Anthropic survey: employees expect AI to be able to complete 75% of their own work by 2030, while the public expects only 28%

[[**An Anthropic survey shows that its employees expect AI to be able to complete 75% of their own work tasks by 2030, while the American public expects only 28%. The two groups are close in their judgments of AI's current capabilities, but their expectations for the pace of technological development and the scope of enterprise adoption have diverged sharply. AI practitioners' judgments about future economic impact are significantly more aggressive.**

**Key points**

**1. The gap between AI practitioners and the public in expectations for AI capabilities in 2030 is enormous**

Among Anthropic employees, 90% expect that by 2027 AI will be able to handle designing and publishing scientific research tasks (only 38% of the public); 91% believe that by 2030 AI will make Nobel Prize-level discoveries (only 39% of the public). The two groups are broadly consistent in their judgments of AI's current capabilities.

**2. The expected replacement rate of AI for personal work differs by nearly three times**

Anthropic employees on average expect that by 2030 AI can complete tasks corresponding to 75% of their current working time, while the average expectation of employed Americans is only 28%. The more people believe their own work can be replaced by AI, the more they tend to predict higher unemployment; the two groups are consistent in this reasoning pattern.

**3. The public that frequently uses AI has expectations closer to internal employees**

Among the roughly one-fifth of Americans who frequently use AI at work, 56% believe that by 2030 AI will be able to complete at least seven of eight tasks; the proportion among all American adults is 41%, and among those who have never used AI it is 28%.

**4. Task automation and company expansion can coexist**

Although Anthropic employees expect AI to take on most work, 95% still predict that the company's total headcount will increase over the next 12 months, with a median expected growth of 66%. Companies can expand their business due to efficiency gains, but entry-level positions for newcomers may narrow—Anthropic's internal data shows that the share of junior technical employees has already fallen from 21% to 14%. 2026-10-20 13:59 The 2030 seen by employees of AI companies is completely different from that of ordinary people. Cold Eye Quick View

If a young engineer who has just joined the workforce opens a job-hunting app, what he cares about most may be: how many junior positions are still available this year? Yet in a frontier AI lab in Silicon Valley, engineers are having agents write code, check vulnerabilities, and perform some of the tasks they used to have to do themselves. The two people may be using similar AI technology, but their feelings about work are not necessarily the same.

Anthropic published "The AI Expectations Gap" on October 16, 2026, turning this difference into a set of comparable data. In August of that year, the researchers surveyed a nationally representative sample of American adults, as well as Anthropic employees who develop and use frontier models every day. They did not ask broad questions like "Do you like AI?" but instead asked: what exactly can AI do, will companies use it, how much time can it save, can it work independently, and how long will it take for displaced people to find their next job.

The result is not simply "insiders are optimistic, outsiders are pessimistic." The two groups' judgments about what AI can do today are actually not far apart; what really separates them is the next four years. Anthropic employees on average expect that by 2030 AI can complete tasks corresponding to 75% of their current working time, while the average expectation of employed Americans is only 28%. But under the same job automation expectations, the two groups' judgments about the unemployment rate are again very close. The issue is not who understands the economy better, but that the technological world they see is already different.

The following is the complete translation of the original report:

AI capabilities are advancing rapidly. Future models may even increasingly participate in developing their own successor models, the so-called "recursive self-improvement," thereby further accelerating technological progress. But when this change will occur and how large its impact will be, no one can give a definite answer at present. Predictions vary widely, and the debate is intense.

The researchers wanted to clarify a more basic question: do people's disagreements about AI's economic impact come from different judgments about the pace of technological progress, or from different understandings of the pace of enterprise adoption and the ways it is applied? If AI really makes the economy grow rapidly, what conditions need to be met?

To this end, the team conducted a survey in August 2026 targeting two groups: one was a nationally representative sample of American adults; the other was Anthropic employees. The survey covered questions such as capabilities, adoption, time savings, degree of autonomy, and difficulty of job transitions. Because the latter develop and use frontier AI every day, their answers can show how the group closest to the technology imagines the future.

