Anthropic 首次公众记录调查结果公布

我们正在开展一项新的调查系列——Anthropic 公众记录，旨在了解公众对 AI 的看法和感受，并展示第一波调查结果的快照，该调查于 2025 年 11 月至 12 月进行，近 5.2 万名美国人参与。

我们发现：

• 近半数（48%）的美国人将治愈癌症或阿尔茨海默症等疾病列为对 AI 的三大期望之一，其次是帮助残障人士（36%），然后是推动技术进步和让生活更轻松（并列 23%）。
• AI 导致的岗位流失在每一个州都是最常见的恐惧，64% 的美国人对此感到担忧。第二突出的恐惧是认知依赖（56%），其次是虚假信息（52%）。
• 支持政府干预 AI 的比例很高：超过 70% 的受访者认为政府应在监管 AI 中发挥作用，且这一支持跨越党派。人们最希望政府在隐私（56%）、儿童安全（52%）和损害责任（49%）等领域对 AI 采取行动。
• 当被问及什么最能确保 AI 有益于人类时，美国人将追究 AI 公司的法律责任（47%）和将安全置于增长之上（44%）列为杠杆作用最大的行动。
• 仅 15% 的美国人表示信任 AI 公司对 AI 开发和使用方式做出决策。

引人注目的是，在大多数问题上，AI 并没有让美国人在典型的党派、地域或教育界限上出现严重分歧。总的来说，各个议题都存在广泛共识：美国人渴望实现 AI 所承诺的益处，但担心其可能带来的颠覆，并且希望构建 AI 的公司承担起责任。即便存在分歧，也主要是在观点强度上的差异。

这项研究基于 Anthropic 正在开展的其他工作，旨在理解人们如何使用 Claude 以及如何看待 AI 的发展。我们最近通过 Anthropic Interviewer 这一大规模深度访谈工具，完成了一项针对 8.1 万名 Claude 用户的全球定性研究。我们还定期发布来自 Anthropic 经济指数 的数据，该指数利用匿名化的 Claude 使用数据，展现世界各地的人们如何使用 AI。Anthropic 公众记录调查是我们首次直接对话普通公众，这使我们能够触达非 AI 用户，并更好地理解态度如何因人口特征而异。

Anthropic 公众记录将定期更新，随着新议题变得更加突出而扩展范围，使我们能够追踪公众对 AI 的态度如何随着模型能力进步和采用深化而改变。未来，我们计划将调查扩展到美国之外。

方法简介

我们于 2025 年 11 月和 12 月进行了一项具有全国代表性的在线调查，调查了 51,993 名美国人，样本来源于 YouGov，并根据美国人口普查基准进行了加权。各州样本量从 232（阿拉斯加州）到 1,902（纽约州）不等，州级误差幅度在 ±2.6 到 ±9.1 个百分点之间。关于方法的更多细节见附录。

美国人对 AI 的期望

我们请美国人从 17 项清单中选择他们对 AI 的三大期望。治愈疾病位居榜首，48% 的受访者将其列入前三，领先第二名帮助残障人士（36%）12 个百分点。像治疗和减少孤独感——或者说期望 AI 替代人际接触——这类选项在所列选项中排名最低。

美国人的担忧

我们向受访者提供了一份包含 20 种 AI 可能造成的危害的清单，请他们标记每一种个人感到关切的危害，然后用五级量表对其担忧程度评分。我们将任何 2 分（稍微担忧）及以上的回答视为担忧。（该方法与上面的问题不同，在上面的问题中我们只让参与者排出了他们的三大期望；这两个数字不可直接比较。）

岗位流失是迄今为止最常见的担忧，近三分之二（64%）的美国人对此感到担忧。紧随其后的是认知依赖——即 AI 融入导致人们无法独立思考——达到 56%，以及虚假信息 52%。岗位流失和认知依赖也是我们 对 8.1 万名 Claude 用户进行的定性研究 中的主要恐惧。

