twelve falsifiable predictions
#  Prediction  Baseline  Predicted in  Falsification  Data source
|     | 2025  | 2028  | threshold  |     |
| --- | ----- | ----- | ---------- | --- |
|     |       |       |            |     |
P1  Salary ratio P90/P10
(knowledge workers)

| United States  | 3.8×  | 12-15×  | <6.0×  | BLS OES  |
| -------------- | ----- | ------- | ------ | -------- |

| European Union  | 3.2×  | 8-10×  | <5.0×  | Eurostat  |
| --------------- | ----- | ------ | ------ | --------- |
|                 |       |        |        |           |
P2  Gini coefficient
(knowledge workers)

| United States  | 0.31  | 0.55-0.58  | <0.40  | BLS CPS  |
| -------------- | ----- | ---------- | ------ | -------- |

| European Union  | 0.29  | 0.48-0.52  | <0.38  | Eurostat  |
| --------------- | ----- | ---------- | ------ | --------- |
|                 |       |            |        |           |
P3  Median variation in
real wage

| United States  | $78,000  | $55,000- | >$73,000  | BLS CPS  |
| -------------- | -------- | -------- | --------- | -------- |
$62,000 (-18
to -27%)

| European Union  | EUR     | EUR 42,000-     | >50,000 EUR  | Eurostat  |
| --------------- | ------- | --------------- | ------------ | --------- |
|                 | 52,000  | 48,000 (-8 to - |              |           |
19%)
| P4  Bimodal wage  | p>0.10  | p<0.01  | p>0.05  | Hartigan 's     |
| ----------------- | ------- | ------- | ------- | --------------- |
| distribution      |         |         |         | immersion test  |
|                   |         |         |         |                 |
P5  Unemployment after
education (bachelor's
degree)

| United States  | 3.5%  | 9.5-12.0%  | <6.0%  | BLS Table A-4  |
| -------------- | ----- | ---------- | ------ | -------------- |

| European Union  | 4.2%  | 10.5-13.5%  | <7.0%  | Eurostat  |
| --------------- | ----- | ----------- | ------ | --------- |
|                 |       |             |        |           |
P6  Code quality
divergence
(errors/100 LOC)

| Junior developers  | 2.3  | 3.5-4.5  | <3.0  | GitHub  |
| ------------------ | ---- | -------- | ----- | ------- |
analysis

| Senior developers  | 0.4  | 0.2-0.3  | >0.35  | GitHub  |
| ------------------ | ---- | -------- | ------ | ------- |
analysis
|         |        |         |        |     |
| ------- | ------ | ------- | ------ | --- |
| Report  | 5.75×  | 12-20×  | <7.0×  |     |

|     |     |     |     |     |
| --- | --- | --- | --- | --- |
P7  AI tool abandonment
rate

| Low-skilled users  | -   | 35-50%  | <20%  | GitHub Copilot  |
| ------------------ | --- | ------- | ----- | --------------- |
Data

| High-skilled users  | -   | 5-12%  | >25%  | Stack Survey  |
| ------------------- | --- | ------ | ----- | ------------- |
Overflow
|     |     |     |     |     |
| --- | --- | --- | --- | --- |
P8  Blind acceptance rate

| Low qualification  | 78%  | 82-88%  | <75%  | IDE Telemetry  |
| ------------------ | ---- | ------- | ----- | -------------- |

| Highly qualified  | 42%  | 30-40%  | >50%  | IDE Telemetry  |
| ----------------- | ---- | ------- | ----- | -------------- |
P9  Self-reporting of skill  42%  58-68%  <45%  Stack Survey
| degradation (heavy  |     |     |     | Overflow  |
| ------------------- | --- | --- | --- | --------- |
users)
P10  Productivity variance  0.35-0.45  0.75-1.0  <0.55  Corporate
| (within firms, CV)  |     |     |     | human  |
| ------------------- | --- | --- | --- | ------ |
resources data
P11  Revenue per  $400- $2.0-4.0 million  <$1.0 million  10-K filings
employee (AI-based  500,000
companies)
| P12  Market concentration  | 42%  | 62-72%  | <50%  | Industrial  |
| -------------------------- | ---- | ------- | ----- | ----------- |
| (profit share of top 10    |      |         |       | reports     |
companies)

Table 3.4: Early Signals - McKinsey, "The State of AI in 2025" (November 2025)

| Metric  | Prediction   | McKinsey findings   | Condition  |     |
| ------- | ------------ | ------------------- | ---------- | --- |
|         | (2027-2028)  | (November 2025)     |            |     |
Productivity gap  71× (simulated)  High performers are 3  Directional confirmed
| (best performers vs.  |     | times more likely to  |     |     |
| --------------------- | --- | --------------------- | --- | --- |
| others)               |     | scale;                |     |     |
gap in EBIT impact (6%
vs 94%)
Workforce reduction  Unemployment of  30% expect a decrease  Acceleration
|     | 9.5-12% (P5)  | in the workforce in  | confirmed  |     |
| --- | ------------- | -------------------- | ---------- | --- |
2025-2026 (compared
to 17% in 2024)
Bimodal distribution  Predicted until  6% "high performance"  Already observable
|     | 2028 (P4)  | vs 94% others; 33%  |     |     |
| --- | ---------- | ------------------- | --- | --- |
scaling vs 67% stuck in
pilots
Three-tier structure  Level 1 (20-25%),  Expected 13% increase,  Matching proportions
|     | Level 2 (15-30%),  | 43% no change, 32%  |     |     |
| --- | ------------------ | ------------------- | --- | --- |
|     | Level 3 (50-60%)   | decrease            |     |     |

Interpretation: The divergence occurs before the stipulated deadline.
The categorical separation between “high performers” (6%) and others (94%) occurred within 36
months of ChatGPT launch (November 2022 - November 2025), not within the modeled 48
months. This suggests that our simulation may be conservative - the actual divergence could be
more severe and faster than modeled.