Three main findings are worth stating first: First, the two groups are broadly consistent in their judgments of AI's current capabilities, but Anthropic employees expect future capabilities to develop faster, with broader application scope, less human involvement, and greater time savings. Second, at the level of personal work, the expectation gap is especially obvious: 75% versus 28%. Third, the two groups reason similarly about employment—respondents who more believe AI can take on their own work are more inclined to predict higher unemployment; when the two groups believe their work has the same degree of automation, their unemployment rate predictions are roughly comparable.

III. AI in 2030, as if from two worlds

The researchers listed eight tasks, from writing a routine email, bookkeeping, filing taxes, drafting contracts, developing software products, running an online business, to designing and publishing scientific research, and making Nobel Prize-level scientific discoveries, and asked when AI could reach the level of a skilled professional.

Regarding the present, the two groups give almost identical difficulty rankings, especially for those complex tasks. For example, about a quarter of American adults and Anthropic employees believe AI can currently design, execute, and publish scientific research. The real gap begins to appear from "next year" and "2030": by 2027, 90% of Anthropic employees expect AI to be capable of such research tasks, while only 38% of the American public do; by 2030, 91% of employees believe AI can make Nobel-level discoveries, compared with 39% of the American public.

In addition, for different tasks, 19%–40% of American adults believe AI "will never be able to do it"; among Anthropic employees, at most 2% hold this judgment.

Figure 1 | The numbers represent the proportion of respondents who "believe AI can reach the level of a skilled professional"; they do not indicate that models have actually passed these tasks, nor that companies will necessarily use them. American public responses are weighted for national representativeness, original sample N=10,980; employee responses are unweighted. Source: Anthropic Institute, October 16, 2026.

Frequency of contact with AI seems to explain part of the difference. In the American public sample, about one fifth use AI frequently at work; among this group, 56% believe that by 2030 AI will be able to complete at least seven of eight tasks. For all American adults it is 41%, and for those who have never used AI tools it is 28%, but even frequent users are far below the 96% of Anthropic employees.

For work that AI is technically capable of doing, American adults expect that by 2030 42% of it will actually use AI; Anthropic employees expect this proportion to be 65%. Both groups agree that AI will continue to enter the workplace, but they differ on the speed of diffusion.

To understand why companies might not use AI, the research team asked American adults to choose up to three from nine categories of obstacles. Ranked first is not computing power or model capability, but "customers prefer to deal with real people," followed by data privacy and security, AI not being reliable enough, and legal, regulatory, and liability issues. That technology can complete a task does not mean customers are willing to hand insurance claims, medical consultations, or financial disputes entirely to machines. The scenarios here are examples to help readers understand, not independent cases from the survey.

Figure 2 | The proportions are the share of American adults who listed that item among the "top three obstacles"; multiple selections were allowed, so the total does not equal 100%; this question was not asked of Anthropic employees. Source: Anthropic Institute.

American adults who frequently use AI at work are more likely to see privacy and security as an obstacle: 42% of them listed it in the top three, versus 37% of other respondents. Conversely, they place slightly less importance on "customers prefer real people": 37% versus 42%. Frequent AI users may be more familiar with its actual limitations, but this is still a survey correlation and cannot prove that using AI changed someone's views.

If AI takes a complex job from start to finish, people may be replaced; if AI only helps look up information, draft, and organize, enabling employees to complete more tasks, it may also become an auxiliary tool that boosts productivity. These two uses have different effects on employment.

Researchers asked respondents to imagine five types of tasks in 2030, assume that AI is already useful for these tasks, and then judge who will ultimately perform them: entirely by AI independently; AI completes the main part and a person checks it; a person completes the main part with AI assistance; or entirely by a person.

Among Anthropic employees, the vast majority expect to fall into the first two situations by then—AI does it independently, or AI does it and a person reviews it. About four in ten American adults still expect that people will complete most of the work, or use only a little AI assistance.

Figure 3 | Shows the proportion of respondents choosing "AI completes independently" or "AI completes the main part, with human review." This is not the actual automation completion rate of AI. Source: Anthropic Institute.

If a job is merely switched to AI without saving much cost or time, then it may change the employment structure but may not significantly increase social output. Economists Daron Acemoglu and Pascual Restrepo once called this type of limited-effect automation "so-so technology," roughly meaning a technology that is "barely good enough." In contrast, if AI can complete the same task with only a small fraction of the original time or cost, the space for macroeconomic benefits will expand significantly.