最常见的危害往往具有近期性和具体性：岗位流失、认知依赖、虚假信息、犯罪滥用和监控。每一种恐惧在 AI 之前就已存在，在更早的技术中就有先例——例如，自动化造成岗位流失、智能手机促进依赖、社交媒体传播虚假信息。总体而言，美国人更担心的是 AI 的滥用，而非 AI 对齐失败，提及犯罪滥用、监控和恐怖主义的频率，高于 AI “失控”等情形。

对于我们所提及的除三种危害外的所有危害，多数受访者描述自己“不担忧”，然而，没有任何一种潜在危害是有不到四分之一的美国人对其至少有些许关切的。

岗位流失的模式

64% 的美国人担心 AI 将取代工作岗位。这种担忧分布得异常均匀。它是民主党人（67%）和共和党人（62%）、有子女家庭（59%）和无子女家庭（66%）以及每个州从爱荷华州的高点（71%）到密西西比州的低点（57%）排名第一的恐惧。

学历越高的美国人越担忧岗位流失

对岗位替代的担忧随受访者教育水平上升。拥有研究生学位的美国人比高中及以下学历者担忧岗位流失的比例高出近 10 个百分点。换句话说，最担心岗位替代的劳动者，恰恰是其工作内容已经与 AI 所被要求的任务更紧密重叠的群体——这一发现 反映 在我们的经济研究团队对全球 Anthropic Interviewer 研究的分析中。

最少使用 AI 的人对岗位流失的恐惧更高

与此同时，每天在工作中使用 AI 的人对岗位流失的担忧 明显低于 根本不使用 AI 的人：54% 对比 70%。

对这一趋势有多种可能的解释。亲手使用 AI 的经验可能有助于人们培养技能和熟练度，使他们能够增强而非自动化的部分工作，让岗位流失看起来不那么像逼近的威胁。亲手经验还可能让人看到 AI 的局限性。很可能是这些因素以及其他因素的结合。

对 AI 能力的感知及工作场所接受度

我们给受访者提供了一份包含 14 项工作场所任务的清单，并对每项任务问了两个问题：“你认为今天 AI 工具或应用能够多好地执行这项任务？”以及“考虑到你自己的岗位，你对以下每项工作任务/目的的 AI 参与程度的偏好是多少？”

对 AI 能力的总体评估相当高。在高端，75% 的美国人认为 AI 在研究方面的能力与人类相当或更好。在低端，44% 的人认为 AI 在服务与支持方面与人类相当或更好。

在大多数任务上，大多数美国人不希望 AI 参与 他们的工作，而且即使在他们认为 AI 能力最强的任务——比如研究和数据分析上——也有近一半的受访者表示不希望 AI 参与自己的工作。然而，工作场所中对 AI 参与的接受度似乎与感知的能力同步变化：人们认为 AI 在某个领域越有能力，就越可能愿意使用它。

认知依赖是一种预期性恐惧

在我们的调查中第二常见的恐惧是对 AI 的认知依赖。为了更好地理解人们是否真的正在经历依赖，我们询问受访者，如果 AI 明天变得不可用，他们会感受到多大程度的干扰，然后比较了担忧依赖和不担忧依赖者的回答。

迄今为止，认知依赖似乎主要是一种预期性恐惧：在表达了一些对依赖担忧的 56% 的美国人中，只有约五分之一的人会在 AI 不可用时感受到显著的干扰。相反，在不担忧依赖的 44% 的人中，更高比例——约三分之一——会 感受到显著干扰。

我们对 8.1 万名 Claude 用户 的定性研究发现，教育工作者报告亲眼目睹认知能力衰退的可能性是平均水平的 2.5 到 3 倍，推测是在他们的学生身上。在 Anthropic 公众记录中，教育工作者同样属于最担忧依赖的职业群体之一，仅次于从事艺术和设计工作的人。

