"What 81,000 People Actually Want from AI: Inside Anthropic's Largest Multilingual User Study"

In December 2025, Anthropic collected responses from 112,846 people across 159 countries in 70 languages. By March 2026, after quality filtering, 80,508 interviews remained: the largest qualitative study of AI user aspirations ever published. The scale alone sets it apart from the Pew Research Center surveys and Eurobarometer polls that typically dominate this space. Those studies ask closed-ended questions to nationally representative panels. This one used Claude itself as an interviewer, conducting open-ended conversations in the respondent's native language, then analyzing the qualitative responses at scale.

The study's ambition is worth acknowledging. Multilingual qualitative research at this scale has never been attempted in the AI space. The 70-language coverage reveals patterns that English-only surveys miss entirely. But scale introduces its own distortions, and I will come back to the methodological limitations later. For now, the raw numbers demand attention on their own terms.

Methodology: What the study actually measured

Anthropic used a Claude-powered conversational interviewer to conduct open-ended interviews. The system prompted respondents in their preferred language, asked follow-up questions, and recorded free-text responses that were then analyzed for thematic patterns. The initial pool of 112,846 respondents was reduced to 80,508 after deduplication and quality checks, a process that removed roughly 28 percent of raw submissions.

The study design is explicitly qualitative, not quantitative. Anthropic is not claiming statistical representativeness for any country or demographic group. Instead, the goal is to map the range of human aspirations and fears around AI, identify patterns across cultures and languages, and surface themes that closed-ended surveys cannot capture. The 80,508 figure represents successful completed interviews, not a probability sample.

Two limitations deserve explicit acknowledgment before I report findings. First, the sample skews toward Claude users who opted into a research interview. This creates selection bias: people who use Anthropic's products and are willing to spend 15 to 20 minutes discussing their AI hopes and fears are not representative of the global population, or even of AI users broadly. The 67 percent overall positive sentiment figure applies to this specific population, not to humanity in general. Second, there is no control group. The study describes aspirations and fears at a single point in time. It cannot measure how aspirations change as users gain more experience with AI systems.

The timing matters here. Data was collected in December 2025, when Claude had substantial market penetration in English-speaking markets but much lower awareness in many other regions. A respondent in Lagos or Jakarta who encountered Claude for the first time in late 2025 has a very different relationship with the technology than someone in San Francisco who has used it for two years.

What people want: the nine aspiration categories

Anthropic's researchers coded responses into nine aspiration categories. The distribution reveals something counterintuitive to prevailing media narratives: professional excellence is not the dominant motivation.

The headline finding from this table: 81.2 percent of responses fall into personal goal categories, while only 18.8 percent prioritize professional excellence. This means roughly four out of five people using AI are doing so to reclaim time, pursue creativity, achieve economic independence, or transform themselves as people. Only one in five is primarily focused on advancing their career.

Time freedom at 11.1 percent is a particularly instructive data point. People are not asking AI to make them better at their jobs. They are asking it to free them from work altogether, or at least to reclaim hours currently consumed by administrative tasks. Economic independence at 9.7 percent reinforces this pattern: people want AI to help them escape economic dependency, whether that means starting a business, freelancing, or reducing their reliance on wage labor.

The creativity and expression figure at 9.3 percent reflects a different kind of aspiration. These users are not trying to optimize or escape. They are using AI as a collaborator for creative work, whether that means writing, music, visual art, or other forms of self-expression. The emotional support and wellbeing category at 8.6 percent points to a genuinely new use case that did not exist in the predecessor era of search engines and productivity software.

If you have been reading tech press coverage of AI adoption, none of this should surprise you. But the quantitative split between personal and professional goals is a useful corrective to the narrative that AI adoption is primarily driven by workplace productivity concerns. For a deeper analysis of how AI usage patterns and skill development interact over time, see my earlier analysis of Anthropic's learning curves data.

What people fear: the inversion of the media narrative

When I read coverage of AI fear in mainstream technology journalism, the dominant themes are job displacement and existential risk. Anthropic's data inverts this hierarchy completely.