Researchers asked respondents to choose a task that today's AI is particularly suited for, such as summarizing documents or drafting emails, and to compare the time needed to complete the task with and without AI. The calculation needs to include the time to check and revise AI results, not just the few seconds for the model to generate text.

Among Anthropic employees, 85% said that using AI to complete such tasks takes no more than half the original time; 59% said it takes only a quarter or less of the original time. Among American adults who frequently use AI at work, the typical answer is "half the time," with 26% saying it can be shortened to a quarter or less; 29% say the time is not reduced, or even longer. Only 5% of Anthropic employees answered this way. For American adults who do not use AI at work, 21% estimate that such tasks can be shortened to a quarter or less.

Figure 4 | The survey asked about specific tasks for which AI is suitable today and required including time for human checking and error correction; it does not mean AI makes all work four times more efficient. Source: Anthropic Institute.

The researchers believe that Anthropic employees' answers may come partly from firsthand experience. Another internal company employment study released at the same time found that within one month of introducing coding agents, engineers' new lines of code reached three times the previous level; after sales staff introduced agents, they closed more customer deals, and some quantifiable output in roles such as legal and recruiting also increased. A reminder is necessary here: lines of code do not equal code quality, nor do they equal business profit, and observations from a single company are not enough to represent all enterprises.

VII. Facing one's own job, the gap between 75% and 28% is the most obvious

When the question shifts from "what capabilities will AI have" to "how much of my work can AI do for me," the divergence widens further. Anthropic employees on average believe that by 2030, AI will be able to complete 75% of the tasks corresponding to their current working hours; the average judgment of employed Americans is only 28%. Among surveyed US workers, only about 5% believe AI can complete more than 80% of their work.

There are also differences among different occupations in the US public: workers in manual technical, production, and transportation roles on average expect AI to take on 25% of working hours; computer, engineering, and science occupations, 33%. Even among the latter, those who frequently use AI expect only 40%, far below the 75% of Anthropic employees.

Figure 5 | The proportion of respondents' working hours that they expect AI to be able to complete by 2030. The proportion of respondents' current working hours that AI is expected to be able to complete by 2030. The US adult sample is limited to employed respondents. Workers answered via a slider, while US adults chose a range, and we used the midpoint of the range. Bar charts show 95% confidence intervals. Source: Anthropic Institute.

There is an easily misread point here: "75% of work can be done by AI" does not equal "75% of people will lose their jobs." An editor may have different tasks in a day, such as finding materials, fact-checking, interviewing, writing drafts, and communicating. Even if most of these can be handled by AI, roles responsible for judgment, gatekeeping, and bearing consequences may still exist. Whether an industry can retain the same number of people after task automation is another question that requires real employment data to answer.

The researchers told respondents that at the time of the survey the US unemployment rate was about 4.2%, and then asked them to predict the unemployment rate in 2030. The study grouped people by "how much of their own working time they expect AI to be able to complete" and compared the average judgments of different groups.

Among surveyed US workers, those who believe AI can take on less than half of their work on average expect a 2030 unemployment rate of 7%–8%. After exceeding half, as the expected share AI can take on rises, the predicted unemployment rate also gradually increases. Those who believe AI can complete 80% or more of their work on average expect the 2030 unemployment rate to reach 12.2%, higher than the peak level once seen in the US during the financial crisis.

The statistical relationship shows that among employed Americans, for every 10 percentage point increase in the share of work an individual expects to be affected by AI, their prediction for the 2030 unemployment rate is on average 0.4 percentage points higher. The study further found that a general belief that AI can solve more tasks has no such connection with unemployment rate judgments; what is truly closely associated with unemployment expectations is how much impact a person believes their own specific job will receive. A similar relationship also appears among Anthropic employees.

Figure 6: Average predicted 2030 unemployment rate by the proportion of respondents' working hours that they expect AI to be able to complete. Respondents were told that the current US unemployment rate was 4.2% and were asked what they expected the unemployment rate to be in 2030. Panel (A) groups employed respondents by the proportion of their current working hours that they expect AI to be able to complete by 2030, and shows the average predicted value and 95% confidence interval for each group. Groups with fewer than 20 respondents are omitted. Panel (B) shows the share of US workers and Anthropic employees in each group. Source: Anthropic Institute.

