“AI will disrupt 50% of entry-level white-collar jobs over 1–5 years,” Dario Amodei warned in an essay previewed by Axios. He has framed the coming wave even more dramatically on Anthropic’s own site as a “country of geniuses in a datacenter.” That warning deserves attention, but it does not deserve blind acceptance.
The question that matters is not whether AI can draft a memo, summarize a brief, or impress investors in a demo. The question is whether it can reliably take over real paid work from start to finish, in the format a client or manager would accept, without a human quietly rescuing the result.
On that question, the evidence is still much more sobering than the rhetoric.
Benchmarks Are Not Payroll Systems
The most important new evidence comes from the Remote Labor Index, a benchmark built from actual freelance projects spanning design, analysis, coding, architecture, and other forms of remote work. This is the right test for worker replacement because it measures end-to-end deliverables against human work, not isolated tricks.
The result is brutal for the replacement thesis. The best-performing agent automated only 2.5% of projects, while GPT-5 reached 1.7%. In other words, frontier systems are still failing on more than 97% of the kind of computer-based work they are supposedly poised to erase.
What sank them was not some mystical human essence; it was the ordinary messiness of real work. The paper’s qualitative analysis found corrupt files, incomplete outputs, weak quality, and glaring inconsistencies across deliverables.
The models did succeed in a narrow band of tasks, especially audio editing, image work, writing, and data retrieval. That matters. Those are commercially valuable capabilities. But a technology that shines in islands of work while collapsing on end-to-end projects is a productivity tool, not a replacement worker.
Most of the Gains Still Need a Human in the Loop
Even Anthropic’s own latest labor-market research makes the same point in a different way. The company finds that actual workplace use is still far below theoretical capability. In the computer and math category, Claude currently covers just 33% of tasks, and 30% of workers are in occupations with zero measured coverage.
That is not what a mature replacement wave looks like. It is what an early, uneven adoption curve looks like.
The strongest workplace evidence also points toward augmentation. In the widely cited NBER study of 5,179 customer support agents, access to generative AI raised productivity by 14% overall and by 34% for novice and low-skilled workers, with much smaller gains for experienced staff. That pattern is revealing.
AI is often best at spreading the habits of top performers, speeding up routine tasks, and helping less experienced workers close gaps. It raises the floor before it removes the building. That is a powerful economic effect, but it is very different from handing the whole job to the machine.
Replacing Workers Means Replacing the System Around Them
Jobs are bundles of tasks as much as they are bundles of accountability, exception handling, judgment, trust, coordination, and institutional memory. The Remote Labor Index actually makes the case against near-term replacement even stronger because it leaves out important parts of real work, including many tasks involving client interaction, tutoring, project management, and team coordination.
If frontier models still perform near the floor on a cleaner, more self-contained slice of remote labor, talk of broad worker replacement inside actual firms is premature.
That caution shows up in the macro data too. The Budget Lab at Yale finds no sign that current measures of exposure, automation, or augmentation are yet translating into economy-wide employment or unemployment effects. Brookings reaches a similar conclusion: stability, not disruption, is the story so far.
And the Bureau of Labor Statistics explicitly notes that technological employment effects tend to arrive more gradually than technologists predict. None of that means the threat is fake; it means the threat is being confused with the timetable.
Amodei may yet prove right about the direction of travel. Entry-level work in law, finance, administration, coding, and customer service is plainly under pressure. But pressure is not replacement, and capability curves are not deployment realities.
Today’s AI can already reshape jobs, narrow hiring funnels, and make some tasks cheaper. What it still cannot do, at anything like an economy-wide scale, is replace workers across the unruly range of real professional work.
The immediate danger is not that AI is already ready…it is that executives may start acting as if it is.
