A shared characteristic of many of the malicious usages of LLMs is that they seek to conceal that the text was LLM-generated. This is notably the case for LLMs employed for the purpose of manipulating individuals on social media or spreading propaganda and misinformation. Or, on a more mundane level, the use of LLMs by pupils and students to expedite various school assignments.
There is currently no reliable method for detecting whether a piece of text was generated through a large language model instead of a human writer. There are, however, a number of so-called watermarking techniques (see for instance Kuditipudi et al, 2024, Dathathri et al, 2024) that can be applied to leave an "invisible" (to a human reader) mark in the selection of tokens, and therefore provide a way to determine the provenance of the text. Those watermarking techniques are now being integrated in various commercial AI services, most notably Claude and Gemini.
Those watermarks can, however, be diluted, or even entirely removed, through relatively simple attacks, such as paraphrasing the text or translating it into another language. The thesis will analyse the robustness of watermarks against such attacks and work on defense techniques, such as back-translation. A particular focus will be on how watermark strength is affected by the output language, such as Norwegian.