3.4.2. Falsification criteria
Strong falsification: ≥4 out of 12 predictions fail → the theory is fundamentally wrong and the
framework is abandoned.
Weak falsification: 2-3 predictions fail → theory is directionally correct, but overestimates
magnitude or misses moderating factors, parameters are revised.
Confirmation: ≥10 predictions are confirmed → theory is provisionally accepted; focus shifts to
intervention (Chapter 4).
3.4.3. Chronology of empirical tests
Phase 1: Early Signals (Q2-Q4 2026)
- Unemployment by education (P5): BLS/ Eurostat quarterly data
- Code Quality Indicators (P6): Academic Studies GitHub
- Artificial Intelligence Use Cases (P7, P8): Enterprise-level analysis
Critical Early Indicator: If unemployment among undergraduates shows no increase by the end
of 2026, the theory is likely falsified.
Phase 2: Mid-term evaluation (Q2-Q4 2027)
- Wage data (P1, P3): BLS/ Eurostat annual releases
- Revenue/employee of the firm (P11): 10-K Statements
Critical medium-term indicator: If P90/P10 remains <5.0× (US) or <4.0× (EU) by mid-2027,
divergence will not occur at the projected pace.
Phase 3: Full Test (Q4 2028)
- Full salary distributions (P2, P4): Annual data + academic analysis
- Multiannual longitudinal follow-up (P9, P10)
Final Verdict: Strong/Weak Falsification vs. Confirmation.
3.4.4. Geographical area-specific predictions
Romania/Eastern Europe (additional predictions):
P13: Accelerating the “ brain drain”
- Baseline (2025): ~15% of Romanian computer science graduates migrate within 5 years
- Forecast (2028): 35-50% of Tier 1 workers (IQ >110, fluent in English) migrate
- Counterfeiting: <25%
- Data source: National Institute of Statistics of Romania, migration data
P14: Widening wage gap (East-West EU)
- Reference level (2025): Romanian median development = 55% of the German median
(€32,000 compared to €58,000)
- Forecast (2028): Romanian median income = 35-40% of the German median (€27,000
vs. €72,000)
- Counterfeiting: >50%
- Data source: Eurostat, national statistics
Preliminary evidence from academic publications suggests accelerating divergence in
academia as well. Analysis of academic research shows explosive growth: from ~50 publications
in 2022 to 1,100 in 2024 (Springer, 2025), with an annual growth rate of 545%. However, this
growth is concentrated in higher-tier institutions, while mid-tier institutions show minimal
representation (Springer, 2025). Nature (2024) reported that computational biologists use
ChatGPT to reduce manuscript editing time, allowing more time for research - a productivity
multiplier available only to those with expertise in the field to critically evaluate AI results.
Systematic data on publication rates by researcher seniority (senior vs. junior faculty) remain
unavailable at the end of 2025.

3.5. Alternative hypotheses and discriminant tests
3.5.1. "The equalization effect persists in the long term"
Claim: AI continues to be more beneficial for low-skilled workers, even over a period of more
than 24 months.
Discriminant predictions:
- P90/P10 should decrease: <3.0× by 2028
- Average salary should increase: >55,000 EUR (EU), >82,000 USD (US)
Tests: P1, P3 discriminate directly. If both fail (ratio decreases, median increases), the equalizer
hypothesis is supported against the divergence theory.
3.5.2. "Political intervention successfully moderates"
Claim: EU law on artificial intelligence, UBI pilots and retraining programs prevent extreme
divergences.
Discriminant predictions:
- Gini increases but remains <0.45 (EU)
- Moderate unemployment, but income threshold maintained (transfers)
- P6 (code quality) shows low improvement in skills
Tests: If P2 (Gini <0.45) AND P5 (moderate unemployment) AND P9 (no increase in skill
degradation), the hypothesis of successful intervention is supported.
3.6. Conclusion: Economic stratification as an implicit trajectory
The projections establish:
1. The three-tier structure appears universally (EU, US, Eastern Europe), differing only
in severity and manifestation (unemployment vs. wage collapse).
2. Geographic variation is significant: the Nordic pattern is moderated (Gini 0.42) by
redistribution; the US faces extreme divergence (Gini 0.57); Eastern Europe faces an
amplifier of brain drain.
3. Romania/Central and Eastern Europe: Level 1 migrates west, Level 3 remains, and the
productivity gap within the country deepens, despite EU membership.
4. Twelve falsifiable predictions define empirical benchmarks for the period 2027-2028.
By Q4 2028, we will know whether the theory accurately describes reality.
The structure described is not inevitable - but it is the default trajectory in the absence of
aggressive intervention. The next chapter examines whether intervention can moderate
divergence and at what cost.
CHAPTER 4: DISCUSSIONS AND INTERVENTIONS
4.1. Introduction: Scope, limits and intervention question
The preceding analysis has built a cumulative empirical case: literature synthesis and behavioral
data (Chapter 1), mathematical modeling and simulation (Chapter 2), economic projections and
falsifiable predictions (Chapter 3). This chapter takes a step back to critically examine the scope
and limitations of the theory (§4.2), evaluate intervention strategies with a focus on the Cognitive
Stimulation Mechanism (§4.3), outline the adaptive MCS 2.0 framework (§4.4), and provide
methodological transparency regarding the AI collaboration in producing this work (§4.5).
Scientific honesty requires recognizing what a theory cannot explain as clearly as what it can.
Cognitive Divergence Theory deliberately focuses on a specific mechanism (cognitive ability
as a determinant of AI productivity) and a specific population (knowledge workers in developed
economies). This chapter clarifies critical limits before examining whether intervention can
moderate catastrophic outcomes.

