Artificial intelligence has arrived with extraordinary expectations attached to it. Many in the technology sector argue that it will unlock a new era of economic expansion, comparable to-or even exceeding-the productivity surge driven by the personal computer and the internet. Enthusiasts point to rapid improvements in generative models, rising corporate adoption, and early efficiency gains in selected industries as evidence that a broad productivity acceleration is imminent.

Yet the historical and empirical record so far suggests a more restrained conclusion. Despite impressive technical progress, there is limited evidence that AI is currently translating into sustained, economy-wide productivity growth. In fact, the underlying dynamics of AI adoption suggest that its impact may be structurally different from previous general-purpose technologies. Rather than simply accelerating work, AI may be shifting the bottleneck from production to verification, creating constraints that significantly dampen its aggregate productivity effects.

To understand why, it is useful to revisit the last major episode of productivity acceleration: the computer revolution. During the late 1990s and early 2000s, the United States experienced a significant increase in output per hour, growing at roughly 3 percent annually. This surge was driven by widespread adoption of personal computers, enterprise software, email, and the internet. These technologies fundamentally reduced the cost of accessing, transmitting, and storing information.

The key characteristic of these earlier digital tools was that they automated processes of retrieval and communication, not the generation of knowledge itself. A spreadsheet did not invent new arithmetic; it merely executed calculations faster. A search engine did not generate original facts; it retrieved existing ones. Email did not create new ideas; it moved messages more efficiently between people. In each case, digital systems substituted slower methods of information handling with faster ones, while preserving the underlying integrity of the output.

This distinction is crucial. Because the outputs of early digital tools were deterministic and verifiable, the productivity gains were relatively straightforward. Users could trust that a calculation in a spreadsheet was correct if the inputs were correct. A document retrieved from a database was the same document regardless of the method of retrieval. Errors existed, but they were generally traceable to human input rather than system-generated ambiguity.

Artificial intelligence systems, particularly large language models, operate in a fundamentally different domain. They do not merely retrieve or transmit information; they generate it. That generative capacity is precisely what makes them powerful, flexible, and widely applicable. However, it is also what introduces a new and persistent economic friction: the problem of verification.

Unlike a spreadsheet or search engine, an AI system can produce outputs that are fluent, plausible, and incorrect all at once. These “hallucinations” are not rare edge cases but structural features of probabilistic generation. The system is optimized for coherence and likelihood, not truth. As a result, its outputs often require external validation before they can be safely used in consequential contexts.

This creates what can be described as a verification tax. Any time AI is used to produce work that carries real-world consequences-legal filings, financial transactions, medical recommendations, engineering decisions-the output must be reviewed by a human or another trusted system. The time saved in generating content is therefore partially or fully offset by the time required to check it.

Recent empirical studies reinforce this tension. In structured environments such as customer support, AI tools can improve productivity meaningfully, particularly among less experienced workers. This is because the tasks are standardized and the outputs are easy to evaluate. Workers can quickly determine whether a suggested response is correct or appropriate, and the AI effectively serves as a real-time assistant distributing best practices.

However, in more complex domains, the benefits diminish significantly. Studies involving experienced software developers working on their own codebases show that access to advanced AI tools can actually reduce productivity. The reason is not that the AI is useless, but that integrating its output requires time-intensive review, debugging, and correction. The cognitive load shifts from creation to evaluation, and that shift can be costly.

The implications become even more pronounced in high-stakes professional environments. A recent incident involving a major law firm filing documents containing fabricated citations generated by AI illustrates the risk clearly. The errors were not caught internally but by opposing counsel. While such incidents may appear anecdotal, they highlight a structural vulnerability: AI systems can produce convincing but false information at scale, and detecting those errors requires expertise, attention, and time.

As AI systems become more “agentic”-capable of taking autonomous actions such as modifying codebases, executing transactions, or interacting with external systems-the consequences of errors increase significantly. A flawed paragraph is an inconvenience. A flawed automated financial transaction or database operation can be materially damaging. This escalation in stakes further increases the verification burden.

In this sense, the limiting factor in AI-driven productivity is not generation capacity but organizational verification capacity. There is a finite amount of human attention, expertise, and accountability available to validate outputs. Even if AI can produce infinite drafts, summaries, or decisions, organizations cannot infinitely scale their ability to certify correctness.

This constraint becomes even more significant when considering how firms may adapt structurally. If companies reduce hiring of junior professionals on the assumption that AI can handle entry-level tasks, they may inadvertently weaken their internal verification systems. Junior employees traditionally serve as both producers and learners, gradually acquiring domain expertise through repetition and supervision. If that pipeline is disrupted, organizations risk losing the very human expertise required to evaluate AI outputs effectively.

The paradox is that the more AI is relied upon to replace early-stage human work, the more difficult it becomes to verify its outputs at later stages. Over time, this could lead to organizations that appear more efficient on paper but are in fact more fragile in practice, with hidden error accumulation that only becomes visible when failures occur in high-stakes environments.

The broader macroeconomic implication is that AI’s productivity impact may be uneven and bounded by task structure. Where outputs are easily testable, low-risk, and standardized, AI will likely deliver meaningful efficiency gains. Where tasks are complex, ambiguous, or high-stakes, gains will be muted or offset entirely by verification costs. This suggests a highly asymmetric productivity landscape rather than a uniform surge.

If AI is to deliver broad-based productivity acceleration comparable to the computer revolution, it will likely require complementary institutional evolution. Specifically, economies will need to develop what might be called verification infrastructure. This includes technical systems such as provenance tracking, audit trails, and model explainability tools, as well as institutional mechanisms such as updated professional standards, liability frameworks, and regulatory certification processes.

Some early signals of this shift are already emerging. In certain legal contexts, courts are beginning to require explicit certification that AI-generated material has been independently verified using traditional methods. Similar practices may eventually extend into finance, healthcare, engineering, and public administration, where the cost of error is high and accountability is non-negotiable.

However, these institutional adaptations typically move far more slowly than technological innovation cycles. AI models evolve in months; legal norms, regulatory systems, and professional standards evolve in years or decades. This temporal mismatch means that the productivity benefits of AI are likely to remain constrained in the near to medium term.

Ultimately, the key insight is that AI does not simply accelerate existing workflows; it redefines them. In doing so, it introduces a new economic constraint centered on trust, verification, and responsibility. The question is not whether AI can produce output quickly-it clearly can-but whether societies and organizations can reliably determine which outputs are correct, safe, and actionable.

Until that verification problem is solved at scale, expectations of a near-term AI-driven productivity boom comparable to the computer revolution may prove overly optimistic. The technology is powerful, but its economic impact is mediated by a slower, more human constraint: the capacity to trust what machines produce.

Please follow Blitz on Google News Channel