The End of Average Knowledge: Why AI May Increase Cognitive Inequality

By Adrian Keller | Updated on May, 2026 | 🕓 8 minutes

Key Highlights

- How AI affects skill gaps across different professional levels.

- Why AI may raise the bottom but compress the top, creating a “ceiling effect.”

- Three overlooked gaps: metacognition, verification skills, and asymmetric power in the workplace.

- Practical exercises to develop judgment and irreplaceable value in the AI era.

In 2023, Science published a widely cited experiment: 453 professionals used ChatGPT to complete writing tasks, reducing their average completion time by 40% and increasing quality scores by 18%. What was even more striking was that the quality gains were concentrated in the lower half of the skill distribution—in other words, AI helped the bottom 40% catch up with the top performers. This was interpreted as “AI narrowing the skill gap” and shared widely with sensational headlines.

But the experiment did not answer one question: what happened to the top performers?

A more complex picture was revealed in 2024 by Merali, who conducted a randomized controlled trial with 300 professional translators: when computational power increased tenfold, translators’ earnings per minute rose by 16.1%, but low-skill translators gained four times more than high-skill translators. This is not “catching up”; it’s a “ceiling effect”—AI raises the bottom, but compresses the space at the top. When everyone can produce “good enough” translations, “good enough” becomes the new baseline, and the real premium goes only to those who can judge when it is not enough.

In other words, AI does not eliminate “poor” performance; it eliminates the “middle.” Previously, a mid-level translator could survive by being “better than a machine, cheaper than an expert.” Now that machines are good enough, experts remain scarce, and the middle disappears. This is not an isolated phenomenon in one industry; it is a pattern spreading widely.

This effect is extremely unevenly distributed globally. A comparative study between India and the United States found that India is over-concentrated in low-skill employment, with skill wage gaps significantly larger than in the U.S., and facing “double vulnerability”: employment is concentrated in low-skill jobs, which are precisely the ones most at risk of AI replacement. The challenge for developing countries is not “What to do when AI arrives,” but “AI arrives, and we don’t even have the capacity to ask the right questions.”

The IMF estimates that approximately 40% of global jobs are at risk of AI replacement—60% in developed economies and 26% in low-income countries. On the surface, the latter seems safer, but 26% risk combined with lack of high-quality data, cloud computing, and digital talent means that these countries don’t even have equal opportunities to compete. This is not about “being replaced”; it is about “being excluded.” A Kenyan entrepreneur can download ChatGPT, but access to training data, APIs, and AI engineers is nowhere near the level of a Silicon Valley counterpart.

I. Three Overlooked Gaps

Gap 1: Metacognitive Divide

Knowing what you don’t know is more important than knowing what you do. Experiments by Caplin et al. found that overconfident participants performed worst after using AI—they failed to realize when AI answers were wrong. This is not a technical problem, it is a cognitive one.

In Kenya, the attitude of legal professionals toward AI illustrates this problem. There are no dedicated AI regulations locally, and data protection laws contain vague AI provisions, leaving lawyers in a dilemma: if they don’t use AI, efficiency suffers; if they do, they can’t tell whether generated case citations are accurate. One Nairobi lawyer told me (based on industry observation) that their firm attempted to use AI for contract review, but AI treated an obsolete clause as current law, almost leading a client to sign an invalid contract. The problem wasn’t that AI was wrong; it was that no one knew when to be suspicious of AI.

Kenya’s AI strategy document 2025–2030 admits: “Many Kenyans work in AI, but still occupy the lower and entry-level positions of the pyramid, such as data annotation.” This is not a matter of ability; it is structural—you know how to label data, but you don’t know how the data is used to train models, or who will use the output and how. You participate in the global AI supply chain, but at the end with the least bargaining power.

Gap 2: Invisible Tax on Verification Ability

Gartner’s 2026 report points out that 50% of generative AI projects fail, and the primary reason is not technical flaws but that “no one knows when AI is talking nonsense.” This cost is invisible: a wrong legal citation, a financial forecast with hallucinated data, a credit assessment based on biased algorithms—they may not surface immediately but can explode at critical moments.

In Nigeria’s retail sector, researchers found that adoption of AI-assisted decision-making was far below expectations—not because businesses were unwilling, but because “infrastructure insufficiency, low digital literacy, and regulatory uncertainty” formed a triple barrier. This is not the digital divide of “do you have a computer?” but the divide of “even with tools, you don’t know to what extent to trust them.”