IX. Work can be automated, but companies may still continue hiring

Surprisingly, although Anthropic employees expect AI to take on most of their work, they do not generally believe the company will shrink its workforce. 95% of surveyed employees expect Anthropic's total headcount to increase over the next 12 months; the median expectation among employees who gave a forecast is growth of 66%, and 27% expect headcount to at least double.

The original report specifically notes: this is only a personal prediction in an internal employee poll in August 2026, not Anthropic's official hiring plan, and even less a company commitment. What it reveals is an idea: automating certain tasks and the company cutting a corresponding proportion of positions are not the same proposition. Enterprises may expand through stronger production capacity, and may also reduce hiring for specific roles; the two can coexist.

But for people who have already lost jobs, the next job is not so easy to find. The study asked respondents: if in 2030 someone loses their position because of AI automation and has to change careers, how long would it take to find work in a new career? The median time given by Anthropic employees is 12 months; the median answer from US adults falls in the 7–12 month range. The two groups are not asserting that this will definitely be the case in the future, but rather expressing expectations about the difficulty of career transitions.

X. If these expectations are put into economic models, what kind of growth would be derived?

The researchers previously proposed a structured framework in "Economic Scenarios for Transformative AI," mapping five types of judgments about AI capabilities, diffusion speed, degree of automation, productivity gains, and worker transition frictions into different macroeconomic scenarios. In this survey, 3,259 US adults and 1,093 Anthropic employees completed the five key questions, so they can be entered into the model person by person.

The results obtained differ greatly. The median of the US public's answers corresponding to the 2030 GDP annual growth rate scenario was 5.4%; the median of the model scenario corresponding to Anthropic employees' answers was 17.1%. Within both groups, the forecast distributions were wide and overlapped considerably.

In addition, Anthropic launched a public interactive economic scenarios tool on September 9, 2026. In the first 48 hours, 23,075 visitors from more than 140 countries voluntarily filled out a short questionnaire. Plugging their answers into the same framework, the median corresponded to a 12.5% growth scenario, situated between the general public and Anthropic employees. Visitors who proactively went to an AI company's website may have already been more interested in AI, so this cannot be treated as a random sample of the global public.

Figure 7: Model-projected 2030 GDP growth rate based on the expectations of US adults, website visitors, and Anthropic employees. Each respondent's answers to the five questions on capability, adoption, autonomy, time savings, and re-employment were all run through the model in "Economic Scenarios: Transformative AI." Thick bars indicate the 25th to 75th percentiles, and thin lines indicate the 10th to 90th percentiles. Dashed lines mark the three scenarios proposed in the paper: modest (2.4% annual GDP growth), significant (5.4% GDP growth), and extreme (15.4% GDP growth). The figure includes respondents who answered all five questions: 3,259 US adults (2030 half only), 1,093 Anthropic employees, and 23,075 website visitors who completed the anthropic.com/institute/econ-scenarios quiz within 48 hours of its release on September 9, 2026. Source: Anthropic Institute.

Some employees are more optimistic about the future, possibly because they can directly see what is happening in model training and R&D. The Anthropic Institute has previously discussed "recursive self-improvement"—AI systems increasingly helping to improve their own subsequent capabilities.

In the study, employees were asked another question: relative to developing a frontier model at the level of early 2026, how much R&D work today is done by AI? And how much will be by the end of 2027 and the end of 2030? The median judgments given by employees were 51%, 80%, and 95%, respectively.

Differences in views among employees within different teams were even larger than the average differences between teams. Employees directly engaged in model research estimated on average that 60% of current AI R&D work is done by AI; other surveyed employees averaged 53%. This is consistent with the explanation that "people who know more about internal progress may have higher expectations," but the researchers also note that researchers may inherently rely more on AI than other roles.

Figure 8: Anthropic employees' estimates of the share of work in building frontier models that is done by AI. Anthropic employees were asked what share of the work of building an early-2026 frontier model today is done by AI, and what they expect that share to be by the end of 2027 and in 2030. Tick marks indicate median answers, boxes indicate the 25th to 75th percentiles, and lines indicate the 10th to 90th percentiles. These answers are employees' personal estimates in an internal survey, not Anthropic measurements or company forecasts. Source: Anthropic Institute.

