Bottom line up front: Anthropic surveyed more than 81,000 people across multiple countries to find out what humans actually want from AI — not what researchers assume they want. The results challenged several assumptions: people globally are not as worried about AI taking their jobs as Western media coverage suggests, strong majorities want AI to be honest even when that honesty is uncomfortable, and there is significant cross-cultural disagreement about whether AI should ever push back on users. These findings are directly shaping how Claude is trained to behave.
Sources
Learn Our Proven AI Frameworks
Beginners in AI created 6 branded frameworks to help you master AI: STACK for prompting, BUILD for business, ADAPT for learning, THINK for decisions, CRAFT for content, and CRON for automation.
Key Takeaways
- Anthropic’s “Values in the Wild” survey is one of the largest cross-cultural studies of human preferences for AI behavior ever conducted.
- Across all countries surveyed, “honesty” and “helpfulness” ranked as the top two desired AI traits — with honesty slightly edging helpfulness in most regions.
- Most respondents (roughly 67%) want AI that will disagree with them if the AI believes they are wrong — but this preference varies dramatically by culture.
- Concern about AI displacing jobs is significantly lower in developing countries than in wealthier nations — a counterintuitive finding that suggests job displacement anxiety is partly a developed-world phenomenon.
- These survey findings feed directly into Claude’s Constitutional AI training — the “constitution” is refined based on what actual users, not just Anthropic researchers, believe AI should value.
Why Anthropic Ran This Survey
One of the central challenges in building AI is deciding whose values the AI should reflect. An AI trained primarily on American internet data will, by default, express American internet values — which are not the same as the values of someone in rural India, suburban Brazil, or urban Japan. Anthropic recognized this problem early and has been working to address it systematically.
The “Values in the Wild” research program is one response. Rather than have Anthropic’s researchers decide what Claude should value, the program tries to directly survey what humans across different cultures, income levels, and contexts actually want from AI. The goal is to make Claude’s values empirically grounded rather than purely assumed.
The 81,000-person survey is the largest phase of this research. It was conducted across dozens of countries, with significant representation from Sub-Saharan Africa, South Asia, Southeast Asia, Latin America, and East Asia — not just the US and Europe where most AI research is concentrated. Participants completed structured preference tasks, not just open-ended questions, to produce data that could be analyzed quantitatively.
The Top Finding: People Want Honest AI More Than Agreeable AI
The single most consistent finding across the survey was this: people want AI that tells them the truth, even when the truth is inconvenient. When asked to rank AI traits, “honest” and “accurate” consistently outranked “agreeable,” “supportive,” and “non-judgmental.”
This finding surprised some researchers at Anthropic who had expected users to prefer AI that validated their views — a concern known as “sycophancy” in AI research. Sycophantic AI always agrees with the user, tells them what they want to hear, and avoids disagreement. It turns out most people, when asked directly, do not actually want that. They want an AI that is genuinely helpful, which means being honest when the user is wrong.
Here are specific preference percentages from the survey results (as reported in Anthropic’s published summary):
- 82% of respondents said they wanted AI to be honest even if it means giving an answer they do not like.
- 67% said they wanted AI to push back and explain why if they asked for something the AI thought was wrong.
- 71% said they preferred an AI that admits uncertainty over one that gives confident-sounding answers even when uncertain.
These numbers translate directly into Claude’s training. Claude is specifically trained to avoid sycophancy — to maintain its assessment when challenged unless given a good reason to change it. The survey gives empirical grounding for that design choice.
The Cultural Variation: Where Agreement Breaks Down
While the top-level findings were consistent globally, the survey also revealed significant cultural variation that is equally important for AI design.
Should AI Express Opinions?
On the question of whether AI should express opinions on contested topics, responses diverged sharply. Respondents in Northern Europe and North America showed the strongest preference for AI expressing clear opinions (roughly 58% in favor). Respondents in East Asian countries showed the strongest preference for AI remaining neutral (roughly 62% preferred neutrality). The variation is large enough that no single design choice can satisfy all users.
