
The AI Expectations Gap
Szymon Sacher, Tess Cotter, Charles I. Jones, Anton Korinek, Peter McCrory
以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。
Szymon Sacher, Tess Cotter, Charles I. Jones, Anton Korinek, Peter McCrory
How will AI reshape the economy in the years ahead? AI capabilities are advancing rapidly, and the possibility that future AI models may be able to build their own successors (recursive self-improvement) could further accelerate the rate of progress. How quickly this might materialize and to what extent is unclear. Forecasts vary widely and are fiercely debated.
What is the core source of disagreement about AI's future impact on the economy? Is it principally due to differences in beliefs about how quickly AI capabilities will improve? Or does it reflect disagreement about how quickly and in what form those capabilities translate to real-world adoption? Which views on the future path of AI are consistent with rapid economic growth?
This note presents new survey evidence on these questions. In August 2026, we surveyed two groups: a sample of US adults that was nationally representative and Anthropic staff.1 The questions covered, among other things, AI’s capabilities, how widely it will be used, how much time it saves, how autonomous it will be, and how difficult it will be for a displaced worker to find a job in a different occupation.2 Anthropic staff build and use frontier AI every day. The second survey therefore shows how people closest to the technology anticipate its future, both in capabilities and in broader economic impact.
We report the following main findings. First, the two groups largely agree about what AI can do today but disagree about the pace of future progress. Anthropic staff expect capabilities to advance faster, AI to be used more widely and with less human involvement, and the time it saves on a task to be much larger. For example, 91% of staff expect AI to be able to make a Nobel-level discovery by 2030, compared to 39% of US adults.
Second, there are sharp differences concerning people's own jobs. On average, staff members expect AI to be able to do 75% of their current work by 2030, compared to 28% for US workers overall.
Third, the two groups seem to reason about the labor market in a similar way: people who expect AI to do more of their own work in the future forecast higher unemployment. And, at the same level of expected AI automation of their own work, both groups forecast about the same unemployment rate.
We first investigate expectations about AI capabilities. We asked when AI will be able to do each of eight tasks as well as a skilled professional, from writing routine emails to making a Nobel-level scientific discovery. US adults and Anthropic staff largely agree on what AI can do today, particularly when it comes to more complex tasks. Both groups rank tasks in nearly the same order of difficulty (Figure 1, panel A).
The groups disagree on the path of future AI progress. For example, around a quarter of US adults and Anthropic staff think that AI can design, conduct, and publish scientific studies today. But by next year, 9 in 10 staff expect this to be feasible for AI as compared to just 38% for the general public. By 2030, 39% of US adults expect AI to be able to make a Nobel-level discovery (panel C), compared with 91% of staff. What is perhaps the most striking difference, depending on the task, between 19 and 40% of US adults say AI will never be able to do a given task, compared with at most 2% of staff.
In part, differences in views reflect differences in frequency of AI use. Among those who use AI at work regularly (about a fifth of respondents), 56% expect AI to be able to do at least seven of the eight named tasks by 2030. This is substantially higher than among all US adults (41%) and among those who have never used an AI tool (28%), but still well short of the staff view (96%).
The impact of AI on the economy is not solely determined by raw capabilities. People and businesses need to actually use AI, which takes time and experimentation. Of the work tasks AI will be able to do, US adults expect that 42% will be done with AI by 2030. Anthropic staff expect noticeably faster diffusion over the next few years, with AI used for 65% of all tasks for which it could be used.
To understand what might hold back AI adoption at work, we asked US adults to choose up to three obstacles from a list of nine that they expect to be the biggest. The general public expects that people's preferences and rules, more than the technological constraints, may be what holds adoption back (Figure 2). “Customers preferring to deal with a person” was the obstacle selected most often, followed by “data privacy and security concerns,” “AI not being reliable enough,” and “laws, regulations or liability.”
Notably, US adults who use AI regularly at work rank data privacy and security concerns first (42% consider it among the top 3 obstacles, compared with 37% among the rest) and see customer preferences to deal with humans as slightly less important (37% vs 42%).
The impact of AI on the labor market will in part depend on whether AI completes complex tasks autonomously or in collaboration with a human. The former tends to displace human workers. Alternatively, when AI complements human capabilities, workers become more productive. To get a sense for anticipated automation versus augmentation in 2030 among Anthropic staff and the general public, we asked how AI and humans will work together on various tasks ranging from email drafting to legal contracting by 2030. At Anthropic, the vast majority of staff say that by 2030 AI will either complete tasks alone or with a worker evaluating the work (Figure 3). By contrast, 4 in 10 US adults instead expect that humans will do most of the work themselves or with some AI help.