与岗位流失一样，随着使用频率增加，依赖恐惧稳步下降。每天在工作场所使用 AI 的美国人对依赖的担忧（46%）比从不使用的人（62%）低 16 个百分点。

重度 AI 用户的情况

截至 2025 年底，约 6% 的美国人在工作和个人生活中都 每天都在使用 AI。这些“融合用户”是 AI 更深度采纳情况的一个预览，也可能预示着随着采纳增长，主流意见的走向。

融合用户偏向年轻、男性、城市居民、在职且受过大学教育。近三分之二的人将自己描述为在大多数人之前尝试新技术，或一旦看到潜力就早期采用的人，而普通公众中这一比例为 30%。

融合用户在我们列出的每一项危害上的担忧程度都低于普通公众，尽管这可能反映了早期采纳者心态上的差异。

美国人希望 AI 治理如何

71% 的美国人表示政府应当参与到 AI 的发展和监管中。该数字在民主党人中为 79%，共和党人中为 68%，独立人士中为 69%——这是一个跨党派的压倒性多数。在我们调查的每个州和领土，多数人都支持政府参与 AI，从哥伦比亚特区的 81% 到夏威夷的 63% 不等。

美国人希望政府在哪些方面行动

接着，我们询问了八个特定领域，以及政府应在每个领域参与的程度。只有两个领域——隐私和儿童安全——获得了超过最低限度参与的绝对多数支持。与此同时，国家安全是所有领域中党派差距最小的，民主党人和共和党人之间只差三个百分点。

美国人对行业的期望

当我们问到应该采取什么措施来确保 AI 的开发符合人类利益时，美国人聚焦在两个答案上：追究 AI 公司的法律责任（47% 将其选入前三）和将安全置于增长之上（44%）。随后是拥有实权的独立监督机构（29%）和为安全放慢 AI 发展（27%）。

信任赤字

仅有 15% 的美国人表示信任 AI 公司对这项技术的开发和使用方式做出决策。这是在我们测试的所有机构中最低的数字，低于联邦政府（20%）、州和地方政府（19%）以及国际机构（20%），并远低于独立专家（43%）。

深度用户同样有监管和监督的需求

融合用户对我们询问的每一家机构都更为信任，包括 AI 公司——并且显著更不倾向于认为应放缓或停止 AI 发展。然而，他们也以与全国比例几乎相同的程度支持政府参与 AI（74% 对比 71%），而且在我们在八个特定治理领域进行的测试中，他们的偏好与公众几乎无法区分。

Anthropic 的行动与下一步计划

公众意见对确保强大的 AI 服务于人类利益至关重要。Anthropic 公众记录、Anthropic Interviewer、Anthropic 经济指数以及我们许多其他研究项目，都是在为更深入地理解并获取公众意见，以便让 AI 转型顺利进行。

我们近期公布了几项与这些发现相关的 政策框架。我们的前沿 AI 框架提出了对前沿模型进行强制性独立安全测试、透明度要求，以及政府阻止或召回危险 AI 部署的权限。我们的经济政策框架阐述了政府如何为 AI 的经济影响做好准备——尽可能减少岗位替代，并为劳动者提供支持，使 AI 的收益得到广泛共享。

AI 的发展方向不应仅由构建它的公司来决定。公众的希望与担忧被衡量得越清晰，我们——以及其他公司——就越能更好地回应它们。

附录：方法

人群与方式： Anthropic 公众记录是一项针对美国 16 岁及以上、居住在 50 个州、哥伦比亚特区或波多黎各的青少年及成年互联网人口的在线调查。田野调查由 YouGov 通过其在线小组于 2025 年 11 月 1 日至 12 月 11 日期间执行。

样本设计： 该研究设计为 52 个并行的州和领土样本，目标为每州约 1,000 份完成问卷（蒙大拿州、北达科他州、南达科他州、怀俄明州、佛蒙特州和波多黎各约 500 份）。每个州内在年龄、性别、教育和种族/族裔上设定了配额。最终的样本量为 51,993 份完成问卷，各州样本量范围从 232（阿拉斯加州）到 1,902（纽约州）。

加权： 数据根据州、年龄、性别、教育程度等变量进行了加权，以代表 16 岁及以上的美国常住人口。

We’re conducting a new survey series, Anthropic Public Record, to understand how the public thinks and feels about AI, and presenting a snapshot of the results from the first wave, fielded in November and December of 2025 with nearly 52,000 Americans.

