The new BLS 2025–35 projections show office and administrative support occupations shedding 752,100 jobs — the largest loss of any major occupational group. BLS says continued workflow automation, including AI tools, is likely to reduce demand for several occupations in the group.

AI-related pressure is now part of the official U.S. long-run employment baseline, not only a finding in vendor or academic exposure models. The projection still does not prove that AI alone will cause the decline; BLS incorporates several structural and economic forces.

Occupation-level employment, hiring, wages and openings for customer service representatives, data entry keyers and administrative assistants; annual projection revisions; and the new BLS AI exposure categories.

BLS combined three theoretical measures with two observed-use measures from Anthropic and Microsoft, then classified all 831 detailed occupations as Low, Moderate, High or Very high relative AI exposure.

The official federal benchmark reduces dependence on any single model provider and makes it possible to compare AI exposure with BLS employment, wage and projection evidence using the same occupation framework.

Whether highly exposed occupations diverge in employment, hiring, wages and openings; how BLS revises the categories with new evidence; and the next 2026–36 release in 2027.

Stanford's revised payroll analysis puts employment for workers ages 22–25 in highly AI-exposed occupations about 19% below the path of comparable less-exposed roles, up from a 15% gap in the earlier data vintage.

The adjustment is showing up mainly through weaker hiring of younger workers rather than broad economy-wide displacement. That makes entry-level access a more useful early-warning signal than headline layoff counts alone.

Whether the hiring gap keeps widening, stabilizes, or spreads beyond early-career workers — and whether automation-heavy occupations continue to diverge from augmentation-heavy roles.

The U.S. administration finalized a voluntary testing initiative with major AI developers focused on the cybersecurity capabilities of advanced models.

As models become better at offensive and defensive cyber tasks, capability testing itself becomes an early indicator of whether AI-driven cyber risk is moving from theoretical concern toward operational policy.

Which capabilities are tested, whether results become public, whether participation remains voluntary and whether testing changes deployment or procurement rules.