4.2. Seven critical limitations
Limitation 1: Cognitive capacity indicators are imperfect
IQ measures pattern recognition, verbal reasoning, and working memory, but not metacognitive
ability (awareness of one’s own thought processes), domain expertise (in-depth knowledge in
specific domains), or personality traits (conscientiousness, openness). Two people with identical
IQs can have drastically different outcomes in terms of productivity using AI if one of them has
superior metacognition or domain expertise. Despite its imperfections, IQ remains the best
available indicator: it is measurable, predictive (r = 0.5-0.7 with complex task performance), and
available through educational/occupational indicators. Future research should develop
assessments of “AI-partnership readiness” that directly measure relevant faculties (prompt
engineering skills, ability to evaluate results, iterative refinement).
Limitation 2: Assumption of static capacity of artificial intelligence
The model assumes that AI remains constant at roughly the equivalent of an IQ of 120 by 2028.
In reality, AI could improve rapidly (reaching an IQ of 140+, replacing even highly skilled workers),
stabilize (reaching diminishing returns), or specialize (different models dominating different
domains). If AI improves to an IQ of 150+ by 2027, even highly skilled workers Level 1 might
experience cognitive atrophy. Conversely, if capacity stagnates, the divergence may be less
severe, as low-skilled workers adapt over longer periods. The static assumption is conservative
and testable - sensitivity analysis (Chapter 2) shows that the qualitative results hold even with a
±25% variation in capacity.
Limitation 3: Assumption of homogeneous adoption of AI
The model assumes that all users adopt identical AI tools. In reality, there is inequality in access:
high-skilled workers can use state-of-the-art proprietary models (GPT-5, Claude Opus 4), low-
skilled workers can use free, weaker models (GPT-3.5). Organizational variation (some firms
mandate AI, others prohibit it) and geographic variation (EU AI law vs. US permissiveness) create
complex adoption landscapes. Inequality in access could amplify divergence (Level 1 gets better
tools) or moderate it (mandatory equal access). A testable prediction: within firms that provide
universal access to identical tools, divergence should still emerge through the cognitive capacity
of the user. If divergence only emerges when access to tools differs, the mechanism is economic
(access to capital) rather than cognitive.
Limitation 4: Labor market feedback was not modeled
The simulation models individual productivity, but not market adjustments: wage equalization
(why would firms hire workers with low productivity?), task reallocation (reassigning workers to
non-AI tasks), or sectoral shifts (knowledge workers who have been replaced and move to
healthcare, education, trade). Findings of “negative productivity” may overestimate real-world
outcomes if markets efficiently reallocate workers, or underestimate them if mass unemployment
triggers demand collapse and political instability. Integrating the simulation with macroeconomic
models (DSGE, agent-based) is essential for future work, but requires dozens of assumptions
about elasticities, adjustment speeds, and institutional rigidities.
Limitation 5: Works without knowledge are excluded
The model focuses exclusively on knowledge workers (30-35% of the US workforce, 15-20%
globally). It does not address manual labor (construction, manufacturing, logistics), personal
services (healthcare, education, childcare), or creative labor (arts, entertainment). The “50-60%

displaced” projection applies only to knowledge workers. Impact on the entire economy: If 55%
of knowledge workers are displaced, this translates to ~18% of the total US workforce, ~9%
globally. These figures assume no spillover to non-knowledge sectors; if displaced knowledge
workers flood manual labor markets, wages in these sectors could collapse, creating indirect
displacement.
Limitation 6: Simplified Cultural and Institutional Variation
The model implicitly assumes Western (especially US) labor market institutions: discretionary
employment, weak safety nets, individualistic culture. Alternative contexts produce different
manifestations: strong labor protections in Europe may generate unemployment (Tier 3
protection, but unproductive) rather than wage collapse; lifetime employment norms in East Asia
may cause firms to absorb low-productivity workers; high taxation and redistribution in
Scandinavia may compensate for divergence (Tier 3 income is supported by transfers) without
preventing underlying productivity gaps. Economic divergence (productivity gaps) may be
universal, but social outcomes (income inequality, employment) vary by institutional context - as
the geographical analysis in Chapter 3 demonstrates.
Limitation 7: Endogenous technology development was not modeled
The model treats AI capability as exogenous (provided by technology developers). In reality, AI
development responds to market signals. If Level 1 demand dominates (highly skilled users pay
premium prices), developers optimize for complex reasoning and strategic planning, further
widening the gap. If Level 3 demand dominates (large user base despite low willingness to pay),
developers optimize for simplicity and reliability, narrowing the gap. The current trajectory (2024-
2025) favors Level 1: OpenAI ChatGPT Enterprise, Anthropic Claude for Work, GitHub Copilot
Business - all premium- priced, feature-rich products. Level 1 generates 7x more revenue, despite
1/3 the number of users, creating self-reinforcing feedback loops that favor sophistication over
simplicity.
4.3. Intervention Analysis: MCS Case Study
4.3.1. Context: Minimum Ethical Governance (GEM)
The Cognitive Stimulation Mechanism (CSM) is Article 3 of the Minimum Ethical Governance
(MEG) framework - a proposed governance standard for conversational AI systems, developed
to address the risks of cognitive atrophy, loss of critical thinking, and degradation of human skills
through the overuse of AI (Stan, 2025).
The MEG architecture comprises four main articles:
Article 1: Audit and transparency
All AI responses are recorded with metadata (task complexity, user interaction patterns, MCS
activation), stored in an immutable audit trail. Users and regulators can verify AI-based decision-
making.
Article 2: Consent and control
Users must explicitly consent to the assistance provided by artificial intelligence, understand
when artificial intelligence intervenes (MCS activation), and can revoke consent or disable
interventions at any time.
Article 3: Cognitive Stimulation Mechanism (CSM)
AI dynamically adjusts complexity to keep the user engaged. Instead of always providing direct
answers, AI can enable “force functions”: asking for clarification, challenging the user to evaluate
alternatives, or asking for synthesis before offering solutions.