We found:

• Nearly half (48%) of Americans ranked curing diseases like cancer or Alzheimer’s as one of their top three hopes for AI, followed by helping people with disabilities (36%), then making technological progress and making life easier in general (tied at 23%).
• AI-induced job loss was the most common fear in every state, held by 64% of Americans. The second most prominent fear was cognitive dependency (56%), followed by misinformation (52%).
• Support for government intervention in AI was high: over 70% of the Americans we surveyed believe the government should play a role in regulating AI, and this support was bipartisan. People were most eager to see the government take action on AI in the areas of privacy (56%), child safety (52%), and liability for harm (49%).
• When asked what would best ensure AI is of benefit to humanity, Americans ranked holding AI companies legally liable for harm (47%) and prioritizing safety over growth (44%) as the highest-leverage actions.
• Only 15% of Americans said they trust AI companies to make decisions about how AI is developed and used.

Strikingly, on most questions, AI did not heavily divide Americans along typical partisan, geographic, or educational lines. In general, there was broad consensus across topics: Americans are eager to realize AI’s promised benefits but fear the disruption it may bring, and they want accountability from the companies building it. To the extent we saw disagreement, it was largely only in the intensity of people's views.

This research builds on other work underway at Anthropic to understand how people use Claude and think about AI development. We recently conducted a global qualitative study of 81,000 Claude users through Anthropic Interviewer, our tool for conducting in-depth interviews at scale. We also regularly release data from the Anthropic Economic Index, which draws on anonymized Claude usage data to show how people around the world are employing AI. The Anthropic Public Record survey marks the first time we’ve spoken to the general public, allowing us to reach non-users of AI and better understand how attitudes differ across demographic lines.

The Anthropic Public Record will be repeated regularly, evolving in scope as new topics become more salient, and allowing us to track how the public’s attitudes towards AI change as model capabilities advance and adoption deepens. In the future, we plan to expand outside the US.

Method in brief

We conducted a nationally representative online survey in November and December of 2025 of 51,993 Americans, sourced from YouGov and weighted to US Census benchmarks. State samples range from n=232 (Alaska) to n=1,902 (New York), with state-level margins of error between ±2.6 and ±9.1 percentage points. More details on the methodology are available in the Appendix.

What Americans hope AI will deliver

We asked Americans to choose their top three hopes for AI from a list of 17. Curing disease topped the list, with 48% of respondents putting it in their top three, 12 percentage points ahead of the second most commonly selected item, helping people with disabilities, at 36%. Items like therapy and reducing loneliness—or hopes that AI might substitute for human contact—were the lowest ranked of the options presented.

What Americans fear

We gave respondents a list of 20 possible harms from AI, asked them to flag each one they felt personally concerned by and then to rate each on a five-point scale of how worried they were. We considered any response of 2 (somewhat worried) or higher as worried. (This methodology differs from the question above, in which we asked participants to rank only their three hopes; the numbers aren’t comparable.)

Job loss was by far the most common concern, held by nearly two-thirds (64%) of Americans. This was followed by cognitive dependency—in which AI integration leaves people unable to think for themselves—at 56%, and misinformation at 52%. Job loss and cognitive dependency were also among the top fears in our qualitative study of 81,000 people using Claude.

