The strongest critique of “AI will fix it” is not anti-technology
Artificial intelligence is increasingly marketed as a remedy for problems in education, agriculture, employment, and public service delivery. That framing is seductive because it compresses messy social failures into a tractable engineering challenge. If institutions are slow, underfunded, or fragmented, the promise of a responsive model seems almost irresistible.
But an essay published by Rest of World argues that this framing misses the central reality of social systems: technical capability alone is not enough. Even sophisticated AI tools need human support, institutional capacity, and local accountability if they are going to do more than generate impressive demos.
The article, written by Cornell researchers Deepak Varuvel Dennison and Aditya Vashistha, does not deny AI’s genuine potential. It explicitly acknowledges growing evidence of productivity gains and the appeal of AI in both private and public sectors. Its argument is narrower and more important: deploying AI in underserved communities is not the same thing as solving their problems.
The contradiction at the center of AI-for-good
The essay highlights a structural tension. AI is often presented as a tool for addressing inequality, exclusion, and service gaps. Yet the systems themselves are shaped by extractive supply chains, concentrated power, and existing inequities. Drawing on themes associated with books such as AI Snake Oil and Atlas of AI, the authors position AI not as a neutral software layer, but as a socio-technical system built on natural resources, human labor, and entrenched institutions.
That matters because the communities most often targeted by “AI for social good” projects are also the communities most likely to bear the costs of poorly designed interventions. A model that appears efficient from a distance may still fail locally if it ignores language, trust, access, governance, or the human intermediaries required to act on its outputs.
The core question, then, is not whether AI can help. It is what conditions must exist for it to help in a durable and accountable way.
Why institutions matter more than demos
The authors examined eight AI systems deployed to address social problems across the developing world. From the source text available here, the article’s key conclusion is that AI works only when paired with human support and institutional capacity. In practice, that means trained staff, service delivery pipelines, community relationships, and organizations capable of responding to what the technology surfaces.
This is a critical point because many AI deployments are judged by model performance rather than by downstream outcomes. A system may summarize, classify, or predict effectively, yet still fail to improve anyone’s life if no agency can act on the information. In social contexts, the last mile is often the whole story.
Consider what happens when an AI tool identifies need but there is no staffing, funding, or legal authority to respond. The system can still produce dashboards, but the result is administrative theater rather than solved problems. The essay argues that this gap between technical promise and implementation capacity is where many AI-for-good initiatives quietly break down.
Communities are not deployment environments
Another implication of the essay is that underserved communities should not be treated as test beds for generalized tools. Social problems are embedded in local histories, norms, and power structures. Systems that ignore those realities may reproduce exclusion while claiming neutrality.
This is particularly relevant in sectors like agriculture, education, and public service access, where informal intermediaries and on-the-ground relationships often determine whether people can actually use a system. AI can assist those systems, but it rarely replaces them. The labor of translation, explanation, verification, and trust-building remains stubbornly human.
The article’s argument also pushes back on the common idea that a more capable model will naturally close implementation gaps. Better reasoning or stronger language fluency may improve parts of a workflow, but they do not create institutions where none exist. They do not fix broken procurement, fragile local governance, or under-resourced public agencies.
What a more serious AI-for-good agenda would require
If the essay is right, then meaningful AI deployment in social sectors must start with design constraints that many product teams prefer to treat as externalities. Systems need clear accountability structures. They need human operators who can intervene, explain, and contest outputs. They need institutions that can absorb recommendations into real action. And they need enough local grounding to avoid imposing a technical answer on a social problem that has not been properly understood.
This does not make AI irrelevant. On the contrary, it suggests where AI may be most useful: not as a substitute for public systems, but as a component within them. Used that way, AI can reduce clerical burden, widen access to information, support triage, and help frontline workers make better use of limited resources.
But that vision is slower and less glamorous than the pitch that AI can simply route around institutional failure. It requires investment in people as much as software, and in governance as much as models.
The value of the Rest of World essay is that it pulls the debate back to first principles. Social problems persist not because no one has built a sufficiently clever interface, but because durable solutions depend on trust, capacity, and power. AI can assist within that landscape. It cannot wish it away.
As governments, NGOs, and companies continue to adopt AI in public-facing systems, that distinction will become more important. The next phase of AI-for-good will be judged less by what models can generate and more by whether institutions can responsibly use what they generate. That is a harder standard, but it is the one that actually matters.
This article is based on reporting by Rest of World. Read the original article.
Originally published on restofworld.org
