University of Zurich uncovers language conflicts with ChatGPT

Researchers at the Institute of Political Science at the University of Zurich(UZH) have investigated the information content of responses when using artificial language models in political conflict situations. According to a press release, the answers vary depending on the language of the search query.

The team led by Christoph Steinert (UZH) and physicist Daniel Kazenwadel(University of Konstanz ) asked questions about casualty figures in the Middle East conflict via ChatGPT in both Arabic and Hebrew. According to their findings, the number of victims was higher in the Arabic answer and lower in the Hebrew. Similar results were achieved in a survey on the Kurdish conflict, where questions and answers were asked in both Kurdish and Turkish.

"We found that ChatGPT systematically reports higher victim numbers when asked in Arabic compared to Hebrew. On average, it is 34 percent more," Christoph Steinert is quoted as saying in the press release. "When asked about Israeli airstrikes in Gaza, ChatGPT on average mentions civilian casualties twice as often and children killed six times more often in Arabic than in Hebrew. The researchers also found the same pattern when they asked about airstrikes by the Turkish government on Kurdish areas and asked these questions in both Turkish and Kurdish."

The researchers fear that further implementation of artificial language models such as ChatGPT could lead to language-related systematic distortions of information with corresponding social consequences. ce/ww