Anthropic researchers surveyed 10,980 nationally representative US adults and Anthropic staff in August 2026 about AI capabilities, workplace adoption, productivity, job automation and unemployment expectations. The groups broadly agreed on what AI can do today but differed over how fast it will advance: 91% of staff expected AI to make a Nobel-level discovery by 2030, compared with 39% of US adults. By next year, 90% of staff expected AI to design, conduct and publish scientific studies, versus 38% of the general public. Depending on the task, 19% to 40% of US adults said AI would never be able to do it; no more than 2% of staff said the same. Among adults who regularly used AI at work, 56% expected it to perform at least seven of eight specified tasks by 2030, compared with 41% of all adults, 28% of people who had never used an AI tool and 96% of staff.

Respondents expected different rates of workplace adoption even when AI could perform a task. US adults expected AI to be used for 42% of tasks it could do by 2030; Anthropic staff expected 65%. Asked to choose up to three obstacles to workplace use, US adults most often selected customers preferring to deal with a person, followed by data privacy and security concerns, AI reliability, and laws, regulations or liability. Among adults who regularly used AI at work, 42% ranked privacy and security among the top three obstacles, compared with 37% of other adults; 37% ranked customer preference for human service, versus 42% of other adults. Staff were not asked this obstacles question. On how work would be done, most staff expected AI to complete tasks alone or with a worker checking its work by 2030, while 40% of US adults expected workers to do most of the work, either alone or with AI assistance.

Staff also anticipated larger time savings from AI on tasks suited to it. 85% said such a task would take half the time or less, while 5% said it would take as long or longer. Among US adults using AI at work, 26% said a task took a quarter of the time or less; 21% of those not using AI at work gave that answer, compared with 59% of staff. The researchers cited a separate Anthropic account reporting that engineers tripled lines of code added in the month after adopting coding agents, while sales representatives closed more client deals after adopting AI agents; output per person also rose in legal and recruiting.

On their own jobs, Anthropic staff expected AI to perform an average of 75% of their current work hours by 2030, compared with 28% among employed US respondents. Only 1 in 20 US workers expected AI to perform at least 80% of their hours. Estimates varied by occupation: 25% for manual trades, production and transport workers, and 33% for computer, engineering and science workers; regular AI users in that latter group expected 40%. The survey told respondents the unemployment rate was about 4.2% and asked for their 2030 forecast. US workers who expected AI to handle less than half their work forecast unemployment of 7% to 8%; those expecting AI to handle at least 80% forecast 12.2%. Across US workers, each additional 10 percentage points of expected work exposure was associated with a 0.4-point higher unemployment forecast. Staff showed a similar relationship, and expectations about how many of eight tasks AI could master were unrelated to forecasts when respondents' own work exposure was held constant.

Despite expecting AI to perform much of their work, 95% of Anthropic staff expected the company's headcount to grow over the next 12 months. The median staff member expected growth of 66%, and 27% expected headcount to at least double. The report says these are staff members' personal expectations from an internal poll, not Anthropic's plans or view. Staff expected a displaced worker who had to change occupations to take a median 12 months to find another job; the median US adult answer was 7–12 months.

Anthropic researchers surveyed 10,980 nationally representative US adults and Anthropic staff in August 2026 about AI capabilities, workplace adoption, productivity, job automation and unemployment expectations. The groups broadly agreed on what AI can do today but differed over how fast it will advance: 91% of staff expected AI to make a Nobel-level discovery by 2030, compared with 39% of US adults. By next year, 90% of staff expected AI to design, conduct and publish scientific studies, versus 38% of the general public. Depending on the task, 19% to 40% of US adults said AI would never be able to do it; no more than 2% of staff said the same. Among adults who regularly used AI at work, 56% expected it to perform at least seven of eight specified tasks by 2030, compared with 41% of all adults, 28% of people who had never used an AI tool and 96% of staff.

Respondents expected different rates of workplace adoption even when AI could perform a task. US adults expected AI to be used for 42% of tasks it could do by 2030; Anthropic staff expected 65%. Asked to choose up to three obstacles to workplace use, US adults most often selected customers preferring to deal with a person, followed by data privacy and security concerns, AI reliability, and laws, regulations or liability. Among adults who regularly used AI at work, 42% ranked privacy and security among the top three obstacles, compared with 37% of other adults; 37% ranked customer preference for human service, versus 42% of other adults. Staff were not asked this obstacles question. On how work would be done, most staff expected AI to complete tasks alone or with a worker checking its work by 2030, while 40% of US adults expected workers to do most of the work, either alone or with AI assistance.

Staff also anticipated larger time savings from AI on tasks suited to it. 85% said such a task would take half the time or less, while 5% said it would take as long or longer. Among US adults using AI at work, 26% said a task took a quarter of the time or less; 21% of those not using AI at work gave that answer, compared with 59% of staff. The researchers cited a separate Anthropic account reporting that engineers tripled lines of code added in the month after adopting coding agents, while sales representatives closed more client deals after adopting AI agents; output per person also rose in legal and recruiting.

On their own jobs, Anthropic staff expected AI to perform an average of 75% of their current work hours by 2030, compared with 28% among employed US respondents. Only 1 in 20 US workers expected AI to perform at least 80% of their hours. Estimates varied by occupation: 25% for manual trades, production and transport workers, and 33% for computer, engineering and science workers; regular AI users in that latter group expected 40%. The survey told respondents the unemployment rate was about 4.2% and asked for their 2030 forecast. US workers who expected AI to handle less than half their work forecast unemployment of 7% to 8%; those expecting AI to handle at least 80% forecast 12.2%. Across US workers, each additional 10 percentage points of expected work exposure was associated with a 0.4-point higher unemployment forecast. Staff showed a similar relationship, and expectations about how many of eight tasks AI could master were unrelated to forecasts when respondents' own work exposure was held constant.

Despite expecting AI to perform much of their work, 95% of Anthropic staff expected the company's headcount to grow over the next 12 months. The median staff member expected growth of 66%, and 27% expected headcount to at least double. The report says these are staff members' personal expectations from an internal poll, not Anthropic's plans or view. Staff expected a displaced worker who had to change occupations to take a median 12 months to find another job; the median US adult answer was 7–12 months.