Article 4: Quality assurance
AI must verify that the results meet minimum quality thresholds before delivery, preventing the
generation of content that the user cannot properly evaluate or integrate. MEG is positioned as a
micro-standard - not as a comprehensive regulation (like the EU AI Law), but as a protocol
focused on conversational AI to prevent specific harms. It can be implemented voluntarily by AI
companies (Anthropic, OpenAI, Gemini, etc.) or imposed through sector-specific regulations
(e.g., enterprise AI tools, educational platforms, professional services).
This article focuses exclusively on Article 3 (MCS) as a proof-of-concept intervention for
cognitive divergence. We test whether MCS, as currently specified in MEG 1.0, can moderate the
71× productivity gap predicted in Chapter 2. The analysis reveals both the effectiveness of MCS
(80% gap reduction for users with IQs of 85–100) and its fundamental limitation (universal
design), motivating the proposal for MCS 2.0.
4.3.2. Cognitive stimulation mechanism (MCS 1.0)
The Cognitive Stimulation Mechanism, specified in the Minimum Ethical Governance (MEG)
framework (Stan, 2025), operates on a fundamental principle: AI should not always provide
the easiest answer, but the answer that best promotes the user's cognitive engagement.
Three activation zones based on the AI effort metric (Tg):
Zone 0 (simple tasks): Direct response
Zone 1 (moderate complexity): Clarification - AI aims to refine the user's thinking
Zone 2 (high complexity): Synthesis/Challenge - AI challenges the user to co -create a solution
The user sensitivity parameter (μS, range 0.5-2.0) allows adjustment of the activation frequency.
All activations are recorded in an immutable audit trail.
4.3.3. MCS simulation results
Rerunning the Monte Carlo simulation from Chapter 2 with MCS enabled reveals dramatic but
heterogeneous effects:
Table 4.1: Aggregate impact of MCS (t=24 months)
Metric Standard AI AI + MCS Change
Average productivity 1,515 1,641 +8.4%
Gini coefficient 0.551 0.600 +8.9%
Top 25% / Giants 50% 70.66× 13.92× -80.3%
% Negative productivity 13.5% 15.1% +1.6 percentage points
Key findings: MCS reduces the productivity gap from 71× to 14× - an 80% reduction that turns
“categorical elimination” into “survival marginalization”. However, the Gini coefficient increases
slightly (MCS differentially benefits mid-range users, squeezing those in the middle while leaving
the extremes unchanged), and negative productivity increases (MCS overwhelms the lowest-
ability users).
4.3.4. Heterogeneous effects according to cognitive quartile
Table 4.2: Impact of MCS by IQ quartile
IQ quartile Standard AI AI + MCS Change % Negative (MCS)
The last 25% (<90) -0.002 -0.045 Worse 59.8%
T2 (90-100) 0.127 0.679 +436% 12.3%
T3 (100-110) 1,543 1,543 0% 2.1%
Top 25% (110+) 4,316 4,316 0% 0.0%

Critical observations:
Last 25%: MCS is actively harmful - productivity continues to decline, 60% produce negative
results. Mechanism: MCS Level 2 (synthesis/challenge) is too cognitively demanding. Users
cannot formulate coherent answers to the questions "What trade-offs do you see?", waste time,
generate confusion.
Q2 (IQ 90-100): MCS is transformative - productivity increases by +436%. Only 12% produce
negative results (vs. 31% without MCS). Mechanism: MCS prevents “cognitive hibernation”.
Forced to engage critically, these users retain their skills, detect errors, and produce higher quality
results. This is the sweet spot: MCS offers massive gains for ~30% of the population.
T3 and top 25%: MCS is redundant - zero productivity change. These users are already engaging
critically, evaluating results, iterating on solutions. MCS adds difficulty (slows them down by
asking questions they've already thought about), but doesn't improve results.
Interpreting the 14× gap: With MCS, the gap is survivable - the bottom 50% produce significant,
if modest, value. Q2 productivity (0.679) is 68% of the baseline, 16% of the top quartile. Economic
analogy: top quartile earns €145,000, Q2 earns €33,000 (poverty threshold, but not deprivation).
Workers can stay employed, earn living wages, maintain their dignity. MCS does not eliminate
inequality (14× is still extreme), but it prevents catastrophic collapse.
4.3.5. Fundamental limitation: Universal solution
MCS 1.0 uses a single threshold formula for all users. User sensitivity (μS) is user-configurable,
but represents preference (how much involvement do I want?), not capacity (what can I handle?).
Result: same MCS Level 2 challenge is presented to users with IQ 80 (overwhelmed, frustrated),
users with IQ 95 (struggling but succeeding, learning), and users with IQ 120 (already considering
this, bothered by redundancy). One protocol, three outcomes: harmful, beneficial, irritating.
MCS is also unidirectional: AI analyzes the difficulty of the task, decides the MCS level, the user
responds - but AI never learns whether MCS helped or hurt this specific user. There is no
adaptation mechanism. If users consistently face Level 2 challenges, AI should reduce the
difficulty. If users consistently excel, AI should increase the difficulty or reduce the frequency.
MCS 1.0 is "fire and forget" - it intervenes, but does not observe the results.
4.4. The way forward: MCS 2.0 (Adaptive Pyramid Framework)
4.4.1. The fundamental problem: Unidirectionality
MCS 1.0 is unidirectional:
AI analyzes the difficulty of the task (Tg)
↓
AI decides MCS level (Zone 0/1/2)
↓
The user responds (or not)
↓
[END - no information to the AI about whether MCS helped or harmed]
What's missing: The AI never learns whether the MCS intervention was appropriate for THIS
specific user on THIS specific task. If the user is struggling in Zone 2, the AI should detect
confusion and simplify. If the user consistently excels, the AI should detect mastery and increase
difficulty or reduce difficulty. If the user ignores MCS prompts, the AI should adjust strategy.
MCS 1.0 is "trigger and forget" - it intervenes solely based on the complexity of the task, without
observing the quality of the user's response or adapting to individual ability.

