AI for Good. Good for Whom?

Why migration is AI’s ultimate stress test.

Written by Florence Kim, Executive Director, AI.dvocacy and former Head of communications and advocacy at IOM/UN Network on Migration

Frank Laczko, Senior Adviser, Multicultural Insights, former Director IOM GMDAC

Artificial intelligence is no longer simply answering questions. It is shaping the questions we ask, the answers we get and the decisions we make. Through agentic systems, it can now act on our behalf. Every day, it decides what information is retrieved, summarized, translated, generated and recommended. It is becoming humanity’s memory.

But whose memory?

A few days ago, when 193 UN member states met in Geneva for the first-ever Global Dialogue on AI Governance, deepfakes were named. Autonomous weapons were named. Children’s safety was named. Migration was not. Major discussions repeatedly referred to inclusion, while migration itself remained largely absent.

That gap is the wager of this piece: if AI cannot work fairly in migration, it is unlikely to work fairly anywhere. Migration concentrates, in one policy area, nearly everything AI struggles hardest to get right.

As AI increasingly shapes how billions of people understand the world, whether the range of human experience it reflects actually resembles the range that exists becomes just as important as intelligence. Call that “representativeness.” Nowhere is the gap more visible than in migration, where the people most affected by AI’s decisions are often the ones least represented in the data that trains it.

The question is no longer simply whether AI is intelligent or biased. It is whether it represents humanity well enough to deserve our trust.

The Myth of Human-Centred AI

The AI conversation often assumes that AI should serve humanity. That phrase has become the default language of UN and policy circles. Who said AI’s purpose was ever to serve humanity? It’s a fine aspiration, and one worth holding AI to. The trouble starts when we assume it’s already true, rather than something to be built, tested and enforced.

“Humanity” is not the primary optimization function of those building or using it.

Technology companies optimize for engagement, compute efficiency and margin. Governments optimize for competitiveness, security and, sometimes, surveillance. Investors optimize for returns. Users optimize for convenience, speed and personalization.

None of these actors optimize to ensure AI reflects reality. But maybe someone should. Maybe we should.

So before asking whether AI is “human-centred,” “fair,” or an engine for good, ask a blunter question: who is AI built for and who is it built by?

Probability vs. Representativeness

Foundation models — the large, general-purpose systems, like GPT, Gemini or Claude, that most AI tools are now built on top of — don’t ask whether something reflects the real world. They ask whether it’s statistically probable.

The internet doesn’t reward accuracy or breadth. It rewards volume, visibility and digitization. The result: the largest languages dominate, the most-documented cultures dominate, the most-digitized countries dominate and the most-repeated narratives become the default account of reality.

The UN’s own Independent International Scientific Panel on AI — a body of 40 members from 37 countries — reaches a similar conclusion. It documents deep language exclusion and a dangerous concentration of technological power. What it barely mentions is migration at all.

AI is not simply built on discrimination; it perpetuates it; it normalizes it.

The Missing Arrow

That normalization has a particular shape when it comes to migration. Most discussions of AI and migration stop at one arrow:

AI → Migration

But the relationship also runs in the opposite direction:

Migration → AI

Migration produces multilingual datasets. Diasporas generate knowledge across cultures. Cross-border communities reshape languages that later become training data for foundation models.

Much of the world’s data-labeling and content-moderation workforce is migrant labor, annotating the very models that will later misrepresent them.

Migrants do not simply use AI. They help build it: as engineers and researchers (a striking share of the field’s top talent is foreign-born), as the diaspora communities whose everyday cross-border exchange keeps reshaping what the internet treats as normal usage, and as the largely invisible workforce that labels and moderates the data those same models are trained on.

Human mobility isn’t something AI merely encounters in its outputs. It runs through AI’s workforce, its inputs and its language, at every level.

Migration Makes AI Real

Around the world, generative AI is entering migration systems directly. Chatbots now field asylum-seekers’ questions in place of overstretched case officers. Large language models draft and summarize case files. AI-generated risk scores flag which visa applications need “further review.” Automated credibility-assessment tools scan interview transcripts for the kind of inconsistencies that get someone’s account disbelieved. This is not digitization. A paper file becoming a PDF is digitization. A model generating a judgment about whether a person’s account is credible is something else entirely.

Used responsibly, these systems can improve efficiency. Used poorly, they can magnify inequality. Migrants often face language barriers, weaker legal protection and fewer opportunities to challenge automated decisions.

When AI discriminates — or simply gets it wrong — who is accountable: the government, the technology company, the engineer or the official?

Migration Is AI’s Toughest Stress Test

Migration concentrates almost every challenge foundation models struggle with in a single policy area: multiple languages, fragmented identities, informal and oral knowledge, legal complexity, limited digital infrastructure and restricted opportunities to challenge decisions. This is what a stress test looks like — few other policy domains ask a single system to get all of this right at once.

That matters because migration is not an exception. It is an early warning system. A foundation model that consistently overlooks minority languages, under-represented communities or informal knowledge does not merely fail migrants — it produces an increasingly incomplete picture of humanity.

Migration is therefore where we can observe, earlier than anywhere else, how foundation models construct reality, whose knowledge they privilege and whose experiences they erase. It is the proving ground where the strengths — and weaknesses — of AI governance become visible.

The Real Blind Spot Is Governance

The UN Global Digital Compact — the 2024 agreement setting shared principles for global digital cooperation — says little about migration. The Global Compact for Safe, Orderly and Regular Migration understandably predates today’s AI revolution while not saying much about technologies either. As already noted, the Scientific Panel highlights language exclusion and technological concentration, yet migration remains largely absent from its findings too.

Two agendas, two vocabularies, one shared blind spot.

Meanwhile, governance still regulates algorithms while society has already moved to foundation models, AI assistants and agentic systems — AI that doesn’t just answer, but acts, carrying out multi-step tasks on a person’s behalf, the way this piece opened by noting.

Beyond Another Global Declaration

High-level principles are not enough. Governments need practical guidance. A rights-based approach requires one coherent framework, not disconnected agendas. The migration benchmark outlined above already points the way forward.

1. Governments remain accountable (Human oversight). Responsibility for AI-assisted decisions cannot be delegated to algorithms or technology providers.

2. People must be able to challenge AI-assisted decisions (Transparency). Individuals should know when AI materially influences a decision and be able to request meaningful human review.

3. Audit for representativeness — not just performance (Non-discrimination). Independent audits should assess multilingual reasoning, representativeness and disparate impacts — not only technical accuracy.

4. Sensitive data deserves heightened protection (Privacy). Personal and biometric data should be subject to stronger safeguards throughout their lifecycle.

5. Bridge migration governance and AI governance (International standards aligning the GCM and GDC). The GCM and GDC should evolve together instead of in parallel.

None of this requires a new treaty. It requires the treaties that already exist to start talking to each other.

Conclusion

AI governance is often framed as a technical challenge. It is, above all, a question of whose experiences become part of humanity’s shared digital memory. Whether AI serves everyone will depend on whether the world’s most mobile, multilingual and often least visible communities are included in the knowledge it learns from. Migration is where that choice becomes impossible to ignore.

In the end, AI will be judged not only by what it can do, but by whose humanity it remembers.