Anthropic Employees Expect AI to Handle 75% of Their Work on Average by 2030

TEXT START: Employed U.S. respondents put AI’s expected capacity for their current work at 28% on average; 5% thought it could handle at least 80% of their jobs.

The Dissection

The text is a labor-displacement warning disguised as a survey report. Its central signal is the perception gap: Anthropic employees expect AI to handle 75% of their work by 2030, while employed U.S. respondents expect 28%. The people closest to frontier capability anticipate far deeper substitution than the population whose livelihoods are exposed to it.

The caveats establish that this is a forecast of capability, not realized automation. The headcount-growth statistic functions as a softening device: it presents expansion during the buildout phase as if it complicates the displacement thesis. It does not. A company can expand while scaling AI infrastructure, then require fewer humans per unit of output.

The Core Fallacy

The text’s main conceptual error is treating self-reported capability forecasts as if they were a labor-market model. “AI can handle 75% of the work” does not specify deployment cost, reliability, liability, coordination, ownership, or whether institutions will continue paying humans to perform automatable tasks.

That limitation cuts both ways. The survey does not prove that 75% of jobs will disappear by 2030. But the headcount increase does not refute displacement either. Under DT mechanics, capability superiority is the trigger; institutional and commercial adoption are lag variables. The article reports the trigger while pretending the lag is the story.

Hidden Assumptions

• A percentage of tasks can be translated cleanly into a percentage of jobs or wages.
• Employee forecasts are calibrated rather than shaped by hype, fear, or insider information.
• Human institutions can preserve large human-only economic domains once AI is cheaper and more capable.
• Anthropic’s expected headcount growth represents durable demand for human labor rather than temporary expansion during an automation arms race.
• Unemployment expectations capture the full effect, while ownership concentration, wage compression, and loss of bargaining power remain secondary.
• Technical capability will diffuse gradually enough for existing labor arrangements to adapt.
• The current wage-to-consumption circuit can survive after productive participation collapses.

Social Function

Classification: partial truth, transition management, and ideological anesthetic.

The article surfaces an uncomfortable fact—frontier AI workers expect much greater automation than ordinary workers—but packages it as polling methodology and corporate growth data. This allows the audience to register the danger without confronting its systemic implication: if AI severs the mass employment-to-consumption circuit, transfers may preserve purchasing power while leaving most people economically unnecessary.

It is not pure propaganda. The data, as presented, are useful evidence of expectation asymmetry. But the framing converts a structural warning into a neutral workplace statistic and lets temporary expansion masquerade as evidence of continued human indispensability.

The Verdict

This is an early-warning document, not proof of a 2030 collapse date. It strengthens P1: the people building cognitive automation expect it to absorb most of their own work. It does not disprove P2 or P3; it merely shows the lag before those mechanisms become socially visible.

The decisive fact is not the exact 75% figure. It is that the frontier workforce already sees its own labor as highly substitutable while the broader workforce remains materially less aware. That is how obsolescence begins: first as an insider forecast, then as a cost curve, then as a labor-market corpse.