4.4.2. Core Insight: Bidirectional feedback through trajectory tracking
MCS 2.0 introduces bidirectionality:
AI analyzes the difficulty of the task (Tg)
↓
AI proposes MCS level based on Tg AND user historical trajectory
↓
User responds
↓
AI analyzes the quality of the response → updates the user trajectory estimate
↓
[FEEDBACK LOOP: The next interaction uses the updated trajectory for the MCS decision]
Instead of estimating the user's IQ (complex, privacy-impairing, potentially discriminatory), the
user's conversation trajectory is tracked: Does Turn(n) show a cognitive ascent relative to
Turn(n-1)? (User engages, learns) Does it show a descent? (Confused, overwhelmed) Does it
repeat without progress? (Stuck)
Trajectory detection is privacy-aware (no IQ labeling), immediate (detectable after 2-3 turns vs.
10+ for IQ convergence), and natural (mimics how human tutors adapt based on student
response).
4.4.3. Four types of trajectories
ASCENDING: Complexity increases from one turn to the next
Signals: Turn(n) references concepts from Turn(n-1), asks deeper questions, vocabulary /
specificity increases
MCS Action: Increase ambition, activate Level 2 earlier, introduce advanced concepts
DESCENDING: Complexity decreases or user seeks simplification
Signals: Turn(n) asks about a simpler concept, "What does X mean?" where X was in Turn(n-1),
expressions like "I don't understand".
MCS Action: Disable Level 2, enable Level 1 (clarification), provide concrete examples
LATERAL: The subject changes without a clear change in depth.
Signals: Turn(n) different topic, similar complexity, exploratory rather than confusing
MCS Action: Neutral (maintains current MCS level), allows exploration
STAGNANT: Repetition without progress
Signals: Tower(n) rephrases Tower(n-1) without new information, same question asked 2-3 times
MCS Action: Interrupt and offer an explicit choice (“Would you prefer A or B?”), suggest a pivot
4.4.4. Adjusting the dynamic threshold
Reference thresholds (from MCS 1.0):
threshold_base_L1 = 10 × Tg_base / μS
threshold_L2_base = 30 × Tg_base / μS
Trajectory-based adaptation:
if ASCENDENT: adaptation_factor = 0.7 (activate MCS earlier)
if DESCENDENT: adaptation_factor = 1.8 (reduces MCS frequency)
if LATERAL: adaptation_factor = 1.0 (no change)
if STAGNANT: adaptation_factor = 0.5 (forces MCS to stop the model)

Adjusted thresholds:
threshold_L1 = threshold_L1_base × adaptation_factor
threshold_L2 = threshold_L2_base × adaptation_factor
Extended MCS levels:
o Level 0: Direct response
o Level 1: Clarification
o Level 2-Easy: Easy guidance (for descending trajectory) - "I see two options: A (simpler)
or B (more complete). Which one fits better?"
o Level 2: Standard synthesis
o Level 2 - challenge: Advanced Synthesis (for the upward trajectory) - "Good thought.
Now think: what second-order effects could this create?"
4.4.5. Expected benefits and open questions
Expected benefits:
1. Prevents overload for low-capacity users (detects descent, switches to Level 2-light)
2. Maximizes difficulty for high-capacity users (detects ascent, escalates to 2 challenges)
3. Fast adaptation (2-3 turns vs. 10+ for IQ estimation)
4. Maintaining confidentiality (no IQ labeling, just "conversation going well/poorly")
Open research questions:
1. Can trajectory detection accurately reflect cognitive ability? (Test: Recruit N=500 with
known IQ, correlate trajectory scores with IQ and performance)
2. Does MCS 2.0 perform better than MCS 1.0? (Test: RCT with N=2,000, Group A: Static
MCS, Group B: Adaptive MCS)
3. What are the optimal parameters? (Test: Optimization based on detection thresholds,
adaptation speed, level granularity)
4. Do users accept adaptive MCS or perceive it as an “AI judgment”? (Test: Qualitative
interviews + large-scale survey on transparency and control preferences)
Draft MCS 2.0 specifications (future work): The full specification would include the trajectory
detection algorithm (formal definitions of metrics, composite score formula), adaptive threshold
calculation (mapping trajectory score to adaptation factor), extended MCS levels (precise
requests for 2-lite and 2-challenge), audit and transparency requirements (trajectory_score
logging, user-oriented explanations), and ethical safeguards (no IQ labeling, user override
capability, opt-out clause).
This specification is not provided here (it would require an additional 20-30 pages). MCS 2.0
is presented as a roadmap for research and development, not as a complete solution. Critical
point: even perfected, MCS 2.0 cannot eliminate divergence - it turns categorical elimination into
a surviving marginalization (gap of “only” 8-12×, instead of 71×), but severe inequality remains.
4.4.6. From MCS 2.0 to MEG 2.0: Implications for the framework
MCS 2.0 is not just a technical update to Article 3 of the MEG - it represents a fundamental
reconceptualization that requires updates across the entire MEG framework.
Current assumptions of MEG 1.0:
- User static capacity (preferred parameter μS)
- One-way intervention (AI decides, user responds, no feedback)
- Universal thresholds
MEG 2.0 requirements:
Article 1 (Audit) must record:
- trajectory_score (per turn)

