← Back to feed

Anthropic published research on September 30 that maps current robot capabilities against every physical job task in the US economy, using Claude to rate 19,000 tasks from O*NET’s database of roughly 900 occupations. The methodology scores each task on how well robots perform in controlled settings versus the unstructured environments actual workers face.

Key Numbers

• 74% of physical US job tasks can be performed by robots today, representing 34% of total working hours
• Robots and large language models together expose 80% of employment by working time
• The remaining 20% is highly interpersonal or requires physical dexterity current robots lack
• Robots are cost-competitive with human workers on 0.3% of tasks today
• At historical robot price-decline rates, that share reaches 10% in 40 years
• Robots gain the ability to perform roughly 2% more physical work annually

Who Is Exposed Now

Exposure is concentrated among male, less-educated, and lower-paid workers. Driving and warehouse jobs are highly exposed to available robots today. Nursing and general repair are not — robots can do little of those tasks even in highly controlled environments. The gap is largely about unstructured physical environments and fine dexterity, not raw compute.

Barriers Beyond Cost

After cost, the main adoption blockers are physical capability (fine dexterity, such as untangling wires), human preferences, and regulations that restrict automation for a significant share of tasks. Most commercially available robots still require highly structured environments to function reliably.

Historical Validation

The methodology is backtested against 50 years of labor data. Jobs with high robot exposure from 1977 onward showed greater wage and employment declines in subsequent decades compared to low-exposure jobs. The research concludes that current robot capability is a reliable leading indicator of future labor displacement: taxi drivers and warehouse workers face automation before nurses and mechanics.