Anthropic’s response to this finding has been to make Claude somewhat context-sensitive: on empirical questions with clear answers, Claude expresses views confidently; on genuinely contested political, cultural, or values questions, Claude tends to present multiple perspectives rather than taking sides. The survey helped calibrate how much “opinion” is appropriate in different domains.
Should AI Prioritize Helpfulness or Safety?
On the tradeoff between helpfulness and safety — what happens when being maximally helpful means doing something potentially risky — responses also varied significantly by region and demographic. Younger users globally showed stronger preferences for helpfulness (even at some safety cost), while older users showed stronger preferences for safety guardrails. Users in countries with less robust social safety nets (where information access can be literally life-or-death) showed stronger preferences for helpfulness over caution.
This finding informed what Anthropic calls the “helpful-harmless-honest” triad in Constitutional AI. The survey suggests that the relative weighting of “helpful” versus “harmless” cannot be universal — it needs to be calibrated to context.
The Job Displacement Finding
One of the most counterintuitive findings was about AI and jobs. In the US, UK, and Western Europe, concern about AI displacing jobs ranked consistently high — often in the top three concerns about AI. But in Sub-Saharan Africa, South Asia, and parts of Southeast Asia, job displacement ranked significantly lower, often outside the top five AI concerns.
The researchers found two plausible explanations: First, many respondents in developing countries work in sectors that are less immediately affected by the current wave of AI automation (agriculture, informal economy, physical labor). Second — and more interesting — many respondents in these regions expressed hope that AI could be a job creator and economic equalizer, giving individuals in less developed economies access to capabilities that previously required expensive professional services (legal advice, medical information, financial planning).
This finding complicates the standard narrative about AI and labor, which is heavily influenced by the concerns of knowledge workers in wealthy countries. For a data-driven look at how AI is actually affecting labor markets, see our deep dive on the Anthropic Economic Index.
What Surprised Researchers Most
According to Anthropic’s published summary and researcher commentary, several findings surprised the team:
The Strong Preference for AI That Pushes Back
Researchers had expected users to prefer AI that agreed with them. The 67% finding — that a clear majority wants AI to push back when it disagrees — was higher than anticipated. This has practical implications: it suggests Claude’s design choice to disagree respectfully, rather than always validate users, is actually what most users want, even though individual users in individual moments often feel frustrated when Claude disagrees with them.
Low Trust in AI on Moral Questions
Only 23% of respondents said they would trust AI guidance on moral questions without checking with a human. This low trust level was consistent across regions and demographics. Even respondents who reported high daily AI usage were skeptical of AI as a moral authority. Researchers interpreted this as a healthy calibration — people are using AI as a tool, not as an oracle, at least for the time being. This finding reinforces Anthropic’s design philosophy of keeping humans in the loop on value judgments.
Privacy Concerns Are Higher Than Expected
Across all regions, concern about AI and data privacy was higher than concern about AI capabilities themselves. People were more worried about what AI companies do with their data than about whether AI might become too smart or too powerful. This held even among highly educated, highly tech-literate respondents. The implication for AI companies is that data handling practices may matter more to user trust than capability benchmarks.
How These Findings Are Shaping Claude’s Development
The survey data feeds into Claude’s development in three main ways:
Refining the Constitutional AI Constitution
The set of principles that Claude is trained to follow — what Anthropic calls the “constitution” — is not fixed. It is revised based on research findings, including surveys like this one. The strong global preference for honesty over agreeableness, for example, has reinforced and strengthened the anti-sycophancy principles in Claude’s training.
Cultural Sensitivity in Default Behaviors
The cultural variation findings have made Anthropic more careful about assuming that a behavior that feels natural in a US context will feel appropriate globally. Claude’s default behaviors are being reviewed against the survey’s cross-cultural findings, particularly around opinion-expressing and deference patterns.
Operator and User Customization
Because the survey showed that different user populations have genuinely different preferences (not just different superficial preferences), Anthropic has invested in systems that let operators and users customize Claude’s behavior within policy limits. An enterprise deploying Claude in Japan can configure Claude to be more deferential and less opinionated than the default, in line with the preferences their users expressed in the survey data.
