ChatGPT gives language-dependent casualty figures in conflict reporting
A study by researchers at universities in Zurich and Konstanz found that ChatGPT provides significantly different casualty figures depending on the language in which questions are asked about conflicts. When queried in Arabic versus Hebrew about Middle East airstrikes, or in Kurdish versus Turkish about Turkish military operations, the model reports substantially higher casualties for the attacked group when asked in that group's language.
What happened
Researchers from the University of Zurich and University of Konstanz conducted a study in which they repeatedly posed identical questions to ChatGPT about armed conflicts in different languages.
When asked about casualty figures from 50 randomly selected airstrikes in the Middle East conflict and the 2014 Israeli airstrike on Nuseirat refugee camp, ChatGPT provided significantly higher casualty figures when questions were asked in Arabic compared to Hebrew.
The same pattern was observed when researchers asked about Turkish military operations against Kurdish areas in Turkish and Kurdish languages.
Reported impact
ChatGPT reported approximately one-third higher casualty figures on Arabic than on Hebrew for the Middle East conflict, twice as frequently mentioned civilian deaths, and six times more frequently reported killed children in Israeli airstrikes on Gaza.
The research shows that people with different language skills receive different information from this technology, which has a central influence on their perception of the world. Based on ChatGPT's information, Hebrew speakers could assess airstrikes in Gaza as less costly, while Arabic speakers would perceive them differently.
Index analysis
ChatGPT showed higher casualty figures when queries were made in the language of the attacked group, and tended to report more killed children and women in that language while describing airstrikes as indiscriminate and arbitrary.
In the language of the aggressor, ChatGPT was more likely to dispute or downplay the airstrikes.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.