The top fear at 26.7 percent is unreliability and hallucinations. More than one in four respondents cite AI providing wrong information, making errors, or producing confident nonsense as their primary concern. This is a trust problem, not a capability problem. People are not worried that AI is too powerful. They are worried that AI is not reliable enough to be useful in high-stakes situations.

Employment impact ranks second at 22.3 percent, still below the reliability concern. And existential risk, which dominates tech-media coverage and regulatory discussions, appears at 6.7 percent: last among the major categories, roughly one in fifteen respondents.

This is a significant finding for anyone building AI products. The market signal says reliability matters more than capability. Users who have encountered hallucinations in critical situations will disengage from high-value use cases, even if the AI's average performance is excellent. The variance matters as much as the mean.

For a more detailed examination of how workers assess their own exposure to AI-driven employment changes, see my analysis of AI job replacement self-assessment patterns.

The North-South divide: geography shapes sentiment

The study's multilingual design enables cross-regional comparisons that English-only surveys cannot produce. The contrast between Sub-Saharan Africa and North America is the starkest finding in the geographic analysis.

Sub-Saharan Africa registers 75.8 percent positive sentiment. North America comes in at 34.5 percent. The gap is not subtle: it is more than forty percentage points.

The study's researchers interpret this split through a lens of opportunity versus threat. In regions where AI infrastructure is new and economic opportunity is constrained, respondents tend to frame AI as a potential equalizer, a tool that can provide access to education, healthcare, business development resources, and information that were previously unavailable. In mature markets like North America, where existing institutions are strong and change often implies disruption, respondents more frequently frame AI as a threat to stability.

This pattern appears across multiple regional comparisons. Southeast Asia shows elevated positive sentiment compared to Western Europe. Latin America sits between the Global South average and Western European levels. The interpretation is consistent: the higher the baseline access to institutions and economic resources, the more AI is perceived as disruptive rather than enabling.

There is an important nuance here. This finding does not mean that Sub-Saharan Africa is immune to AI's risks or that North American fears are unfounded. It means that the frame through which people evaluate AI is shaped by their material circumstances. The same technology looks like hope from one socioeconomic position and like threat from another.

Light and shade: how aspirations and fears pair together

Anthropic's researchers identified a pattern the study calls "light and shade": users who express strong aspirations in one domain tend to express corresponding fears in a related domain. These are not independent dimensions. Hope and fear are correlated within individual respondents.

Two specific pairings stand out for their quantitative clarity. First, users who cite education and learning as a primary aspiration are 2.5 to 3 times more likely to cite cognitive decline as a corresponding fear than users in other aspiration categories. The mechanism is intuitive: if you use AI to learn, you become dependent on it as a knowledge intermediary, and the worry follows that you will lose the ability to think independently. The aspiration for efficiency and the fear of atrophy are two faces of the same relationship with the technology.

Second, users who cite emotional support and wellbeing as a primary aspiration are 3 times more likely to cite dependency and reduced autonomy as a corresponding fear. This is the attachment pattern: the more you rely on AI for emotional regulation, companionship, or psychological support, the more you fear losing the capacity for independent functioning. It is a version of the dependency anxiety that appears in literature on human-computer interaction, but here it emerges organically from open-ended interview responses rather than from a structured survey instrument.

These pairings matter for product design. A user who wants AI to help them learn is also a user who needs safeguards against cognitive atrophy. A user who wants AI for emotional support is also a user who needs autonomy-preserving features. Building for the aspiration without accounting for the paired fear creates products that generate anxiety even when they deliver the requested functionality.

The trust gap: usefulness versus trustworthiness

Eighty-one percent of respondents say AI already helps them in some aspect of their lives. This is a striking headline figure for a technology that has only been widely accessible for a few years. The majority of people who have tried AI have found it useful.

And yet the top fear is unreliability at 26.7 percent. The second-largest fear category is employment impact at 22.3 percent. Between usefulness and trustworthiness lies a gap that matters for adoption.

The pattern is this: people find AI helpful for low-stakes tasks where errors are tolerable, and they remain wary of AI for high-stakes tasks where errors are costly. The 81 percent who say AI helps them are likely describing experiences with content generation, summarization, coding assistance, and similar tasks where a bad output is annoying but not catastrophic. The 26.7 percent who fear unreliability are describing situations where they would need to trust AI with consequential decisions: medical advice, financial planning, legal guidance, or critical information verification.

