The Flynn Effect and the Age of AI

I recently asked which camp people were in when it came to AI. The responses were split almost exactly three ways: the models will improve, people will adapt, or reliability has to be engineered around the systems available now.

The argument for human adaptation is appealing because history gives us many examples of people learning to work at a higher layer of abstraction. I went back to a framework that David Epstein discusses in *Range* and found a useful connection to the Flynn Effect.

The connection is worth exploring, but it needs one important qualification: the Flynn Effect does not prove that technology caused people to become more intelligent. It gives us evidence that performance on intelligence tests changed across generations and that the pattern varied across cognitive domains and environments.

What the historical evidence actually shows

The Flynn Effect is the long-run rise in intelligence-test performance observed across many countries during much of the twentieth century.

A 2015 meta-analysis of 271 samples, covering almost four million participants in 31 countries from 1909 to 2013, found gains across fluid, spatial, full-scale, and crystallized test performance. The gains were not uniform: they varied by country and domain, were larger in some forms of fluid and spatial reasoning, and weakened in more recent decades. The authors also noted that the causes remain unsettled.

That is more precise than saying every person became three IQ points smarter every decade or that tests simply became harder. Intelligence tests are periodically re-standardized, and a population-level change in test performance does not have one proven explanation.

The part I find relevant to AI is narrower. Human cognitive performance is not fixed independently of the environment. People learn the abstractions that their education, work, tools, and institutions require them to use.

A working hypothesis about abstraction

The model I am considering is straightforward: as work moves from physical objects to symbols, code, systems, and now AI-generated recommendations, the valuable human skill moves toward reasoning across those layers.

This is a working hypothesis, not a causal history of the Flynn Effect. Education, health, nutrition, economic conditions, test familiarity, and other environmental changes also matter. The visual helps isolate the question I care about: what happens to human judgment when the tool performs more of the lower-layer reasoning?

Why range matters in enterprise work

Epstein uses the distinction between kind and wicked learning environments to explain why narrow specialization dominates some problems while range helps with others.

Most enterprise AI decisions live closer to the wicked side. The system may need technical context, operating policy, customer impact, organizational history, and an understanding of the consequence for people. The ability to recognize a pattern is useful; the ability to know when the pattern does not apply is more important.

This is the strongest version of the human-adaptation argument. AI becomes the next abstraction layer. People stop doing some lower-level work, learn to reason across a larger system, and contribute at a different level.

I think that argument has real merit. I also think it is incomplete.

The analogy holds only when people still reason

The adaptation argument works only when people still engage with the new layer. A high-level programming language removes the need to manage some machine details, but developers still have to understand the program, test its behavior, and own the result.

An AI system can create a different pattern. A person may review a recommendation generated by a model they did not build, using data they did not curate, across actions they cannot fully reconstruct, at a pace that discourages careful examination.

The Flynn Effect tells us that human performance has changed as environments changed. It does not show that people automatically get better at slowing down, finding the consequential edge case, or understanding an action approved at machine speed.

Generalists thrive when they can connect domains and test one frame against another. That requires time, context, and responsibility. If AI removes those conditions, the organization does not gain range. It gains distance.

The practical question is therefore not whether humans have adapted to technology before. They have. The question is whether we are designing AI-assisted work so that people still have to understand the problem deeply enough to know when the system is wrong.

If the answer is yes, AI can help people move up the stack. If the answer is no, we are only moving further from the ground.