What 81,000 People Want from AI: Anthropic's Global Interview Study

Executive Summary

Anthropic interviewed 80,508 Claude users across 159 countries and 70 languages in a single week of December 2025, published the results March 18, 2026. The headline: 81% said AI is advancing the life they want. The smaller, more useful number: people average 2.3 distinct worries each, and the top concern — unreliability — is named by 26.7%. Hope and alarm live inside the same person, not in separate camps.

For an executive planning a rollout, the study is a qualitative complement to BCG’s AI at Work 2025 (n=10,635) and OpenAI’s State of Enterprise AI 2025 (n=9,000). It gives language for what workers actually want and fear, segmented by region and role type. It does not replace controlled productivity evidence — vendor caveat applies in full.

Source credibility: MEDIUM. Anthropic’s customers skew toward technical early adopters and the self-selected users who agreed to a Claude-moderated interview. The sample is large and globally distributed, the methodology is transparent, and the categorization was done by Claude classifiers — which is both the method’s strength (scale) and its weakness (the interviewer and the coder are the same vendor’s model). Treat percentages as directional, not precise. Published March 18, 2026 — TIER 1.

Key Data Points

Regional net positive sentiment

• Sub-Saharan Africa: 75.8%
• Latin America & Caribbean: 73.7%
• South & Central Asia: 69.2–69.4%
• North America: 65.5%
• Oceania: 64.5%
• Western Europe: 64.4%

Five tensions where benefits and harms entangle

1. Learning (33% benefit / 17% atrophy) — high co-occurrence.
2. Decision-making (22% benefit / 37% unreliability) — harm is larger than benefit, and primarily experienced directly.
3. Emotional support (16% benefit / 12% dependence) — 3x higher co-occurrence than any other pair.
4. Time-saving (50% benefit / 18% illusory productivity) — self-employed report both most.
5. Economic empowerment (28% benefit / 18% displacement) — lowest co-occurrence (+0.16).

What This Means for Your Organization

The practical signal for a 200–5,000 person company is that concerns cluster by role type, not by whether someone is “pro-AI” or “anti-AI.” Educators see cognitive atrophy 2.5–3x more than average. Self-employed and tradespeople see the economic upside without the atrophy. Institutional employees — the workforce most companies are planning rollouts for — report economic benefit at 14% vs. 50%+ for self-employed. That gap is not about the tool. It is about who captures the value when the tool is deployed inside a structure someone else controls.

Three design implications:

1. Lead with reliability, not capability. Unreliability (26.7%) is the single largest concern and the one workers experience firsthand — 37% cite it in decision-making contexts. Any rollout that opens with model benchmarks instead of error-handling workflow will land wrong.
2. Acknowledge the tensions out loud. Emotional support use correlates 3x with dependency concern. Time-saving correlates with “illusory productivity” — the France quote (“you just run faster to stay in place”) is the one people recognize. Naming the tension in training is more credible than denying it.
3. Segment the message by role. Wealthier-region employees frame AI as life-management. Developing-region employees frame it as economic opportunity. Inside a US company, knowledge workers, frontline staff, and external contractors will each map onto a different vision profile.

Mid-market leaders planning a rollout who want to pressure-test where their workforce actually sits on these tensions — and how that should shape training, guardrails, and communications — can reach brandon@brandonsneider.com.

Sources

• Anthropic. “What 81,000 People Want from AI.” Published March 18, 2026. https://www.anthropic.com/81k-interviews
• Raw fetched content: sources/01-ai-native-landscape/anthropic-81k-interviews-raw.md
• Companion corpus sources for triangulation: BCG “AI at Work 2025” (n=10,635), OpenAI “State of Enterprise AI 2025” (n=9,000), Anthropic Economic Index (Nov 2025 / Mar 2026).

Vendor caveat: Anthropic customers skew technical and self-select into a Claude-moderated interview. Classifiers were Claude-powered — the interviewer and coder share a vendor. Percentages are directional. This is a qualitative worker-voice study, not a controlled productivity measurement. Cross-reference against METR RCT (experienced developers 19% slower), CMU study (40.7% code complexity increase), Atlan 200-deployment analysis (median +159.8% ROI requires workflow redesign first) before using any stat as an ROI claim.

Brandon Sneider | brandon@brandonsneider.com

April 2026