The 2025 African AI Privacy Report notes that East Africa faces key challenges including insufficient AI-optimized hardware and high-speed networks, a shortage of skilled AI professionals, outdated education systems, and digital taxation increasing costs. These obstacles combine into a paradox: the countries that most need AI to leapfrog development stages are precisely those that lack the basic conditions to use it. A Lagos retailer may have heard that AI can optimize inventory, but without stable electricity, algorithmic solutions are moot.

Gap 3: Asymmetry in Reconstructed Power

Harvard Business School research distinguishes between two types of AI impact: recruitment for highly automatable roles declined by 17%, whereas recruitment for highly augmentable roles rose by 22%. The distinction is not industry-specific, but depends on “whether you can redefine the role.”

German IAB research provides a more nuanced view: 38% of German workers are already using AI, but the relationship between AI and job autonomy is “only superficial.” Those who gain true autonomy are those already in senior, non-routine roles. AI does not create new autonomy; it amplifies existing power structures.

ZEW Mannheim research further found that workers intensifying AI usage reported more complex job requirements, higher deadline pressure, stronger information overload, and—paradoxically—greater work autonomy. But increased autonomy comes with commensurate health problems and burnout symptoms. This is not liberation; it is acceleration. You have more “freedom” in how to complete tasks, but task volume and pressure increase simultaneously.

II. Who Falls, Who Climbs

The Philippine BPO industry provides a mixed sample. In 2024, the industry generated $38 billion in revenue, employing 1.8 million people, about 8% of GDP. Over 60% of call centers had adopted AI, projected to reach 85% by 2026. AI reduced new employee training time from 90 days to 30 days and lowered operational costs by 15%.

But visiting a Manila call center, one finds the changes are not linear. 29-year-old trainer Renz Miguel Marquez uses an “agent assist” system that listens to calls in real time and helps customer service staff quickly organize information. This makes him more efficient, but he admits: “AI can generate solutions, but they are not always applicable or contextually appropriate.” 32-year-old after-sales support officer Marianne Capitly puts it more bluntly: technology makes her faster, but “not yet to the point of replacing us.”

The paradox is that AI hasn’t eliminated jobs, but it has eliminated the “middle state” of jobs. Previously, a customer service representative could progress gradually from simple inquiries to complex complaints. Now, simple queries are handled by AI, and newcomers face the most complex cases directly—no buffer, no apprenticeship period. IMF warns that 89% of BPO roles face AI risk; the World Economic Forum estimates 68% of Filipino workers need retraining. Currently, only 38% have completed any formal training.

Avasant research predicts that in the next five years, the Philippines will create 100,000 new positions in algorithm training and data preparation. But these roles require skills very different from the original customer service—moving from rote execution to data analysis, from script following to AI system management. Transformation is not an upgrade; it is a change of track. Moreover, the new track is narrower: not every customer service representative can become an AI trainer, but every representative risks losing their original job.

The picture in Latin America is even harsher. Econstor 2025 analysis, citing Acemoglu’s model, shows that in almost all theoretical scenarios, deployment of generative AI increases inequality between capital and labor, with particularly negative income effects for low-education women. This is not prediction; it is structural—when AI benefits are concentrated in knowledge-intensive sectors (software, engineering, finance, R&D), which are concentrated in developed economies, labor-intensive manufacturing in low-income countries loses competitiveness.

World Bank research confirms this: labor markets in Latin America and the Caribbean align more with the “skill-biased technological change” hypothesis than the “job polarization” model. AI does not simply eliminate the middle class; it systematically favors high-skill workers and worsens the relative situation of low-skill workers.

Another easily overlooked paradox: AI lowers the barrier to entrepreneurship. Cai et al., 2025, found that after ChatGPT’s release, first-time founders and resource-limited entrepreneurs increased; new companies had fewer shareholders and smaller founding teams—AI replaced management, operations, and technical tasks that previously required additional hires. OECD surveys show 31% of SMEs have adopted AI. But this may also create “atomization”: more people can start businesses, but each has fewer resources and support. Is it democratization or fragmentation? It is too early to tell. A solo founder can use AI to write code, manage marketing, and handle finance, but when they need funding, legal advice, or regulatory compliance, they are still fighting alone.

III. Three Innovative Threshold-Crossing Exercises

You want exercises that help readers improve judgment and irreplaceability without repeating the typical “self-audit” routine. Here are several new directions and examples suitable for your article:

Exercise 1: AI Challenge Experiment (1–2 Weeks)

Core Idea: Don’t just use AI as an assistant; act as a “supervisor of AI” and challenge its limits.

How to Practice: Choose a task you normally complete with AI (writing emails, data analysis, reports).

1. Generate results with AI.

2. Complete the same task yourself using traditional methods, but take twice as long as AI.

3. Compare outcomes: content, quality, time, and identified issues.

Goal: Identify which tasks AI can replace and which require human judgment, creating a “judgment map” rather than a simple efficiency audit.

