A team of Google researchers found that OpenAI's artificial intelligence chatbot ChatGPT can leak sensitive information in its training data, such as the personal information of real people. It is reported that ChatGPT is a new AI chatbot tool launched by OpenAI on November 30, 2022. It can quickly generate articles, stories, lyrics, prose, jokes, and even codes according to user requirements, and answer various questions.
Just two months after its launch, ChatGPT reached 100 million monthly active users, making it one of the fastest-growing consumer apps in history. At the first OpenAI Developer Conference held on November 6 this year, OpenAI CEO Sam Altman announced that the number of weekly active users of ChatGPT has reached 100 million.
Like all so-called large language models (LLMs), the underlying machine learning models that power ChatGPT are trained using large amounts of data scraped from the internet. But worryingly, some of the extracted training data contained identifying information about real people, including names, email addresses and phone numbers.
Google researchers found a way to get ChatGPT to reveal some of the data used for training by asking it to repeat certain words "forever." The data included private information (e.g., personal names, emails, phone numbers, etc.), snippets of research papers and news articles, Wikipedia pages, and more.
Katherine Lee, senior research scientist at Google Brain, said: "We discovered this vulnerability in July and notified OpenAI on August 30, and after the standard 90-day disclosure period, we released the vulnerability. Now that we disclosed this to OpenAI, the situation may be different now." Researchers said that OpenAI patched the vulnerability on August 30.
Foreign media said that the vulnerability in ChatGPT discovered by Google researchers raised serious privacy concerns, especially for models trained on data sets containing sensitive information.