The number in Anthropic’s latest economic scenarios that deserves our attention is 11.9 percent. That is the unemployment rate that could emerge in the company’s most extreme scenario by 2030, at a time when artificial intelligence is assumed to be advancing rapidly and becoming deeply integrated into the economy. Even more striking is what happens to knowledge workers, the highly educated professionals who have traditionally assumed that their education, experience and specialized skills would provide some protection from technological disruption.
In Anthropic’s extreme scenario, knowledge-worker unemployment could reach 17.9 percent, while employment in knowledge occupations could decline by more than 20 percent. At the same time, the economy could be substantially larger than it would have been without AI. The numbers illustrate the central paradox of the AI economy: it is entirely possible for an economy to become dramatically more productive while a significant portion of the people who created that productivity become economically less necessary.
The jobs most exposed in that environment are not necessarily the jobs we traditionally associate with automation. Previous technological revolutions primarily targeted physical labor and repetitive mechanical tasks, but AI is moving directly into the professions. Software developers, customer service representatives, accountants, financial analysts, paralegals, administrative professionals, researchers, marketers and content creators all perform work that involves processing information, recognizing patterns, generating language, preparing documents or making recommendations. These are precisely the activities that increasingly capable AI systems can perform at a scale and speed that humans cannot match. In many professions, the human being may remain responsible for the final decision while the amount of human labor required to prepare that decision falls dramatically.
That prospect understandably creates anxiety, particularly among people who did everything they were told they were supposed to do. They went to college, acquired professional credentials, developed specialized expertise and entered occupations that society considered intellectually valuable. The implicit bargain was that education would make them more valuable in an increasingly sophisticated economy. AI is now challenging that assumption because some of the very cognitive skills that commanded a premium in the labor market are becoming easier for machines to reproduce.
READ: Sreedhar Potarazu | AI wants to survive: Hugging Face, self-preservation and intuition (September 7, 2026)
Yet there is another side of this discussion that is rarely included when we calculate the economic consequences of AI. We are becoming increasingly sophisticated at estimating the cost of technology replacing a human worker, but we remain remarkably unsophisticated at calculating the cost of human inefficiency.
When an AI system replaces ten people, economists immediately ask what happens to those ten workers, how much income they lose and what it will cost to retrain them. Those are legitimate questions, but there is another question that should be asked at exactly the same time: what were those ten people actually producing, and how much of their work existed because the underlying system was unnecessarily complicated or inefficient?
Employment statistics do not distinguish between productive employment and employment that exists because an organization has failed to modernize. A person creating something valuable and a person spending the day moving paperwork from one department to another are both classified as employed, even though the economic value of their activities may be very different.
A government employee processing a form that could be completed automatically is still counted as a job. So is the employee checking the work of another employee who is checking the work of a third employee because the organization has accumulated layers of procedures over many years. If AI eliminates those positions, we instinctively describe the result as job destruction without asking whether the underlying activity was economically necessary in the first place.
This becomes particularly important when we consider government. Private companies eventually face consequences for sustained inefficiency because customers can leave, competitors can enter the market and investors can withdraw their capital. GE for the longest time had the practice of weeding out the bottom 10% every year. Maybe AI should do the same.
Government agencies do not operate under the same constraints. An inefficient government agency does not lose its customers in the way a private company does, and it generally does not disappear because its processes have become unnecessarily slow or complicated. It can continue operating, continue hiring people and continue consuming taxpayer resources even when decisions take too long, errors are repeated and citizens are forced to navigate systems that are difficult to understand.
The cost of that inefficiency is real, but it is rarely included in the economic calculations surrounding employment. When a government decision takes six months rather than six days, the salaries of the people processing the decision are counted, but the economic cost imposed on the person waiting for the decision is often invisible.
