The Geography of AI Optimism, and What the Development Agenda Misses

The standard framing of public AI opinion splits people into two camps: optimists who think this is the most transformative technology since the internet, and pessimists who think we are building something we will regret. This framing gets reproduced constantly because it is easy to argue about, easy to assign to people you disagree with, and easy to fit into a headline.

Martin Fowler’s March 26 fragment points to an Anthropic study that cuts through it. Anthropic had Claude interview roughly 80,000 users about their hopes and fears around AI, and the picture that emerged is more complicated than the camps model.

Before getting to the findings: the methodology is itself interesting. Qualitative research at scale typically means Likert scales and fixed-response surveys, which trade depth for volume. Using the model as interviewer allows for conversational follow-up, for probing what someone means when they say they are “worried” or “hopeful,” in a way that structured instruments cannot. The tradeoffs are different ones: consistency across conversations, interviewer effect, the question of whether people respond differently to a model than to a human. But the output is a richer characterization than a five-point scale.

What the study found: people group their AI attitudes around what they value. Financial security, access to learning, human connection. Within each cluster, the same person holds both optimism and concern at the same time. Fowler describes his own position as “yes to both” on the optimist/doomer question, and the data suggests this is the norm rather than the exception.

This maps to my own experience discussing AI with people outside the tech industry. The question “are you optimistic about AI?” produces wildly different answers depending on which AI application you actually mean. A system that helps someone’s kid get personalized reading instruction: yes. A system that makes synthetic media indistinguishable from recorded reality: no. AI-assisted triage in places where trained physicians are scarce: genuinely yes. The downstream effects on medical residency pipelines: much less so. These are not contradictory positions held by confused people. They are coherent responses to different applications of the same general technology.

The Geographic Variance

The more striking finding in the Anthropic data is the geographic pattern. The less economically developed a country, the more optimism people tend to express about AI.

Pew Research surveys on AI attitudes from 2023 found the same contour: respondents in Nigeria, India, Indonesia, and the Philippines viewed AI products and services more favorably than respondents in Germany, the Netherlands, the United States, and Canada. IPSOS’s annual global AI surveys have tracked a similar divide for several years. The Anthropic data, gathered conversationally, appears to confirm the same geographic shape.

The pattern is not unique to AI. When mobile phones expanded across Sub-Saharan Africa in the late 1990s and 2000s, populations that lacked existing landline infrastructure adopted the technology at faster rates and with more enthusiasm than populations that had to abandon established systems. M-Pesa, launched in Kenya in 2007, became the canonical case study: Kenya went from low formal banking penetration to majority digital payment usage in under a decade. There was no incumbent telco defending landline revenue, no bank lobby defending branch infrastructure. The people who needed the capability most, and had the fewest existing arrangements to disrupt, moved fastest.

AI presents the same structure. In a country where trained physicians are unevenly distributed and often concentrated in urban centers, a diagnostic support tool that can extend clinical reach into remote areas is transformative in a direct, concrete sense. In the United States, that same tool enters a system organized around professional licensing boards, malpractice liability frameworks, insurance billing codes, and economic interests that have accumulated over decades. The capability is the same. The stakes of disruption are different.

The same comparison holds for education. An AI tutoring system that provides personalized pacing and feedback in a context where classroom ratios are forty to one addresses a real gap. In a system that already has functioning educational infrastructure, however imperfect, the value proposition is incremental rather than foundational.

What the Optimism Gap Implies for Development

The geographic variance has a practical implication that rarely surfaces in mainstream AI discourse. The features that get built, the safety guardrails that get prioritized, the use cases that receive research investment: all of this is shaped by where the engineering workforce is concentrated and what risks look salient from there.

If AI development is primarily driven by teams in wealthy countries, and those teams are asking “what might this break or threaten,” the resulting systems get optimized for harm mitigation within wealthy-country contexts. Content moderation for social media. Bias audits for hiring algorithms. Privacy protections against surveillance infrastructure that already exists. These are legitimate concerns. They also reflect a particular starting point.

A parallel conversation has been running in global health technology for decades. Medical device and pharmaceutical development gets funded around the disease burden of wealthy populations, not the global disease burden. The result is that conditions like malaria and tuberculosis receive less research investment than their actual global impact would justify. The AI equivalent might be heavy investment in detecting AI-generated content in news articles or improving recommendation system transparency, while the question of how AI can deliver basic educational or diagnostic services to people who currently lack them goes comparatively underexplored.

None of this means the wealthy-country concerns are misdirected. Concentration of economic and political power through AI, labor market disruption at scale, erosion of shared epistemic infrastructure, erosion of privacy: these are real harms worth sustained attention. The question is whether the development agenda, shaped by who is doing the developing and where they live, adequately reflects what would be genuinely transformative for the majority of the world.

Values as the Organizing Frame

The decision to organize the Anthropic findings around values rather than around AI attitudes is analytically useful beyond the immediate survey. Most people do not spend time forming opinions about AI as a general phenomenon. They think about what they need, and AI is becoming relevant to whether they can get it.

Financial security: tools that give small business owners access to financial planning that currently requires expensive advisors, fraud detection for thin-file credit markets, income smoothing for informal workers.

Learning: tutoring systems that provide individualized attention at scale. This is not speculative; working implementations exist now, and the question is whether they get deployed where the need is highest or where the paying customers are.

Human connection: translation tools that reduce the friction of communicating across languages, communication aids for people with disabilities, systems that extend the effective reach of healthcare workers across geographic distance.

The answer to “will AI benefit people like me” differs because starting points differ. A developer with stable employment at a well-resourced company, reliable healthcare, and functioning civic infrastructure is asking a genuinely different question than a smallholder farmer in Tanzania asking whether AI can help identify crop disease from a phone photo, or a teacher in a rural school asking whether AI can give her students the feedback she does not have time to provide herself.

The Spec Problem in Context

Fowler’s fragment also points to something Julia Shaw raised about specification-driven development with LLMs. The excerpt cuts off before the full argument, but the setup is familiar: the advice to write a spec before you prompt has spread widely, and something in the follow-through is missing. The spec gets written, the conversation begins, decisions accumulate through the session, and the spec becomes historical context rather than a live guide. The reasoning behind why the code looks the way it does gets buried in chat history that expires.

This is a real problem in AI-assisted development workflows, and worth solving. It is also the kind of problem that occupies developer attention primarily in contexts where the base infrastructure already works. The bigger population of people interacting with Anthropic’s survey, organizing their AI attitudes around financial security and learning and connection rather than around workflow optimization, is asking whether the capability reaches them at all. Both questions matter. The data suggests which one has more people waiting on the answer.