ChatGPT Was Supposed to Save the World. Four Years Later, It Hasn’t.

November 2022 – OpenAI released ChatGPT to the public.
July 2026 – The U.S. Bureau of Labor Statistics published non‑farm business sector output per hour: up 1.2% year‑on‑year. That is exactly in line with the average annual growth rate of the past decade.

Between these two lines lies an entire cognitive chain of what we might call the “productivity illusion.”

I. Fast Does Not Mean Valuable

What ordinary people call “productivity” and what economists call “productivity” are not the same thing.

Ordinary people: amount of work completed per unit of time. Using AI to write code, design slides, or generate marketing copy—speed triples or quadruples. That experience is real.

Economists: economic value of output divided by economic value of inputs. They don’t care how fast you write. They care how much the final product sells for, and how much total cost went into producing it.

The problem is this: AI makes output faster, but the quality distribution of that output is highly uneven.

A designer using Midjourney generates one hundred images in a day. Eighty are stylistically repetitive, fifteen have perspective errors, three raise copyright concerns, and two just don’t “feel right” to the client. The total time spent filtering these images, communicating revisions, and managing legal risks may well exceed hiring a junior designer to draw from scratch.

This is not an isolated case. Internal evaluation data from GitHub Copilot suggests that developers using AI assistance increase their commit speed by about 55%, but the bug rate and security vulnerabilities of their code rise correspondingly. The time spent later fixing those vulnerabilities virtually cancels out the time saved upfront.

More subtly, these same developers typically report in subjective surveys that they feel 20%+ more productive, while objective metrics show their effective delivery quality dropped by roughly 18%.

That is the first layer of the “productivity illusion”: you think you are making progress, but you are actually generating more mess to clean up.

II. AI’s Reliability Cannot Support Business‑Grade Tolerance for Error

Generative AI is, at its core, probabilistic prediction, not logical reasoning.

It excels at packaging information in fluent language. But it takes no responsibility for verifying whether that information is true. This is the “hallucination” : an output that is grammatically correct, structurally coherent, yet factually invented.

In a poem, a hallucination is poetic license. In a commercial contract, a hallucination is legal liability.

A U.S. law firm once faced judicial reprimand and disciplinary action because a lawyer used ChatGPT to draft a brief that cited six entirely fictitious precedents, cases that never existed. This is not a question of tool efficiency. It is a question of untrustworthiness at a critical juncture.

The corporate response to unreliability is “human‑in‑the‑loop”: every piece of AI‑generated content must be reviewed by senior professionals.

This creates a cost inversion: AI churns out large volumes of semi‑finished work, and senior staff spend substantial time reviewing and correcting errors. Senior staff charge far higher hourly rates than junior staff. If the cost of correction approaches or exceeds the cost of hiring juniors to do the work from scratch, then introducing AI becomes economically negative.

Unless the task has a very high error tolerance, AI does not cut costs and boost efficiency; it merely shifts costs from the “execution end” to the “review end”—and the review end is more expensive.

III. Organizations Are Slower Than Tools

Take a step back. Suppose the hallucination problem is completely solved—AI is absolutely reliable, with zero errors.

Would productivity then necessarily rise?

Not necessarily. Because productivity gains depend not only on the speed of the tool, but also on the speed at which the organization can absorb the output.

Researchers at Asana’s lab introduced the concept of “organisational absorption capacity”: an organisation’s effective output is determined by the slower of two rates—production rate and absorption rate.

AI raises production rate tenfold. But after a report is generated, it still has to go through a three‑day internal approval workflow, be scrutinised by ten people in weekly meetings, and shuttle between departmental silos. The absorption rate has not changed. The bottleneck shifts from “can’t write it” to “can’t decide on it.”

This phenomenon has happened before.

In the 1890s, American factories began replacing steam engines with electric motors on a large scale. Factory owners removed steam engines, installed motors, but kept the old plant layouts designed around central power sources and line‑shaft belts. The result: for the next thirty years, U.S. manufacturing productivity barely grew.

It was not until the 1920s that a new generation of entrepreneurs completely abandoned the old factory designs and reorganised plants around the logic of “distributed power and assembly‑line workflows,” and only then did productivity take off.

In those thirty intervening years, the tool changed, but the organisation did not.

Today, we may be in a similar thirty‑year window: we plug AI into old workflows as if it were a super‑employee, while the processes, performance metrics, accountability structures, and information flows remain industrial‑era legacies.

IV. Replacement, Not Creation

There is a deeper issue, beyond the technical and organisational layers.

Past technological revolutions, like cars replacing horse‑drawn carriages, eliminated old jobs but created entirely new industries. Automobile manufacturing, oil extraction, road construction, maintenance, insurance, driving schools. The entire industrial ecosystem expanded, and total employment grew. This was “creative destruction”: the pie got bigger.

The replacement logic of this wave of generative AI is different.

It targets digitalisable, standardisable cognitive labour in the service and digital economies: programmers, graphic designers, translators, copywriters, VFX artists. It does not create new consumer demand; it merely allows companies to obtain, at a lower price, services they previously had to pay a premium for.

A gaming company reduces its art team from 100 to 20 people, supplemented by AI tools, achieving the same or even greater output. The company’s profit statement looks better. But society has lost 80 high‑paying jobs and gained 80 people who need to find new work.

The purchasing power of those 80 people evaporates. The street vendor selling pancakes downstairs finds he sells 40 fewer each day. The clothing store in the mall sees its revenue drop. This chain does not need to be labelled with the term “multiplier effect”—it happens every day.

The pie has not grown; only the way it is sliced has changed, and the slice now tilts more toward capital holders.

This is not a moral judgment; it is a structural description. And it explains why macroeconomic productivity data show no obvious change—localised efficiency gains are offset by wider consumption contraction.

V. Not Yet Time for a Final Verdict

None of the above is to say that “AI is useless.”

History repeatedly reminds us that the diffusion cycle of general‑purpose technologies is measured in decades.

The Internet entered the public eye in 1995, but it was not until the mobile‑Internet boom of 2010 that it truly permeated the capillaries of the real economy. In those fifteen intermediate years, the Internet was also “everywhere except in the productivity statistics.”

AI may well be at a similar stage. When the first generation of truly “AI‑native” enterprises—those designed from birth around human‑AI collaboration in their architectures, incentive systems, and business models—become mainstream, that productivity leap may finally arrive.

But that is for later.

These past four years have been a painful period of friction between new tools and old systems. AI has increased individual output speed but has not yet changed how organisations absorb; AI has reduced execution costs in some links but shifted costs to more expensive review ends; AI has replaced several roles without creating an equal number of new ones.

The real problem lies not with AI itself, but with the fact that “the world that uses AI” is not yet ready.

A company that introduces AI while simultaneously redesigning its performance metrics, accountability rules, promotion pathways, and error‑tolerance mechanisms may be on the right track.

A company that merely stuffs AI into old workflows, using new technology to do old things, is only adding a footnote to the productivity illusion.