Overall, ordinary US adults and Anthropic employees do not differ much in their understanding of AI's current capabilities; but there is substantial disagreement about the pace of technological progress in the coming years, the actual scope of adoption, and the magnitude of declines in task time.

Under the economic scenarios model, the median answers of the US public correspond to an economic scenario significantly above usual growth rates; the answers of Anthropic employees correspond to extremely unusual rapid growth. At the same time, the more people believe AI can complete more of their personal work, the more they tend to predict higher unemployment. This holds in both groups.

The two groups differ significantly in their day-to-day exposure to AI. Many Anthropic employees already collaborate with AI capable of handling a large amount of work content, while the vast majority of workers in other US industries have not yet reached this level. Ordinary respondents who frequently use AI at work already have expectations closer to those of Anthropic employees. This means that as AI truly enters more industries, public views may change; but how large the change will be still depends on whether the ways of working in the labs can be replicated at other companies and in other roles.

What may become difficult first may not be keeping a job, but entering the workforce

Imagine a recent software engineering graduate. He can use AI to build pages, fix simple errors, and has spent considerable time learning the latest tools. But when companies interview him, they want him not only to write code, but also to independently judge product requirements, understand the client's business, and take responsibility for online failures. The problem is: the experience that takes years to accumulate was originally gained by handling basic tasks and constantly correcting mistakes. If companies hand such tasks to AI, where should newcomers complete their first round of training? This is a scenario inference based on job structure, not a specific job seeker interviewed for this article.

This is not just an abstract concern. A study updated by the Stanford Digital Economy Lab on August 12, 2026, using ADP payroll records for millions of US employees, observed an approximately 19% gap in employment performance for young people aged 22–25 in occupations highly exposed to AI, relative to the employment trends of same-age peers in occupations with lower AI exposure. The study also notes that the overall labor market has not yet shown widespread, large-scale job displacement attributable to AI; changes in youth employment are reflected more in slower hiring rather than mass departures of incumbent employees.

The meaning of "19%" needs to be understood carefully. It is not "19% of American graduates are unemployed because of AI," nor is it a 19% year-over-year decline in employment for all young people. It refers to a relative comparison: if employment among young people in occupations highly exposed to AI had followed the same growth trajectory as their peers in occupations with lower exposure, the actual level would be about 19% below that reference path. The Stanford team also explicitly states that these are descriptive relationships, not final estimates of AI's causal effect. Factors such as the economic cycle, industry hiring adjustments, and educational differences may still be at work together.

Echoing this, another independent internal study published by Anthropic on October 16 shows that since early 2025, the share of junior technical staff in the technology department has fallen from 21% to 14%; the share of junior technical positions among newly hired technical staff has fallen from 27% to 9%. At the same time, the company continues to expand. It does not prove that "AI has already left 90% of graduates unable to find work," but it shows a corporate organizational change worth noting: total headcount can grow while career entry points may narrow.

When a customer service center hands common questions to AI, what is most likely to disappear may not be the entire customer service department, but rather the junior tickets that can be resolved independently by following scripts and standard answers. Senior customer service staff still have to handle escalated complaints, special customers, and cross-departmental coordination; newcomers, however, find it harder to gradually build experience through low-risk tickets. Similarly, paralegals may once have started with document retrieval and preliminary contract review, editors with organizing materials and fact-checking, and programmers with testing and small fixes. AI's ability to share these tasks is, in principle, an efficiency gain, but it also forces companies to answer a new training question. The above are examples of the mechanism, not industry layoff statistics from this survey.

This is also why "tasks completed by AI" does not equal "jobs disappearing." Real work often includes explaining things to clients, coordinating with colleagues, taking responsibility, and handling exceptions. An important finding of the Anthropic survey is precisely that the American public considers "customers prefer to deal with real people" to be the biggest barrier to adoption. Interpersonal trust and legal liability do not automatically disappear from business processes just because models become more capable.