The most common harms tended to be near-term and concrete: job loss, cognitive dependency, misinformation, criminal use, and surveillance. Each of these fears also predates AI, having precedent in an earlier technology—for example, automation causing job loss, smartphones fostering dependency, and social media spreading misinformation. In general, Americans tended to be more concerned with the misuse of AI than AI misalignment, citing criminal use, surveillance, and terrorism more frequently than, for example, AI “going rogue”.

For all but three of the harms we mentioned, a majority of respondents described themselves as “not worried,” but there was no potential harm about which less than 1/4 of Americans had at least some concern.

Patterns with job loss

Sixty-four percent of Americans are worried that AI will displace jobs. The concern is remarkably evenly distributed. It is the top-ranked fear among Democrats (67%) and Republicans (62%), in households with children (59%) and without (66%), and in every state from Iowa at the high end (71%) to Mississippi at the low end (57%).

Job loss concerns are higher among Americans with more education

Concerns over job displacement rise with a respondent’s education level. Americans with postgraduate degrees are nearly 10 percentage points more worried about job loss than those with a high school education or less. The workers most worried about displacement, in other words, are the ones whose work already overlaps more closely with what AI is being asked to do—a finding reflected in our economic research team’s analysis of our global Anthropic Interviewer study.

Fear of job loss is heightened among those who use AI least

At the same time, people who use AI at work every day are notably less worried about job loss than people who don’t use AI at all: 54% versus 70%.

There are many possible explanations for this trend. Hands-on experience with AI may help people develop skills and fluency that allow them to augment rather than automate parts of their job, making job loss seem like less of a looming threat. Hands-on experience may also reveal AI's limitations. Likely it’s a combination of these and other factors.

Perceived capabilities and acceptance of AI at work

We gave respondents a list of 14 workplace tasks and for each asked two questions: “How well do you think an AI tool or application could perform this task today?” and “Thinking about your own job, how much AI involvement would you prefer for each of the following work tasks/purposes?”

Overall assessment of AI capabilities was fairly high. At the high end, 75% of Americans said AI was as good or better than humans on research. At the low end, 44% said AI was as good or better in service and support.

On most tasks, a majority of Americans did not want AI involved __ in their jobs, and even on the tasks they rated AI most capable—such as research and data analysis—nearly half of respondents said they want no AI involvement in their own work. However, acceptance of AI involvement within the workplace seems to move in lockstep with perceived capabilities: the more competent AI is perceived to be in a given domain, the more likely people are to be willing to use it.

Cognitive dependency is an anticipatory fear

The second most common fear in our survey was cognitive dependency on AI. To better understand whether people might actually be experiencing dependency, we asked respondents how much disruption they would feel if AI became unavailable tomorrow, then compared the answers of those worried about dependency with those unconcerned by it.

So far, cognitive dependency appears to be a mostly anticipatory fear: of the 56% of Americans who expressed some worry over dependence, only roughly 1/5 would feel significant disruption if AI became unavailable. Conversely, among the 44% who don’t worry about dependency, a higher percentage—roughly 1/3— would feel significant disruption.

Our qualitative research among 81,000 Claude users found that educators were 2.5 to 3 times more likely than average to report having witnessed cognitive atrophy firsthand, presumably in their students. In Anthropic Public Record, educators are likewise among the occupations most worried about dependency, second only to people working in arts and design.

As with job loss, the dependency fear falls steadily with usage. Americans who use AI daily at work are 16 points less worried about dependency (46%) than those who never do (62%).

What the heaviest AI users tell us

As of late 2025, about 6% of Americans used AI every day for both __ work and personal life. These integrated users are a preview of what more intensive adoption of AI looks like, and possibly of where mainstream opinion is headed as adoption grows.

Integrated users skew young, male, urban, employed, and college-educated. Nearly two-thirds describe themselves as people who experiment with new technology before most others or adopt it early once they see its potential, compared to 30% in the general public.

Integrated users are less worried than the general public across each of the harms we listed, though this probably reflects differences in the outlook of early adopters.

