Why the extraordinary expectations surrounding large language models have yet to show up in the economic data
Large language models may be the fastest-diffusing digital technology in decades. Within only a few years, they have become everyday tools for hundreds of millions of people, entered enterprise software and triggered extraordinary investment in chips, data centers, energy infrastructure and AI startups.
The problem is not that the technology does not work. It does, and its capabilities continue to improve at remarkable speed.
The problem is that financial expectations have moved even faster.
Markets have already priced in the assumption that artificial intelligence will become one of the defining productivity transformations of the twenty-first century. By mid-2026, AI-related companies represented roughly half of the market capitalization of the S&P 500, up from about one quarter in 2022. Global corporate AI investment reached roughly $582 billion in 2025 alone.
That creates the central paradox of the current AI boom: financial markets have already priced much of the revolution, while the revolution remains difficult to see in aggregate productivity data.
The extraordinary gap between investment and measured returns
The United States is the clearest testing ground because it combines the world’s largest AI investment boom with some of the fastest adoption of the technology.
Yet U.S. nonfarm business labor productivity grew 2.2% between the second quarter of 2025 and the second quarter of 2026. That is a respectable rate. It is not an industrial revolution.
Average productivity growth in the current U.S. business cycle has been around 2.1% a year, essentially equal to the long-run historical rate.
In other words, there is no economy-wide productivity discontinuity yet remotely comparable to the discontinuity visible in investment, infrastructure spending or equity valuations.
This does not mean that AI produces no productivity gains. It clearly does. Controlled studies have found substantial improvements in customer support, programming, writing and other structured forms of knowledge work.
But accelerating one task by 30% does not increase the productivity of an organization by 30%.
Between the model and the firm’s bottom line sit workflows, databases, approvals, legacy software, security rules, management structures and human accountability.
AI saves labor, but it also creates a verification tax
A language model can draft a report in seconds. Someone still needs to determine whether the report is correct.
That distinction becomes more important as the cost of error rises.
A mistake in a marketing draft may be trivial. A mistake in a tax interpretation, credit decision, medical recommendation, legal document or government decision may be extremely expensive.
This creates what might be called a verification tax. The more consequential the output, the more human review is required.
The relevant productivity metric is therefore not how fast a model produces an answer. It is whether the entire process, including verification, integration, correction, governance and exception handling, becomes cheaper or better.
That is a much harder test than a benchmark.
Businesses are adopting AI much faster than they are realizing productivity gains
Recent firm-level evidence makes the gap particularly striking.
A 2026 survey of almost 6,000 senior executives across the United States, United Kingdom, Germany and Australia found that roughly 69% of firms were actively using AI.
Yet more than 80% reported no impact so far on either productivity or employment.
That is perhaps the most important fact in the current debate.
We are not looking at a technology that firms refuse to use. We are looking at a technology that firms are adopting rapidly without yet producing an economic transformation remotely proportional to the expectations placed upon it.
That distinction matters because the investment bill keeps rising.
The more capital that flows into accelerators, data centers, electricity generation and model development, the larger the future stream of economic value required to justify that capital.
The question is therefore shifting. It is no longer merely whether people will use AI. They already do. It is whether sufficiently valuable uses will emerge at sufficient scale to generate returns commensurate with the enormous infrastructure being built around them.
Public-sector transformation is harder still
The same mismatch is visible in government.
The promise is compelling: automated citizen support, faster permits, document processing, fraud detection and better public services.
But government cannot operate like a probabilistic chatbot.
An administrative answer must reflect the law. A decision affecting a citizen must be explainable and contestable. Personal data must remain protected. Someone must remain accountable when the system is wrong.
OECD evidence illustrates the problem. Governments are creating AI strategies and pilots quickly, but only 10 of 36 surveyed OECD countries report conducting any financial or non-financial impact measurement of government AI use cases.
The bottleneck is therefore no longer experimentation. It is proving that experimentation produces public value.
AI can make government substantially more capable, but only when it sits on top of redesigned processes, authoritative data, interoperable systems, open standards, human oversight and measurable service objectives.
Simply adding a language model to a dysfunctional administrative process mostly creates an AI-enabled dysfunctional administrative process.
A transformative technology can still produce a financial bubble
There is an important historical misconception in the current debate: that describing financial excess somehow implies that the underlying technology is worthless.
History suggests precisely the opposite.
Railways transformed modern economies and produced speculative bubbles. The internet transformed civilization and still produced the dot-com crash. Houses are indispensable economic assets, yet housing markets can produce bubbles.
Technology can be consequential while investors are simultaneously wrong about the speed of adoption, the amount of future profit or the identity of the eventual winners.
That is now the central question facing AI.
Today’s valuations do not merely reflect today’s AI revenues. They embed assumptions about enormous future productivity gains, rapidly expanding demand and very large profit pools.
Those assumptions may ultimately prove correct.
But they have not been demonstrated yet.
That makes today’s AI valuations, to a significant degree, claims on future productivity.
The next phase must be about evidence, not spectacle
The useful response is neither to dismiss AI nor to accept every forecast made in its name.
Organizations should stop measuring progress by the number of AI pilots they launch and start measuring it by processes improved, costs reduced, waiting times shortened and outcomes made better.
Governments should require the same discipline. Public agencies should prefer interoperable architectures, open standards and, where appropriate, open models and local deployment so that systems can be inspected, replaced and operated without permanent dependence on a single supplier.
The most important AI question of the next few years will therefore not be whether the next model scores higher on another benchmark.
It will be whether hundreds of billions of dollars of investment begin producing an unmistakable and sustained rise in real productivity.
Financial markets have already priced in the revolution.
Now the real economy has to deliver it.