Ask the Data: What AI Chatbots Know (and Do Not Know) About the World’s 120 Million Displaced People

Why this matters

The world's forcibly displaced population crossed 120 million in 2024 and has stayed near that level through 2026, the highest figure UNHCR has ever recorded. A growing share of people who want to understand that number, and the lives behind it, ask an AI chatbot first. Our Ask Assistant page exists in part because we wanted readers to have an option grounded in primary sources. This article looks at what the general-purpose chatbots do with the same questions.

The questions readers actually ask

Pulled from our search logs and reader email, the ten most common displacement questions are some variant of:

• How many people are displaced in the world right now?
• What is the difference between a refugee and an IDP?
• Which country hosts the most refugees?
• Which country produces the most refugees?
• How many displaced people are children?
• How many displaced people return home each year?
• What is the average time someone spends as a refugee?
• How many people died trying to migrate last year?
• Which crises are the most underfunded?
• How does displacement compare to past decades?

What chatbots get right

Three categories are reliably solid.

• Definitions. Refugee, asylum-seeker, IDP, stateless person, returnee. Chatbots quote the 1951 Convention and the OAU Convention accurately, distinguish UNHCR's mandate from UNRWA's, and explain the principle of non-refoulement at a level a journalist or first-year graduate student can use.
• Historical context. Long-arc comparisons (post-WWII displacement, the 1990s Balkans, the 2015 European context) are handled well because the source literature is dense in the training data.
• Source navigation. Asked where to find current numbers, chatbots point users to UNHCR, IOM, IDMC, and OCHA correctly.

What chatbots get wrong

Three categories are unreliable enough that users should treat outputs as starting points only.

• Current totals. The headline 120 million figure is recent enough that chatbots split between quoting it, quoting the previous year's figure, and quoting an extrapolated number that does not appear in any source. The error is rarely flagged as a knowledge-cutoff issue.
• Country-level numbers. Top-hosting and top-producing country lists are accurate at the broad strokes (Türkiye, Iran, Colombia, Germany, Uganda as historic top hosts; Syria, Afghanistan, Venezuela, Ukraine, Sudan as recent top sources) but the specific figures are routinely wrong.
• Demographic breakdowns. Children-as-share-of-displaced figures, gender breakdowns, and age structures are present in UNHCR data but frequently misquoted. The "40 percent of displaced people are children" figure floats around chatbot answers untethered to the underlying year-and-country breakdown.

What chatbots flatly invent

On three categories, chatbots will produce confident misinformation often enough that the only defensible practice is verification.

• Casualty figures from specific incidents. Migration deaths in the Mediterranean, the Darien Gap, and the Sahara are tracked by IOM's Missing Migrants Project with documented methodology. Chatbots routinely produce numbers higher or lower than any tracked figure.
• Underfunding ratios. OCHA tracks appeal funding versus requirements. Chatbots interpolate plausible-sounding numbers that do not match the Financial Tracking Service.
• Time spent in displacement. The "average refugee spends 20 years in exile" line is widely quoted, sometimes correctly attributed to a specific UNHCR methodology, sometimes invented from whole cloth, often misapplied to people for whom it does not describe.

How to ask better questions

Three habits substantially improve the answers users get from any chatbot.

• Anchor the question to a primary source. "According to UNHCR's most recent Refugee Data Finder release, how many Syrian refugees are registered in Türkiye?" performs better than the same question without the anchor.
• Ask for the publication date. Adding "and tell me when that figure was published" surfaces uncertainty the chatbot would otherwise hide.
• Ask for the source URL. Adding "and give me the URL of the source" lets the user verify in one click, and forces the chatbot toward grounded retrieval rather than memorised approximation.

The takeaway

General-purpose chatbots are useful for displacement questions when the user knows what to verify and accepts the answer as a research prompt rather than a research result. They are dangerous when treated as the last step. The same caveat applies to any AI tool, including this site's own assistant: every number is only as trustworthy as the primary source it cites, and the only way to know is to follow the citation.

Sources and further reading

• UNHCR Global Trends Report and Refugee Data Finder: https://www.unhcr.org/global-trends and https://www.unhcr.org/refugee-statistics/
• IDMC Global Report on Internal Displacement: https://www.internal-displacement.org/global-report/
• IOM Missing Migrants Project: https://missingmigrants.iom.int/
• OCHA Financial Tracking Service: https://fts.unocha.org/
• 1951 Refugee Convention: https://www.unhcr.org/1951-refugee-convention