How much AI raises output depends not only on which tasks it can do and how widely it is used, but also on how much faster and better it does them. A technology that is just good enough to replace workers—what Acemoglu and Restrepo (2019) call a "so-so" technology—displaces work without much increase in GDP. By contrast, a technology that completes tasks at a small fraction of cost or in a fraction of time can meaningfully raise GDP.
To measure perceptions about the productivity lift from AI, we asked the general public and Anthropic staff how long a task for which AI is well suited takes with AI, compared with doing it without AI. A task done in half the time corresponds to a doubling of output per hour.3
Anthropic staff report much larger expected savings, with 85% saying that it takes half the time or less. Among US adults who use AI at work, the typical answer is half the time, and 29% say the task takes just as long or longer, while only 5% of staff say so. Among US adults, 26% of those who use AI at work say a task takes a quarter of the time or less, and 21% of those who do not use AI at work say the same. For staff, that share is 59% (Figure 4).
Staff may be describing what they experience at work. As discussed further in How Employment is Changing at Anthropic, Anthropic employees increased productivity on some dimensions after adopting agentic tools: the lines of code engineers added tripled in the month after adopting coding agents and sales representatives closed more deals with clients after adopting AI agents. Output per person also rose in other business functions, like legal and recruiting.
Perhaps unsurprisingly, we find a large gap in perceptions between the general public and Anthropic staff when it comes to AI’s impact on their own jobs (Figure 5). On average, Anthropic staff members say AI will be able to do 75% of their current work hours by 2030. US workers (respondents who report having a job) put the share at 28% on average, and only 1 in 20 say that AI could do 80% or more of their work hours by 2030.
Across US workers, expectations vary by occupation: AI is expected to be able to do 25% of work hours among manual trades, production and transport workers and 33% for workers in computer, engineering and science occupations. For this latter group, those who use AI regularly expect only 40%, a belief that is closer to the average US worker than to the average Anthropic staff member (75%).
How much of their own work people expect AI to do goes a long way toward explaining what they expect for the labor market. We told respondents that the current unemployment rate was about 4.2% and asked them to forecast the rate in 2030. Figure 6 groups respondents by the share of their own hours they expect AI to be able to do by 2030, and shows each group's average unemployment rate forecast.
US workers who expect AI to be able to do less than half of their work forecast unemployment of 7 to 8%. Above half, the forecast increases with the share of AI-exposed work. Those who expect AI to do 80% or more of their work expect, on average, an unemployment rate of 12.2% in 2030, above its peak during the Great Financial Crisis. Across US workers, ten more percentage points of work affected is associated with 0.4 points more forecasted unemployment. General optimism about AI does not explain this. Holding their own work exposure fixed, the share of the eight tasks a worker expects AI to master is unrelated to the forecast.
Even though their own reported exposure is much higher, the same gradient appears among Anthropic staff, with forecasts closely mirroring those of the general population. Therefore, these two groups do not seem to hold different theories of how AI will simultaneously affect their own jobs and the broader labor market. They have different experiences of what AI capabilities are and, in turn, expectations.
Even though Anthropic staff expect AI to be able to do most of their work, they also expect Anthropic to keep hiring. 95% of Anthropic staff expect headcount to grow over the next 12 months, the median staff member expects it to grow by 66%, and 27% expect it at least to double. These are the personal expectations of staff members in an internal poll and do not represent Anthropic's plans or its view of its headcount. Even so, they show that automating tasks and shedding jobs are different forecasts, at least at a frontier AI lab.
Anthropic staff, however, are slightly more pessimistic than US adults about how quickly workers displaced by AI can find new work. We asked how long it would take a person who loses their job to AI automation in 2030—and who needs to switch occupations—to find a new one.4 The median Anthropic staff member expects that a worker displaced by AI would take 12 months to find a role in a new occupation. This compares to the median US adult response in the range of 7-12 months.
In Economic Scenarios for Transformative AI we introduced a structured framework for assessing how different future paths for AI capabilities and diffusion might translate into differences in economic outcomes. The model allows us to translate each respondent’s answers across five dimensions — capabilities, diffusion, automation, productivity, and worker transition frictions — into model-implied economic outcomes.5 Across both our US adult and Anthropic staff samples, we identify every respondent who answered all questions related to the key model inputs (3,259 US adults and 1,093 staff). We then evaluate the economic scenario implied by those answers for each respondent, focusing below on implications for GDP growth in 2030.
The median US adult's answers imply GDP growth of 5.4% in 2030, according to the model. The median staff member's answers imply GDP growth accelerating to 17.1% by 2030. There is significant variation within each group, so the distributions overlap widely (Figure 7). Figure A1 in the appendix shows how much each belief contributes to this gap.
Figure 7 also shows the economic scenarios implied by inputs provided by site visitors to an interactive scenario explorer based on this economic model that was published at www.anthropic.com/institute/econ-scenarios on September 9,
正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。