What Americans want from AI governance

71% of Americans say the government should be involved in the development and regulation of AI. The figure is 79% among Democrats, 68% among Republicans, and 69% among Independents—a bipartisan supermajority. A majority in every state and territory we surveyed supports government involvement in AI, from 81% in the District of Columbia to 63% in Hawaii.

Where Americans want government to act

We then asked about eight specific domains, and how heavily the government should be involved in each. Only two—privacy and child safety—draw outright majority support for more than a minimal role. National security, meanwhile, has the narrowest partisan gap of any domain, just three points between Democrats and Republicans.

What Americans want from the industry

When we asked what should happen to ensure AI is developed in humanity's interest, Americans converged on two answers: hold AI companies legally liable for harm (47% chose it among their top three) and prioritize safety over growth (44%). Independent watchdogs with real power (29%) and slowing AI development for safety (27%) followed.

The trust deficit

Only 15% of Americans said they trust AI companies to make decisions about how the technology is developed and used. That was the lowest figure for any institution we tested, below the federal government (20%), state and local government (19%), and international bodies (20%), and far below independent experts (43%).

Integrated users have the same appetite for regulation and oversight

Integrated users were more trusting of every institution we asked about, including AI companies—and were markedly less inclined to say AI development should be slowed or stopped. However, they also support government involvement on AI at essentially the national rate (74% versus 71%), and across the eight specific governance domains we tested, their preferences are nearly indistinguishable from the public's.

Anthropic initiatives and what comes next

Public input is critical to ensuring that powerful AI serves humanity’s interests. Anthropic Public Record, Anthropic Interviewer, the Anthropic Economic Index, and many of our other research projects are all efforts to gain greater understanding and input from the public about how to make the AI transition go well.

We recently announced several policy frameworks relevant to these findings. Our Advanced AI Framework proposes mandatory independent safety testing for frontier models, transparency requirements, and government authority to block or recall dangerous AI deployments. Our Economic Policy Framework lays out how governments can prepare for AI's economic impacts—minimizing job displacement where possible, and supporting workers so that AI's benefits are broadly shared.

The direction AI takes should not be set only by the companies building it. The more clearly the public’s hopes and concerns are measured, the better we—and other companies—can meet them.

Appendix: Methodology

Population and mode: Anthropic Public Record is an online survey of the US late-teen and adult internet population, age 16 and over, resident in the 50 states, the District of Columbia, or Puerto Rico. Fieldwork was conducted by YouGov from its online panel between November 1 and December 11, 2025.

Sample design: The study was designed as 52 parallel state and territory samples with a target of approximately 1,000 completes per state (approximately 500 in Montana, North Dakota, South Dakota, Wyoming, Vermont, and Puerto Rico). Quotas were set within each state on age, gender, education, and race/ethnicity. The achieved sample is 51,993 completes, with state sample sizes ranging from 232 (Alaska) to 1,902 (New York).

Weighting: Data are weighted to be representative of the US resident population age 16 and over on state, age, gender, education, and race/ethnicity. Weights are normalized to the achieved sample size (mean 1.00, sum 51,993).

Margin of error: The national margin of sampling error is ±0.6 percentage points at the 95% confidence level for a proportion of 50%. State-level margins of error range from ±2.6 points (California, New York, Texas) to ±9.1 points (Wyoming). Margins of error for subgroups are larger.

Reporting conventions: All percentages in this report are weighted. Unless noted otherwise, percentages use the full segment as the denominator (respondents who answered "don't know" or skipped remain in the base). "Worried" on the fears battery is the top four boxes of a five-point worry scale. "As good or better" on the capability battery is the top three boxes of a five-point performance scale. "Integrated users" are respondents who report using AI one or more times daily for work and one or more times daily for personal purposes (unweighted n=2,717). Party affiliation groups include leaners.

Limitations: Anthropic Public Record captures what Americans believed about AI in late 2025. We are treating it as a baseline.