There is another side to consider as well. AI-driven efficiency gains may allow companies to take on more orders, develop more new products, and even create occupations that did not previously exist. That 95% of Anthropic employees expect the company to continue expanding is an example of how "task automation and employment growth may coexist." But whether this possibility can be realized across the entire labor market depends on whether new demand is sufficient to offset displaced work, and whether new positions are open to young people.

A company can believe two things at once: AI makes certain processes that once required a hundred people increasingly automated, but because its products sell better and it serves more customers, it still needs to hire in sales, customer support, legal compliance, or new product lines. This is not contradictory. The problem is that the skills required for new positions may be completely different from those for the old ones.

When task execution becomes cheaper, what is truly scarce in organizations may shift toward task definition, outcome validation, complex judgment, and external accountability.

In Anthropic's internal employment report, employees began describing themselves as managing a group of AI assistants that constantly demand attention. The content employees need to read, check, and coordinate has actually increased. Rather than saying broadly that "everyone will need to be an AI manager in the future," it is better to ask specifically: which decisions really require humans? Which steps carry error costs too high to be left to models to complete automatically?

For ordinary workers, this does not mean everyone must switch to researching large models. Understanding the company's business and customer needs, driving things to completion across teams, and making judgments when evidence is incomplete may become increasingly important experience. However, these abilities also need to be trained in actual work. If young people's first positions are cut, companies will need new apprenticeship, rotation, internship, and training arrangements, rather than simply hoping job seekers will "learn AI on their own."

A 17.1% GDP growth rate is not 17.1% employment growth

The number in this study most likely to generate sensational headlines is the 17.1% GDP growth scenario mapped from Anthropic employees' expectations. It must be said again: it is the output of researchers putting survey answers into a specific economic model. It is not a direct prediction by employees, not government-published data, and certainly not proof by Anthropic that AI can make the economy reach this growth rate.

Even if AI does drive rapid productivity gains in the future, it does not follow that everyone's income will grow in tandem. Companies may increase profits, consumers may benefit from lower prices, and employees with certain scarce skills may receive higher pay; at the same time, other people's positions are cut, and switching jobs takes longer. Macro-level aggregate growth and individual income distribution are not the same report card.

For workers, the impact is not just "whether I can keep earning a salary," but also the cost of transition. Anthropic employees estimate that for people who lose their jobs because of AI and need to change occupations in 2030, the median time to find a new job is 12 months. This is a subjective expectation, but it points to a real problem: if switching jobs takes six months to a year, how are rent, family expenses, and career confidence to be sustained? And will companies and public policy bear part of the cost?

Perhaps AI will not take away everyone's job first. It may first change how people enter occupations: companies continue to grow, senior employees remain busy, and young people find their first door getting narrower and narrower. At the same time, for other industries and positions, AI may help people who originally lacked technical resources enter the field faster, start businesses independently, or gain new customers. Both directions are possible, and the future cannot be concluded on the basis of a single survey.

In the end, for an ordinary person who has just graduated, the most important question may not be whether AI in 2030 can do 75% or 28% of the work, but whether next summer they can still find a position that allows them to go from not knowing to knowing, from being inexperienced to becoming mature. The speed of technological progress is worthy of amazement, but whether a society can preserve room for newcomers to learn and grow is equally worthy of careful calculation.

Sources:

1. Core translation source document: Szymon Sacher, Tess Cotter, Charles I. Jones, Anton Korinek, Peter McCrory, The AI Expectations Gap: Public and Frontier Lab Views on AI's Impact on Work, Anthropic Institute, October 16, 2026. https://www.anthropic.com/research/ai-expectations-gap

2. Anthropic Institute, How Employment Is Changing at Anthropic, October 16, 2026. Used for the extended reporting at the end of the article on the sections about hiring and internal work changes. https://www.anthropic.com/research/employment-at-anthropic

3. Erik Brynjolfsson, Bharat Chandar, Ruyu Chen, No Widespread Displacement, but the AI Employment Gap for Young Workers Has Widened to 19%, Stanford Digital Economy Lab, August 12, 2026. https://digitaleconomy.stanford.edu/news/canariesaug26/AI Industry Signal Channel: Frontier Technology

30 articles published on Huxiu. This content may not be reproduced without permission. For licensing matters, please contact hezuo@huxiu.com. Everyone who is changing and wants to change the world is on the Huxiu app[ ]