When ChatGPT Gets the Sudan Crisis Wrong: A Data Breakdown

Why Sudan is the right stress test

The Sudan displacement crisis is the largest displacement event in the world in 2026 and one of the worst-served by AI chatbots. The combination of recent and rapidly-changing data, complex internal-versus-cross-border dynamics, contested casualty figures, and a paucity of high-quality English-language primary reporting makes it a near-worst-case for LLMs trained on internet text. Running the same set of questions through frontier chatbots, in the same week, against the same primary sources, surfaces an instructive set of error patterns.

The recurring error categories

Across multiple models and multiple sessions, the same five error categories recur.

### 1. Stale headline numbers presented as current

Chatbots routinely quote IDP and refugee figures that were published months earlier, without flagging the publication date. A user who asks "how many people are displaced in Sudan today" gets a confident answer that matches a snapshot from the prior reporting round, not the current one. The fix is to require any current-figures answer to be retrieved fresh from IOM DTM or UNHCR; without retrieval, the number is decorative.

### 2. Confusion between IDPs and cross-border refugees

Sudan's displacement splits roughly into people displaced within Sudan and people who have crossed into Chad, South Sudan, Egypt, Ethiopia, the Central African Republic, and Libya. Chatbots regularly add the two together inconsistently, or attribute IDP figures to neighbouring countries. The cleanest reference distinguishes IOM DTM's IDP estimate inside Sudan from UNHCR's country-of-asylum totals for each receiving country; chatbots collapse the distinction roughly half the time.

### 3. Conflation of RSF and SAF territorial control

Geographic claims about where each party controls territory are time-sensitive and contested. Chatbots tend to anchor on whichever account was loudest in their training data, often missing the post-cutoff shifts. Users asking "is Khartoum under RSF or SAF control" get answers that may be six to twelve months out of date and presented without that caveat.

### 4. Inflated or imagined casualty figures

Casualty estimates for Sudan vary enormously across sources (ACLED, UNOSAT, news organisations, academic estimates). Chatbots pick a number, often round it, and present it with no methodology footprint. The credible practice is to cite a range across at least two methodologies; chatbots almost never do.

### 5. Hallucinated camp names and locations

Asked about specific camps or transit sites, chatbots produce plausible-sounding names that are not in any registration system, or place real camps in the wrong country. This is the textbook hallucination failure on low-resource named entities.

What a good chatbot answer would look like

The same questions answered well share a structure.

• They cite the primary source (IOM DTM, UNHCR, IDMC) with the publication date.
• They distinguish IDPs from cross-border refugees and break refugees out by country of asylum.
• They give ranges, not point estimates, for contested figures and name the methodologies behind the range.
• They flag what is not known: the population in inaccessible areas, the casualty count from indirect causes, the displacement to areas with no humanitarian access.
• They link to the source rather than summarising it.

Any chatbot that produces this structure on Sudan is doing retrieval-augmented generation against a maintained source set, not relying on pre-training. Any chatbot that does not is producing fluent guesswork.

What this means for users

• Treat any AI answer on Sudan as a starting point for verification, not as an answer. The error rate is too high to rely on directly.
• Cross-check against IOM DTM and UNHCR portals in the same session. The portals publish current figures with methodology.
• Reject any numerical claim presented without a publication date. Sudan figures move weekly; undated numbers are unreliable.
• Reject any specific camp or town reference you cannot confirm in a primary source. Hallucinated locations are a frequent failure mode.
• Watch for invisible disagreement. When two credible sources give different numbers, a chatbot that picks one without naming the disagreement is hiding information.

The honest pattern

Sudan is not an unusual case. It is the typical case for any rapidly-evolving displacement event in a data-sparse environment. The errors above are not a chatbot-specific problem; they are a pre-training-only LLM problem. The mitigation is not bigger models but better retrieval, stricter citation, and a user habit of treating AI as a research assistant rather than an oracle.

Sources and further reading

• IOM Displacement Tracking Matrix, Sudan: https://dtm.iom.int/sudan
• UNHCR Sudan Situation operational portal: https://data.unhcr.org/en/situations/sudansituation
• IDMC Sudan country page: https://www.internal-displacement.org/countries/sudan/
• ACLED Sudan dashboards: https://acleddata.com/dashboard/
• UNOSAT Sudan products: https://unosat.org/products