- adaptation_factor (how the thresholds were adjusted)
- mcs_level_selected (0, 1, 2-easy, 2, 2-challenge)
- user_response_quality (for calculating the trajectory of the next turn)
Article 2 (Consent) must inform users:
- "AI learns your cognitive trajectory through conversation patterns"
- "You can request the current trajectory estimate at any time"
- "You can disable trajectory tracking (reverts to MCS 1.0 static thresholds)"
Article 3 (MCS) = MCS 2.0 specification (this section)
Article 4 (Quality) must be adaptable to the trajectory:
Quality thresholds cannot be universal (same standard for all users)
- Lower quality threshold for DESCENDENT users (simpler but correct results are accepted)
- Higher quality threshold for ASCENDENT users (requires more in-depth analysis)
The full MEG 2.0 specification is beyond the scope of this article, but is identified as essential for
future work. MCS 2.0 cannot be implemented in isolation - it requires consistent updates across
all four MEG articles to maintain governance integrity.
4.5. Methodological transparency: AI as co-author
This article is self-referential in a productive sense: it analyzes the cognitive stratification induced
by artificial intelligence, while demonstrating the symbiotic partnership that defines the
emerging cognitive elite.
Human contribution (Adrian Stan):
1) Theoretical intuition (cognitive divergence as a mechanism, the concept of Homo
Symbioticus)
2) Empirical observations (GitHub / Stack behavioral models Overflow identified through
community participation)
3) Conceptual frameworks (MCS pyramidal adaptation, dynamics of the Eastern European
brain drain)
4) Critical evaluation (identifying weak arguments, requesting revisions, challenging
assumptions)
5) Strategic direction (what questions to pursue, which to postpone, maintaining theoretical
coherence)
6) Geographical contextualization (European institutional analysis, specific projections for
Romania)
AI contribution (Claude, Anthropic):
1) Literature synthesis (integrating over 40 studies in the fields of psychology, economics
and HCI)
2) Mathematical formalization (translating intuitions into testable formulas, calibrating
parameters)
3) Carlo simulation design (10,000 agent model, temporal dynamics, sensitivity analysis)
4) Structural coherence (organizing the analysis in logical progression, identifying
redundancies)
5) Generating alternative hypotheses (challenging the author's assumptions, proposing
falsification criteria)
6) Quantitative accuracy (generating tables, performing calculations, ensuring internal
consistency)
The collaborative process: Over three weeks, through approximately 200 rounds of
conversation, we engaged in an iterative refinement. We proposed frameworks (cognitive
divergence, MCS adaptation); Claude formalized them mathematically. Claude generated
analyses (simulation results, literature synthesis); we critiqued and redirected (“this misses

institutional variation”, “add European context”, “and Romania?”). We solicited evidence; Claude
synthesized the literature and empirical models. Claude identified gaps (“predictions need
numerical thresholds”); we provided domain knowledge and strategic priorities.
Critical point: This partnership places us squarely in the “Homo Symbioticus ” category theorized
in Chapter 3. We possess the metacognitive capacity to:
- Critically evaluate Claude's results (detect logical gaps, exaggerated claims, US-centric
biases)
- Integrate content generated by artificial intelligence into a coherent whole (not blind
acceptance)
- Iterate on weak sections (request review with specific feedback)
- Strategic collaboration leadership (focus on falsifiable predictions, avoid speculative futurism)
- Recognition of limits (when Claude does not have data about the Romanian labor market,
completing it with his own knowledge)
A researcher lacking these capabilities would either produce (a) a superficial summary by blindly
accepting AI, or (b) an abandoned project (overwhelmed by complexity). This asymmetry - our
ability to collaborate productively with AI versus someone else's inability - is precisely the
mechanism that drives the cognitive divergence analyzed in this article.
The article thus demonstrates its own thesis: without Claude, producing such a profound analysis
would have required 6–12 months of full-time work (literature review, mathematical modeling,
simulation coding, sensitivity testing). With Claude, it emerged in three weeks of intensive but
time-limited collaboration. This 10÷20× productivity multiplier is available only to those with the
cognitive capacity to critically oversee AI results - the exact asymmetry that creates the 71× gap.
Ethical note: All contributions to artificial intelligence are explicitly acknowledged. Claude is
acknowledged as a methodological collaborator (not a co-author in the traditional sense, as he
cannot consent to authorship nor be held accountable for errors). Responsibility for errors,
omissions, interpretations, and conclusions lies solely with Adrian Stan. This work adheres
to emerging academic norms regarding the disclosure of information about artificial intelligence
(transparent acknowledgement, human accountability, critical oversight).
4.6. Broader policy recommendations (in brief)
At the organizational level: Mandatory inclusion of quality-adjusted productivity indicators in
performance evaluations (tracking error rates, rework cycles, not just production volume).
Requesting human involvement in high-stakes decisions (code generated by artificial intelligence
must pass human review before implementation).
Societal level: Educational reform towards metacognitive skills (critical evaluation, problem
decomposition, solution synthesis - 15–20-year time horizon). Income redistribution (UBI or
negative income tax financed by taxing profits from artificial intelligence to support the replaced
level 3). Investment in research (500 million - 1 billion euros for longitudinal studies, development
of MCS 2.0, randomized clinical trials (RCTs) of interventions).
EU specific: Leveraging the EU AI Law as a natural experiment (mandatory human oversight in
high-risk systems provides a testable comparison to market-driven adoption in the US).
Strengthening active labor market policies (investing in reskilling, albeit with realistic expectations
of cognitive capacity constraints). Addressing brain drain (migration from Eastern Europe, Tier 1,
creates a policy challenge for cohesion funds).
Critical limitation: For the bottom 15-25% of the population (IQ <85), technical interventions like
MCS are insufficient. This population requires either artificial intelligence tools designed for
complete automation (minimum human input) or a transition away from knowledge-based work,
supported by income redistribution and expanded safety nets.