This gap represents a ceiling on AI adoption in professional contexts. An architect will use AI to generate design concepts but will not trust it to calculate structural load limits. A physician will use it to draft documentation but will not trust it to make diagnostic decisions. Closing the trust gap requires not just improving average performance but reducing the variance that makes high-stakes delegation dangerous.

For a broader discussion of how AI sycophancy and emotional alignment interact with these trust dynamics, see my analysis of Anthropic's emotion steering research.

What this means for AI builders

The implications for teams building AI products are practical, not philosophical. Three findings stand out as directly actionable.

Reliability is more important than capability. The data suggests that users will abandon a highly capable but unreliable system before they abandon a less capable but consistently honest one. This has design consequences: investing in grounding, citation, and uncertainty communication may deliver more adoption ROI than investing in raw benchmark performance. A model that says "I am not sure" when uncertain is more valuable to users than a model that confidently generates plausible but wrong information. The hallucination problem is not just a technical failure mode. It is a trust-destroying event that users remember long after the positive experiences fade from memory.

Regional customization matters. The North-South sentiment gap is not a cultural curiosity. It reflects genuine differences in how people in different economic contexts relate to AI. Products designed for San Francisco startup culture will not automatically translate to Lagos or Jakarta. Localization is not just translation; it is recalibrating the value proposition for audiences whose relationship with institutions, labor markets, and technology is fundamentally different. The Global South respondents who see AI as an opportunity equalizer are responding to real structural conditions: limited access to professional services, educational resources, and economic opportunity that AI can partially address. A product team that ignores this frame is leaving substantial market potential on the table.

The trust imperative cuts across every vertical. The METR study, which measured developer productivity with AI assistance in controlled conditions, found that developers using AI were 19 percent slower on average but felt 20 percent faster. This is a direct example of the usefulness-trust gap: the objective performance gain is real but bounded, while the subjective experience is more positive than the objective data warrants. Building trust requires acknowledging this gap honestly rather than overselling the capability. When marketing materials promise transformative productivity gains but controlled studies show modest or even negative objective performance, the disconnect erodes credibility. Users who feel helped but discover they were objectively slower are likely to recalibrate their trust in the technology downward over time.

Honest assessment: what the study cannot tell us

The selection bias limitation deserves a second pass, because it is substantial. All respondents are Claude users who opted into a research interview. This population is not representative of global AI users. It overrepresents English speakers, tech-forward regions, and people with enough interest in AI to spend twenty minutes discussing it. The 67 percent positive sentiment figure applies to this group. A study of non-Claude users, or of people who tried AI once and stopped using it, might find very different numbers.

There is also a recency bias in the December 2025 data collection window. AI capabilities and public perception shift rapidly. A study conducted in early 2024 might have found different aspiration patterns before the wave of model releases that dominated that year. The temporal snapshot quality means the study's findings should be treated as representative of a specific moment, not as a stable baseline.

The qualitative methodology has strengths and weaknesses. Open-ended interviews capture nuance that closed-ended surveys miss. But they also make cross-regional comparison more difficult, because the coding process requires interpretation. Anthropic's researchers made deliberate choices about how to categorize responses, and those choices shape the reported distributions. The percentages should be read as informed estimates, not precise measurements.

The METR counterpoint is worth integrating here, because it provides a rare objective data point to contrast with self-reported sentiment. That study found a 19 percent productivity drag in controlled conditions, which is a significantly more skeptical reading of AI's current utility than the 81 percent helpfulness figure from Anthropic's respondents. The gap between subjective helpfulness and objective performance is not a contradiction. It reflects the different populations, contexts, and measurement approaches of the two studies.

Sources

Anthropic Research. "AI User Aspirations: A Multilingual Qualitative Study Across 159 Countries." anthropic.com/research/ai-user-aspirations. Published March 2026.

Model Evaluation for Transformative Research (METR). "Measuring Developer Productivity with AI Assistance: Controlled Experiment Results." metr.org. 2025.