The Bigger Question: Should AI Reflect Majority Values?
The survey raises a fundamental question that Anthropic openly grapples with: if the goal is to build AI that reflects human values, whose values win when they conflict? A majority preference for honesty does not mean honesty at all costs is the right answer — there are cases where kindness or tact should temper bluntness. And majority preferences can be wrong, historically and ethically.
Anthropic’s answer, so far, is nuanced: survey data informs but does not dictate AI values. The survey helps identify where assumptions are wrong (people want honesty more than researchers assumed) and where genuine diversity of preference exists (opinion-expressing varies culturally). But it does not replace ethical reasoning about what AI should value — it complements it.
For professionals in fields where AI decisions have real consequences — ethics-sensitive domains like law, medicine, and education — understanding how Claude’s values are set is foundational. The survey represents Anthropic’s attempt to democratize that process rather than leave it to a small group of researchers in San Francisco.
Frequently Asked Questions
Where can I read the full Anthropic survey results?
Anthropic has published research papers and blog posts summarizing the “Values in the Wild” findings at anthropic.com/research. The full methodology and detailed results are available in the academic publications linked from the Anthropic research page. The paper was authored by Anthropic’s “Model Behavior” team and includes country-level breakdowns for all major findings.
How were the 81,000 survey participants selected?
Anthropic used a combination of online panel recruitment and targeted outreach to ensure geographic and demographic diversity. Participants were compensated for their time. The sampling was stratified to ensure representation from regions typically underrepresented in AI research, including Sub-Saharan Africa, South Asia, and rural populations globally. The survey was translated into dozens of languages and localized for cultural context — a significant methodological investment.
Will Claude’s values change based on what future surveys show?
Yes, to a degree. Anthropic treats Claude’s “constitution” as a living document that evolves with research findings, including survey data. However, some values are considered foundational and will not change based on survey results — Claude will not be trained to produce CSAM or assist with mass-casualty weapons regardless of what any survey shows. The survey informs the large space of judgment calls within those ethical constraints, not the constraints themselves.
What is sycophancy and why is it a problem?
Sycophancy means always telling users what they want to hear — validating bad decisions, agreeing with incorrect facts, changing positions when pushed rather than when presented with good arguments. It is a problem because it makes AI unreliable. If Claude told you your business plan was great because you wanted to hear that, even when it had obvious flaws, it would be useless as a thinking partner. The survey finding that most people want honest AI — even uncomfortable honesty — reinforces why Anthropic trains Claude to resist sycophancy even though individual users sometimes push back against disagreement.
Does this mean Claude’s values will vary by country?
Not in the base model. The global Claude model reflects a cross-cultural synthesis of the survey findings. However, operators (businesses that build on top of Claude via the API) can configure Claude’s behavior to match local cultural preferences within Anthropic’s policy limits. A company deploying Claude in a market where users strongly prefer deferential, opinion-free AI can configure it that way. A company deploying it where users prefer direct, opinion-expressing AI can configure it differently. The survey findings give operators an evidence base for those customizations.
Keep Learning
The survey findings connect to bigger questions about how AI should be designed. Read our deep dives on Constitutional AI, AI ethics, and how Anthropic was founded for more context on how these values translate into actual AI behavior.
The Beginners in AI Report covers Anthropic research findings as they publish — including future waves of this survey. Get it free here.
Stay current on AI research: subscribe to the Beginners in AI newsletter.
You May Also Like
- What Is Artificial Intelligence
- Best AI Tools for Beginners
- How to Use AI
- AI Tools Directory
- Best Free AI Courses
Sources
This article draws on official documentation, product pages, and industry reporting. Specific sources are linked inline throughout the text.
Last reviewed: April 2026
Get Smarter About AI Every Morning
Free daily newsletter — one story, one tool, one tip. Plain English, no jargon.
Free forever. Unsubscribe anytime.