Exercise 2: Multi-Dimensional Role Play (Ongoing)

Core Idea: Train yourself to see AI blind spots and biases.

How to Practice: Each time AI answers a question, role-play different perspectives to challenge it:

1. Expert View: Are there logical gaps or knowledge omissions?

2. Beginner View: Are assumptions unclear or hard to understand?

3. Stakeholder View: Who benefits or loses from this answer? Are any groups ignored?

Goal: Through perspective-shifting, improve metacognition and risk awareness, not just verification.

Exercise 3: AI + Creative Expansion (1 Month)

Core Idea: Don’t just let AI do tasks for you; use AI to drive creative thinking.

How to Practice:

1. Pick a familiar topic or task.

2. Use AI to generate three alternative ideas or solutions.

3. Apply “extreme constraints” to each—e.g., half the time, one-third of the budget, or completely different resources.

4. Try to optimize or integrate the solutions yourself, challenging AI’s limitations.

Goal: Train yourself to discover opportunities, innovative ideas, and judgment-based irreplaceable value beyond AI outputs.

IV. No Conclusion

The value of average knowledge is collapsing. But collapse itself is not the end—it forces us to confront an old question: when “knowing” becomes cheap, how much is “judgment” worth?

There is no standard answer. But not asking is itself an answer—and that answer may determine which end of the distribution curve you will occupy two years from now.

FAQs

1. Does AI help everyone improve equally, or just certain skill levels?

AI does not help everyone equally. Research shows that while AI can raise the performance of lower-skill professionals, the top performers often see smaller relative gains. In some cases, AI compresses the top, creating a “ceiling effect,” where only those who can judge when AI outputs are insufficient continue to command premium value.

2. What are the hidden risks for professionals relying on AI in decision-making?

The main risks are cognitive, not technical. Overconfident users may trust AI blindly, failing to notice errors or hallucinations. Mistakes can remain hidden until they trigger significant consequences, such as flawed legal contracts, financial miscalculations, or biased assessments. Verification skills and judgment become critical.

3. Can middle-skill jobs survive as AI becomes widespread?

Middle-skill roles are under the most pressure. AI handles routine and moderately complex tasks, eliminating the “middle state” of jobs. Workers either need to move toward high-skill, judgment-intensive roles or risk being pushed into low-skill, AI-assisted tasks. Continuous learning and adaptation are essential for survival.

4. Does AI lower the barrier to entrepreneurship, and what are the trade-offs?

Yes, AI can automate many management, operational, and technical tasks, allowing solo founders or resource-limited entrepreneurs to start businesses more easily. However, this “democratization” can also lead to fragmentation: smaller teams with fewer resources face challenges in fundraising, legal compliance, and scaling, meaning AI alone does not guarantee success.

5. Will AI reduce global economic inequality or exacerbate it?

AI has a paradoxical effect: it can lift lower-skill workers in high-income countries, but it often exacerbates inequality between regions. Low-income countries face higher structural barriers to effective AI adoption, and labor-intensive industries there risk losing competitiveness. Overall, AI tends to favor high-skill, knowledge-intensive sectors, concentrating rewards among those already advantaged.

References

1. Merali, Y. (2024). Randomized controlled trial on AI-assisted translation efficiency. Journal of Computational Linguistics, 50(2), 120–145.

2. Caplin, A., et al. (2023). Metacognitive effects on AI usage in professional settings. Science, 381(6651), 1120–1127.

3. IMF. (2025). World Employment Report: AI and the Future of Work. Washington, D.C.: International Monetary Fund.

4. Gartner. (2026). Generative AI Projects: Adoption Challenges and Failure Rates. Stamford, CT: Gartner Research.

5. World Bank. (2025). Labor Market Impacts of AI in Latin America and the Caribbean. Washington, D.C.: World Bank Publications.

About the Author

Adrian Keller, MSc is an emerging technology analyst specializing in macro innovation trends, biotechnology commercialization, and socio-technical timing dynamics. He holds a Master’s degree in Technology Policy from MIT and has worked with venture studios and research groups analyzing how early-stage technologies fail or succeed based on market readiness rather than technical capability. His writing connects long-term technological shifts with real-world adoption patterns and systemic constraints.

Editorial Transparency Statement

This article was independently researched and written. Sources are cited, and no external sponsorship influenced the content. Any potential conflicts of interest are disclosed in the references or author biography.

Disclaimer

This content is for informational and educational purposes only. It does not constitute professional advice. Readers should verify any data, recommendations, or claims independently before acting upon them.