4.7. Conclusion: Intervention is necessary, but insufficient
The theory has seven critical limitations: imperfect cognitive indicators, static AI assumption,
homogeneous adoption assumption, lack of labor market feedback, exclusive focus on
knowledge-based work, simplified institutional variation, and exogenous technology
development. Despite these, the central mechanism - cognitive ability driving AI productivity gains
- is robust and testable.
MCS 1.0 simulation demonstrates an 80% reduction in the gap (from 71× to 14×), but only for
users with IQs between 85 and 100. MCS 2.0 (adaptive pyramid framework) addresses the "one
size fits all" limitation through trajectory tracking and dynamic threshold adjustment, but remains
a research agenda that requires empirical validation.
Methodological transparency: This article exemplifies the phenomenon it describes. The depth
of the analysis resulted from the symbiotic human-AI partnership over three weeks - a productivity
multiplier available only to those with the metacognitive capacity to critically oversee AI outcomes.
Asymmetry in the ability to collaborate productively with AI is the mechanism driving the 71×
divergence.
MCS 2.0 is necessary but insufficient: even perfected, it reduces the gaps to 8-12× (still severe
inequality). Preventing catastrophic divergences requires simultaneous intervention at three
levels: technical (adaptive governance of artificial intelligence), organizational (quality indicators,
human oversight) and societal (education reform, income redistribution, labor market policy).
The action period is 2026-2028. By the end of 2028, the twelve empirical predictions will be
testable. If the theory is confirmed and interventions are not implemented, the divergence will
be entrenched and potentially irreversible.
CHAPTER 5: CONCLUSIONS
Cognitive Divergence Theory began with a simple observation: if cognitive ability determines
how effectively people use tools, and if AI is the most demanding cognitive tool ever created, then
the adoption of AI should not produce gradual inequality, but rather categorical stratification. Four
chapters of analysis transformed this observation into a comprehensive and testable framework.
The evidence is cumulative and converges towards a consistent pattern. Chapter 1 established
that intelligence predicts tool mastery, that complex tools amplify performance variance, and
documented behavioral patterns across real-world platforms - GitHub, Stack Overflow, Kaggle -
showing that low-skilled users overuse AI, blindly accept the results, and produce low-quality
work that requires extensive rework, while high-skilled users selectively use AI and achieve real
productivity gains. Chapter 2 formalized these observations in a mathematical model and
demonstrated through Monte Carlo simulation that within 24 months of AI adoption, the bottom
50% of the cognitive distribution experiences a collapse in productivity (the bottom quartile
effectively producing zero value), while the top 25% triple their output, producing a 71-fold gap -
not inequality, but economic elimination.
Chapter 2 also resolved the apparent contradiction with the “equalization effect” studies through
the dual-effect framework: AI temporarily equalizes performance on simple, well-defined tasks
over short time horizons (2–3 months), but permanently amplifies divergence on complex tasks
over longer horizons (12–24 months). The equalization studies measured what was
methodologically accessible within their constraints - they were not wrong, they were incomplete.
The long-term outcome is determined not by short-term equalization but by the dynamics of task
migration: as AI automates simple work, workers must move to complex work, where cognitive
ability becomes the mandatory constraint. High-IQ workers succeed in this migration; low-IQ
workers fail, creating a categorical separation.
Chapter 3 projected these dynamics onto economic structures, anticipating the emergence of a
stable three-tier economy by 2027-2028: Homo Symbioticus (20-25%, median salary 130-145k

EUR / 180-210k USD, productivity 3-4x the initial value), a precarious middle class (15-30%,
median salary 42-48k EUR / 55-65k USD, constantly threatened by automation), and a
disadvantaged majority (50-60%, median income 26-30k EUR / 32-35k USD, have left
knowledge-based work). Crucially, this structure manifests itself differently depending on the
institutional contexts. The European Union moderate divergence through labor protection and
generous safety nets (predicted Gini index 0.48-0.52, P90/P10 ratio 8-10×), while the United
States faces more severe outcomes (Gini index 0.55-0.58, P90/P10 ratio 12-15×). The Nordic
model demonstrates that redistribution can limit inequality (Gini index 0.40-0.45) without
preventing underlying productivity divergence.
Eastern Europe faces a unique challenge: the amplification of the “brain drain”. Romania,
Poland, Bulgaria and the Baltic states are facing a “subsistence drain” as level 1 workers (IQ
>110, fluent in English) migrate to Western Europe or the United States for 3-5 times higher wage
premiums (€28,000 Romania vs. €80-120,000 Germany/Netherlands and $150-200,000 US).
This migration is accelerated by artificial intelligence, which makes location less relevant for
knowledge-based work. The result is a concentration of low-productivity workers in Eastern
European countries, despite EU cohesion policies designed to reduce the gap between East and
West, creating a political paradox in which artificial intelligence simultaneously enables
development (productivity tools) and undermines it (talent extraction).
Chapter 3 specified twelve falsifiable predictions with numerical precision: wage ratios,
unemployment rates by education level, productivity variation within firms, code quality indicators,
behavioral patterns in AI use, and firm-level outcomes. These predictions define empirical
milestones for the period 2027–2028. By the end of 2028, we will know whether the theory
accurately describes reality. If four or more predictions fail, the theory is falsified. If ten or more
succeed, it is provisionally confirmed. The predictions distinguish between Cognitive Divergence
Theory and alternative hypotheses: persistent equalization effects, minimal impact of AI, or
successful policy moderation.
Chapter 4 examined whether intervention can prevent catastrophic divergences. The answer is
yes - but only partially. Simulating the Mechanism Cognitive Stimulation (MCS 2.0)
demonstrates an 80% reduction in the productivity gap (from 71× to 14×) by forcing users to
critically engage with AI outputs, rather than blindly accepting them. However, this intervention is
only effective for users with an IQ of 85–100, representing about 30% of the population. It actively
harms users below IQ 85 by overwhelming them with cognitive demands they cannot meet, and
provides no benefit to users above IQ 110 who are already self-regulating. This heterogeneity
exposes the fundamental limitation of the current design of MCS: it is unidirectional and static,
while it should be bidirectional and adaptive.
The solution presented is MCS 2.0 - an adaptive framework that tracks the user's cognitive
trajectory in real time by analyzing conversations from one round to the next, detecting whether
each round represents an ascent (building knowledge, asking deeper questions), a descent
(returning to fundamentals, expressing confusion), a lateral movement (exploring new topics), or
a stagnation (repetition without progress). AI adjusts the intensity of the intervention accordingly.
This approach respects privacy (no IQ labeling), is fast (2-3 rounds for detection), and mimics
how human tutors adapt naturally. MCS 2.0 is not a fully specified solution, but a research agenda
that requires empirical validation through experiments, user studies, and algorithmic optimization.
However, even when perfected, MCS 2.0 cannot eliminate divergence altogether. It can
transform categorical exclusion into survivable marginalization - a gap of 8–12× instead of 71×,
a Gini coefficient of 0.45 instead of 0.58 - but severe inequality remains. For the least-affected
15–25% of the population, with an IQ below 85, technical interventions are insufficient. This
population requires fundamentally different approaches: either artificial intelligence tools
designed for complete automation with minimal human input, or a transition from knowledge-

