The money flowing into artificial intelligence is now measured in figures that sound like national budgets. Data centres, model training runs, valuations that rewrite the definition of a large company overnight. Against that backdrop, a quieter request landed this summer, and it deserves more attention than it got: the International Rescue Committee asked the technology industry to point some of that formidable capability at the people who most rarely appear in a product roadmap. Not a new frontier model. Just the existing one, aimed somewhere useful.
It is a reminder that the technology’s biggest gaps are not only technical. They are gaps of intention, of who gets designed for, and of who gets asked what they actually need.
The call
According to ReliefWeb, the International Rescue Committee urged the tech industry to put AI to work for the world’s roughly 118 million displaced people, directing capability that already exists toward concrete humanitarian needs: translation, case management, and crisis response. The framing matters. This is not a plea for a moonshot or a bespoke ‘humanitarian AI’. It is a request to redirect tools that are already shipping to enterprise customers and consumers, and to treat displaced populations as a legitimate user base rather than an afterthought.
That distinction is the strategic heart of the argument. The industry has spent enormous energy chasing marginal gains for well-served users, while the marginal gain of a working translation tool for a family that has just crossed a border can be the difference between finding shelter and sleeping outside. The capability is not the bottleneck. The allocation of attention is.
There is also a hard-headed case buried in the appeal. Displacement is not shrinking. Conflict, climate shocks, and economic collapse keep the number stubbornly high, and humanitarian agencies operate under permanent funding strain. If AI can extend the reach of overstretched caseworkers, it is not charity, it is capacity, and the sector needs capacity more than it needs another pilot that never scales.
Where AI can genuinely help
Strip away the hype and the useful applications are unglamorous, which is usually a good sign. Three stand out.
Translation and access to services. Displaced people frequently arrive somewhere they do not speak the language, needing to navigate registration, healthcare, legal status, and schooling within days. Real-time and asynchronous translation, especially across the low-resource languages that commercial products have historically ignored, can collapse the distance between a person and the service they are entitled to. This is one place where large language models genuinely outperform the previous generation of tools.
Case management and crisis logistics. Aid workers spend a punishing share of their time on documentation, triage, and coordination. AI that drafts case notes, summarises long intake histories, flags urgent needs, or helps route supplies through a chaotic response can free humans for the parts of the job that require human judgement. The value here is not replacing caseworkers, it is giving each one more hours in the day.
Information in low-connectivity settings. Much of the value of AI assumes a fast connection and a modern device. Displacement rarely offers either. The genuinely hard, genuinely worthwhile engineering problem is delivering reliable information, on rights, on services, on safety, in settings with intermittent power and thin bandwidth. On-device models, offline-capable tools, and lightweight interfaces are less exciting than the frontier, but they are where impact actually lives.
The cautions
None of this is safe by default, and it would be dishonest to pretend otherwise. Serving vulnerable populations raises the stakes on every ethical question the industry has already fumbled with better-resourced users.
Industry analysis published alongside the IRC’s call is blunt about the guardrails: beneficial use requires careful attention to privacy, consent, and dignity. These are not compliance checkboxes. A displaced person is, almost by definition, someone who has lost control over their circumstances. Building a tool that harvests their data, tracks their movements, or nudges their behaviour without meaningful consent does not help them, it deepens the asymmetry that put them at risk.
The sharpest line to draw is between assistance and surveillance. The same capabilities that enable case management, biometric records, movement tracking, sentiment analysis, can be turned into instruments of control, whether by hostile states, by border regimes, or simply by well-meaning agencies that collect more than they can protect. Extractive uses, where the vulnerable become a training resource or a data source rather than the intended beneficiary, must be treated as a red line, not a grey area.
The corrective is not a longer terms-of-service document. It is a change in method: designing with affected communities, not merely for them. That means involving displaced people in defining the problem, testing the tool, and holding the organisation accountable, with a real ability to say no. It is slower and less scalable than shipping a product and measuring adoption. It is also the only version of ‘AI for good’ that earns the phrase.
The India read
For an Indian audience, the debate is not abstract, and it is not only about refugees. India runs some of the world’s largest humanitarian and inclusion challenges inside its own borders: disaster response across flood and cyclone-prone regions, migrant labour that moves across states with little continuity of documentation, and vast populations underserved by formal services. The ‘displaced’ framing extends naturally to the internally displaced and the digitally excluded.
India also happens to be unusually well-placed to build the specific kind of AI this moment calls for. The country’s central AI problem, serving hundreds of millions of people across dozens of languages, many of them low-resource, is precisely the problem humanitarian applications need solved. Vernacular AI is not a nice-to-have here, it is the mainstream use-case. The models, datasets, and speech tools being built for Indian languages are, almost as a by-product, exactly the infrastructure that translation and access-to-services applications for the vulnerable depend on.
That creates an opportunity and a responsibility. The opportunity is that responsible AI for underserved populations can be a genuine Indian strength, exported rather than imported. The responsibility is that the same cautions apply with force. India’s own debates over data protection, consent, and the use of digital identity in welfare delivery show how quickly inclusion tooling can drift toward exclusion when a system fails and there is no human recourse. Designing with communities, building offline-first, and refusing surveillance creep are not Western imports, they are prerequisites for anything that claims to serve people who cannot afford for the technology to fail them.
The IRC’s request is modest in one sense, use what you already have, and radical in another, decide that these users count. The frontier will keep getting the capital and the headlines. Whether any of that capability reaches the 118 million is a choice, not a technical limitation, and choices are the kind of thing a whole industry can still be argued out of, or into.
