The extent of plagiarism within ChatGPT and similar language model that can generate content based on a few simple prompts are worried by the researchers. Became more common after the tool went public, instances of plagiarism within universities and educational institutions.
More than one ways a research team from Penn University states that language models like ChatGPT plagiarise content. ” In different flavours Plagiarism comes” said Dongwon Lee, professor of information sciences and technology at Penn State. If language models not only copy and paste but resort to more sophisticated forms of plagiarism without realising it we wanted to see.
Within language models Studying plagiarism
Three forms of plagiarism – verbatim; paraphrasing; and idea the researchers attempted to identify. The content that was directly lifted, paraphrasing corresponds to rewording content without citing the original source, and idea refers to using the key thought of a research without the right attribution the first kind includes.
To detect plagiarism and run it on OpenAI’s GPT-2 model, allowing researchers to compare AI-generated text to 8 million documents that were used to pre-train GPT-2 they created a pipeline to automatically.
For plagiarism – divided by specific topic areas that are scientific documents, scholarly articles about Covid-19, and patent claims about 210,000 generated texts were tested.
The research team retrieved top 10 training documents that were similar to AI-generated text, using an open-source search engine. All three types of plagiarism and this rate was higher with larger datasets they found that GPT-2 committed.
Verbatim plagiarism, instances of paraphrasing and idea plagiarism were still up while language models may have managed to keep down. That language models often exposed individuals’ private information through all three forms of plagiarism the researchers also found.
ACM Web Conference that is set to take place in Austin, Texas findings from the study will be presented at the 2023. Its process may be applied to newer language models like ChatGPT, even though the study only takes into account GPT-2.