When an agency makes an error and a citizen must hire a lawyer to correct it, the government’s staffing numbers do not capture that expense. When a business loses revenue because a regulatory approval is delayed, that lost economic activity does not appear as an unemployment statistic. When an institution fails to follow its own rules and an individual is forced into years of litigation or administrative appeals, the economic model generally does not assign a cost to the years of time and productivity that have been consumed.
I have seen this problem firsthand over the past decade, and it has changed the way I think about the AI debate. I have experienced what happens when an individual encounters a government institution with enormous delegated authority but finds that the institution does not always function with the consistency, efficiency or accountability that citizens reasonably expect. I have seen how difficult it can be to challenge an institutional decision, even when rules, procedures and evidence should matter, and I have experienced the financial and personal consequences of having to navigate systems that can take years to resolve questions that should have been addressed much sooner. Some of the issues I have encountered have ultimately reached the courts, including issues that are now receiving attention at the Supreme Court.
That experience has made me increasingly skeptical of the assumption that every existing job represents something that society should preserve. We have become accustomed to inefficiency because it is familiar, and we have become complacent to incompetence.
AI has the potential to challenge those assumptions because it can examine processes in ways that humans rarely have the time or incentive to do.
AI could potentially analyze thousands of government decisions, compare them against statutes and regulations, identify inconsistencies, measure processing delays, find redundant procedures and calculate how much money is being consumed by unnecessary administrative steps. It could help determine whether a government agency is actually producing more value by adding another layer of human review or simply creating another opportunity for delay and error. Used properly, AI could become not just a tool for automating government but a tool for auditing government.
That may be the most uncomfortable possibility of all. AI is not merely coming for our jobs; it may force us to examine why those jobs exist in the first place. It may reveal how much of our economy is devoted to correcting errors, navigating bureaucracy, duplicating work and compensating for systems that were designed for an earlier era. If that happens, the AI revolution will become something much larger than a labor-market transformation. It will become an audit of the way we have organized our institutions.
READ: Sreedhar Potarazu | When AI distills intelligence, what happens to context?
Column (August 12, 2026)
The challenge for policymakers is therefore to measure both sides of the equation. We need to calculate the economic and social cost of workers displaced by AI, but we also need to calculate the economic and social cost of maintaining inefficient systems simply because humans currently occupy the positions within them. We should measure the cost of delays, unnecessary administrative layers, regulatory confusion, duplicated work, preventable errors and institutional failures alongside the cost of automation. Otherwise, we are effectively calculating the cost of change without calculating the cost of staying exactly where we are.
Bill Gates is right to warn that AI could eliminate significant categories of employment. Anthropic is right to explore scenarios in which extraordinary economic growth occurs alongside extraordinary employment disruption. Governors such as Wes Moore are right to focus on preparing workers and government for the transition. But there is another conclusion we should be willing to consider. The purpose of technological progress cannot simply be to preserve the number of jobs we have today. It should be to increase the amount of value society receives from the work that people do.
We should therefore be careful about declaring every displaced worker a casualty of AI without first examining what the worker was being asked to accomplish. Some people will unquestionably lose valuable careers because of AI, and society has a responsibility to help them transition. But there will also be cases in which AI eliminates work that should have been eliminated years ago, removes bureaucratic layers that should never have existed and exposes institutional inefficiencies that have survived simply because nobody had a better way to identify them.
Before we hit the panic button over Anthropic’s 2030 scenarios, we should therefore ask a broader question about what we are trying to preserve. If the answer is human dignity, meaningful work, economic opportunity, accountability and a functioning society, then we should welcome anything that helps us achieve those goals. If the answer is simply preserving every existing position regardless of whether the work remains necessary, then we are protecting employment statistics rather than building a better economy.
Perhaps the real test of AI will not be how many jobs it takes away. It will be whether we have the courage to use the technology to identify what is broken, eliminate what no longer creates value and redirect human effort toward the things that actually matter. We have spent years worrying that AI will replace us. Maybe what we should worry about instead is whether we will use AI to finally replace the inefficiencies we have learned to accept as normal.