When Everyone Looks Like an Expert
AI is enabling us to do work we could not do before. But the gap between being able to produce that work and being able to judge whether it is sound is widening. This is where a new problem for business begins.
A comprehensive investment report lands on a company board's desk. The financial projections are ready, the industry analysis is complete, and the risks have been classified. The charts are in order, and the calculations look consistent.
Work that once took days can now be prepared in a few hours.
But if the report's underlying assumption is wrong, who will notice?
This is an issue we do not discuss enough in the debate about AI. The technology makes it easier to produce work that requires expertise. But it may not be spreading the expertise needed to evaluate that work at the same pace.
Nor is this simply a technological issue. It is an economic problem that could change how companies assess productivity, the skills they expect from employees, and the way they make decisions.
Four different experiences of AI in the same world
On October 11, 2026, Andrej Karpathy, who was part of OpenAI's founding team and led AI work at Tesla, shared an interesting observation.
In Karpathy's view, the world is divided into people who experience what AI can do at very different levels.
By his own rough estimates, around 6 billion people have had almost no contact with large language models. Between 1 and 2 billion use chatbots occasionally. Around 20 million use advanced models for professional projects, while only a few thousand may have internal access to the most advanced systems.
It is important to note that these are not measured usage statistics, but Karpathy's estimates to illustrate the scale.
The numbers, though, are not the most interesting part.
One person uses AI to write an email, while another uses advanced versions of the same technology to run software projects, solve mathematical problems, or manage research processes.
Both say they use AI. Yet what they experience is almost like two different technologies.
The divide Karpathy describes is emerging not only in access to technology, but also in people's understanding of what it can do.
I think there is a second, less-discussed divide here: the gap between the complexity of the work AI can produce and the human capacity to evaluate that work independently.
Using AI is one thing; changing how you work is another
According to Stanford University's 2026 AI Index Report, 88% of the organizations surveyed use AI in at least one business function. Yet the use of AI agents remains in single digits across almost all business functions.
In other words, trying a technology and rebuilding business processes around it are not the same thing.
The figures from Türkiye also point to a significant shift.
According to data released by the Turkish Statistical Institute (TÜİK) on October 2, 2026, the share of people aged 16–74 using generative AI rose from 19.2% to 37.6% in one year. Among businesses with at least 10 employees, the share using any AI technology reached 14%.
These percentages cannot be compared directly, because different types of use are measured for individuals and businesses. They are nevertheless useful in understanding that the technology's entry into everyday life and its integration into corporate processes are not moving at the same pace.
One detail is even more striking.
Among businesses that do not use AI but are considering it, 72.3% cite a lack of expertise as one reason for not using it. Legal uncertainty concerns 66.4%, while 65.4% are concerned about data security and privacy risks.
Companies need more than additional AI tools. They need an organization that knows what to use those tools for, can assess their results, and can define its responsibilities.
Does using a skill mean you possess it?
A study published in Nature Human Behaviour on October 26, 2026, makes this question even more interesting.
In a controlled experiment involving 487 management consultants, participants who received access to ChatGPT and training in its use performed better on data science tasks outside their own areas of expertise.
However, technical knowledge assessments conducted after the experiment, without AI assistance, found no evidence that participants had independently acquired those same skills.
The study shows that technology-enabled performance gains and lasting learning may be different things.
This does not mean that people stop learning when they use AI. Such a generalization would go beyond the scope of the research.
After all, technology is not always intended to make us experts in the work it does. You do not need to perform every calculation in your head to use a calculator, or be a pilot to travel by plane.
But there is an important difference.
You can check mathematically whether a calculator is working correctly. In a complex investment decision, correct calculations alone are not enough. The assumptions must also be consistent with economic reality.
Consider, for example, a 13-week cash flow forecast prepared with AI.
The model may have applied every formula correctly. But if it assumes that customers will pay faster than they have in the past, with no concrete development to support that expectation, the resulting liquidity outlook is misleading.
The model has not made a calculation error.
The error lies in the assumption it was instructed to accept as true.
This is precisely where expertise comes in. Not just to perform the calculation, but to question the story behind the numbers.
The cost of speed that never appears on the balance sheet
We have concrete research on AI's contribution to productivity.
In a field experiment published by the NBER, covering 66 companies and 7,137 knowledge workers, employees who regularly used generative AI integrated into their work applications spent around two fewer hours a week on email.
However, providing AI access at the individual level did not produce a significant change in the quantity or composition of employees' tasks.
So there is no automatic link between an employee working faster and a company becoming a more productive organization.
Imagine a finance team that used to prepare five comprehensive analyses a month and can now produce twenty.
At first glance, this is a major increase in capacity.
But has the management team's capacity to evaluate those analyses, question their assumptions, and turn their findings into decisions also increased fourfold?
If not, the company may now have more information. But it does not necessarily have better decisions.
In some cases, the opposite may even happen.
Every new report that becomes easier to prepare creates new assumptions to check, new scenarios, and new possibilities for error.
A kind of burden may accumulate without appearing in the company's financial statements: judgment debt, created by decisions that have not yet been sufficiently questioned.
I use this term to describe how the capacity to evaluate and verify can fall behind as the volume of AI-produced work grows.
And the danger is not limited to wrong answers. It includes answers that look perfectly reasonable, are well written and correctly calculated, but rest on incomplete context.
An obvious error is likely to be corrected. An error presented convincingly is more likely to find its way into the decision-making process.
Where will the new competitive advantage emerge?
In economics, when the cost of one production input falls, a constraint elsewhere in the system becomes more visible.
AI is reducing the cost of producing knowledge-based output. Drafting a report, developing software, or calculating different scenarios is becoming easier.
But assessing the reliability of that output, deciding whether it is suitable for a decision, and taking responsibility for the consequences still require considerable expertise.
Rather than eliminating professions altogether, this may change which activities account for most of the value within some professions.
In finance, for example, value may shift from simply building a model to understanding the conditions under which it will no longer hold.
In law, from drafting a contract to anticipating the commercial and legal consequences of a particular clause.
In management, from gathering information to making a reasoned choice between conflicting pieces of information.
Of course, AI can also be used in verification and oversight. Software testing, data reconciliation, consistency checks, and comparative analyses are important examples. The point is not to make humans redo all the work, but to design how each result will be verified and who will take responsibility for each risk.
It may therefore be misleading to judge tomorrow's successful companies solely by the power of the model they use or the number of prompts their employees write.
There are more meaningful questions:
Which decisions has the company improved through AI? Which errors has it caught earlier? Which processes has it redesigned? Which results can it reliably verify?
And perhaps most importantly, who can challenge the system's output when necessary?
On one side of the great divide Karpathy describes are people who have yet to become acquainted with AI; on the other are those pushing its limits.
But that distinction alone will not determine the future of business.
Very soon, producing a well-prepared report, an impressive analysis, or a complex financial model may no longer be a distinguishing skill in itself.
When everyone can prepare a report, the real value will lie with the person who can spot the one flawed assumption within it.
Sources
Andrej Karpathy, post on X, October 11, 2026. Growing differences in AI use and perceptions of its capabilities.
Stanford HAI, AI Index Report 2026, Economy. Organizational adoption, AI agents, and economic impacts.
Wiles et al., Generative AI and the temporary upskilling of knowledge workers. Nature Human Behaviour, October 26, 2026.
Dillon et al., Shifting Work Patterns with Generative AI. NBER Working Paper 33795, 2025.
Comments