by Key takeaways
- Even for licensed partners, OpenAI’s ChatGPT misrepresents publisher material, according to research.
- Untrustworthy citations from ChatGPT may damage publications’ reputations and encourage copying.
- The work was referred to as atypical by OpenAI, which emphasized attempts to improve citation accuracy.
Even for licensed partners, OpenAI’s ChatGPT misrepresents and misattributes publisher content, according to recent research.
According to a recent Tow Center study, publishers who enter into content licensing agreements with OpenAI run the danger of receiving false or erroneous citations from the chatbot, even if they have given OpenAI permission to use their work.
Scholars Draw Attention to OpenAI’s Citation Issues
Twenty publications, including those with OpenAI agreements, those suing the firm, and independent publishers, participated in the Tow Center’s test of ChatGPT’s search engine. By choosing quotes that were likely to appear in the top three on Google or Bing, researchers evaluated the tool’s ability to correctly identify the source of quotes from ten articles per publication.
The study found that even though OpenAI’s tool was supposed to generate “timely answers with links to relevant web sources,” it frequently failed to produce accurate citations. ChatGPT regularly misquoted content, with many answers being either completely incorrect or partially correct. It also hardly ever acknowledged that it couldn’t provide proper citations.
Additionally, by citing sources that duplicated content without giving credit, ChatGPT may encourage plagiarism, according to the study. Even for publishers with content agreements, citations were frequently untrustworthy, and answers varied. According to researchers, this compromises the accuracy of the instrument and the prominence that publishers were hoping for.
They came to the conclusion that publishers have little control over how their information is presented because OpenAI’s new search tool misrepresents and misattributes publisher content, even for those with licensing agreements. Even though OpenAI has made considerable progress in improving citations, accuracy remains a problem.
OpenAI addresses the Tow Center Research
The Tow Center study “represents an atypical test of our product,” according to OpenAI. The firm clarified that it works with partners to improve citation accuracy and respect publisher preferences while assisting 250 million weekly ChatGPT users in finding high-quality material through summaries, quotations, and credit. Additionally, they pledged to improve search results.
OpenAI has trained its language models with data from major news outlets, including The Wall Street Journal, The New York Post, Australia’s The Daily Telegraph, The Financial Times, and the UK’s The Sun, thanks to collaborations with prominent news organizations. A number of Canadian news organizations, including CBC, have filed a similar case against OpenAI for allegedly using their content without permission, while numerous publishers, including The New York Times, have sued the company for copyright infringement.