based work to manual labor and personal services, supported by income redistribution through
universal basic income, negative income taxes, or expanded safety nets.
Chapter 4 also provided methodological transparency: this article exemplifies the phenomenon it
describes. The depth of the analysis – literature synthesis from over 40 studies, mathematical
formalization, Monte Carlo simulation with 10,000 agents, institutional comparison in
EU/US/Northern Europe/Eastern Europe contexts – resulted from intensive collaboration with
Claude (Anthropic ’s AI assistant) over three weeks of iterative dialogue. This partnership
demonstrates the cognitive amplification available to those with the metacognitive capacity to
critically evaluate and integrate AI results – precisely the asymmetry that drives the 71-fold
divergence that this article analyzes. A researcher lacking this capacity would have either
produced a superficial summary by blindly accepting AI, or abandoned the project as
overwhelming. This is not speculation, but a lived experience of the Homo Symbioticus level.
The broader policy response must work simultaneously at three levels. Individual design of AI
systems must include adaptive cognitive stimulation, quality safety barriers that prevent the
generation of outputs that users cannot evaluate, and ways of developing skills that gradually
reduce the underlying structure as competence develops. Organizational policies must require
human involvement in high-stakes decisions, rebalance productivity metrics in favor of quality
over quantity, and require employees to maintain their independent capacity through periodic
work without AI. Societal interventions should include regulations that impose incentive protocols
in consumer AI products (the EU Law on Artificial Intelligence provides a natural experiment),
educational reforms that shift school curricula towards metacognitive skills (15–20-year time
horizon), massive investments in retraining, with realistic expectations regarding cognitive
capacity constraints, and income redistribution to prevent poverty among displaced people.
The timeline is unforgiving. Empirical predictions specify observable outcomes by the end of
2028. If the theory is confirmed and interventions have not been implemented by then, the
divergence will be entrenched. Workers who have spent 24-36 months in workflows heavily
dependent on AI will have experienced enough cognitive atrophy that reversal becomes difficult
or impossible. Firms will have optimized processes with respect to extreme productivity variance.
Labor markets will have adjusted wage structures to reflect the new reality. Political coalitions will
have consolidated around distributive conflicts. The intervention period is 2026-2028 - precisely
the period in which empirical data will test the validity of the theory.
This creates both urgency and opportunity. Early signals - patterns of unemployment by education
level, divergence in code quality in GitHub data, behavioral patterns in AI usage - will be
observable by the end of 2026. If these early indicators confirm the theory, intervention should
begin immediately, even as longitudinal studies continue. The research agenda and intervention
development can continue in parallel, with adaptive refinement as evidence accumulates. We
don’t have to wait for perfect data before acting, but neither can we justify a massive intervention
without empirical validation.
The stakes extend beyond economics. A society with a Gini coefficient of 0.58 and a 15-fold
income gap has no modern democratic precedent. Historical periods with comparable inequality
- Gilded Age America, pre-revolutionary France, apartheid South Africa - were characterized by
political instability, social unrest, and, in some cases, violent unrest. The emergence of a cognitive
aristocracy controlling 3-4x productivity through AI symbiosis, geographically segregated by a
displaced majority surviving on transfer payments or subsistence wages, poses profound
challenges to social cohesion and democratic legitimacy.
Yet the outcome is not predetermined. Technology enables divergence, but politics
determines whether it occurs. The successful future of AI is not one that surpasses humanity,
but one that learns to walk alongside us, at our own pace, adapted to our varying capacities. This
requires abandoning techno-utopian fantasies of universal superintelligence and accepting the

harsher truth: human cognitive diversity is real, largely immutable in adulthood, and must be
accommodated, not ignored. Different users need different tools. MCS 2.0 represents an
approach to adaptation - offering cognitive challenges to those who can benefit, while providing
support to those who need it.
The question before us is whether we have the collective wisdom to implement such
adaptations before divergence becomes irreversible.
At the time of this writing (December 1, 2025), early signs confirm the theory’s core predictions.
McKinsey ’s November 2025 survey of nearly 2,000 organizations globally shows that only 6%
qualify as “high AI performers” (achieving ≥5% EBIT impact), while 61% report zero enterprise-
wide impact, despite widespread adoption of AI. The predicted categorical divergence - not a
gradual stratification, but an economic elimination - is already observable three years after
ChatGPT. Expectations for workforce reductions have accelerated from 17% (2024) to 30%
(2025-2026), consistent with the attrition mechanism modeled in Chapter 2. The most important
aspect is the gap between those who can fundamentally redesign their workflows (high
performance, 3 times more likely) and those who cannot, reflecting the cognitive capacity
asymmetry that this article identifies as the main driver. The intervention period is not 2026-
2028. It is now. "Alea iacta est." (Julius Caesar)
The theory is testable. The predictions are falsifiable. The interventions are designable.
The timeline is clear.
What remains uncertain is... whether we will act.