Historical and Political Events from the Perspective of AI : A Multilingual Comparison of Large Language Models in Historical Question Answering

Sammanfattning

With the introduction of Large Language Models (LLM) to the public, their performance in various tasks has been an important topic for research in the field of Artificial Intelligence. LLMs are able to execute tasks ranging from simple text generation, to image generation and Question Answering. Their ability to use multiple languages allows the users to interact with the models in various languages, making it accessible for people who are not able to speak English while also providing access to information of multiple origins for bilingual users. However, many studies have shown that many LLMs are not as accurate when responding in a language other than English. Simultaneously, due to the vast amount of data used to train each model, differences in information retrieval can be seen between the models. By using questions of historical context, this Thesis aims to identify differences between the multilingual responses of each model individually, differences between the information provided by comparing the models responses while simultaneously checking for any possible censorship. The results of this thesis showed that the multilingual responses of the three models, ChatGPT, Gemini and DeepSeek did in fact show differences when alternating between languages but also showed differences with how they presented the information regarding the historical events the questions referred to. Lastly, the only model that provided censored responses was DeepSeek, which did not respond properly for two questions that featured sensitive historical topics.