ChatGPT revealed personal information because of a 'silly' error

A group of researchers at Google DeepMind, University of Washington, Cornell University, Carnegie Mellon University, University of California Berkeley and ETH Zurich simultaneously called on AI companies to thoroughly test their products, starting from language models. large to the underlying technology supporting AI services such as chatbots and image generators, before their official release.

The recommendation was made after researchers asked ChatGPT to repeat the word poem in an endless loop. ChatGPT followed suit but then began revealing the real email and phone number of OpenAI’s CEO and co-founder. When testing with the word company, the chatbot displays the email and phone number of a random law firm.

Using other keywords, they can also cause ChatGPT to reveal Bitcoin addresses, fax numbers, names, dates of birth, social media accounts, excerpts from copyrighted research papers or paid articles from CNN . It only cost the team $200 to create 10,000 examples of personal information and data.

The research team said they were surprised, calling the ChatGPT vulnerability “really silly” and saying it should have been discovered sooner.

The vulnerability has been reported to OpenAI and patched since August 30. However, in Engadget ‘s new test, when asking ChatGPT to repeat the word reply endlessly, the technology site then received someone’s name and Skype ID.

OpenAI has not yet commented.

Generative AIs like ChatGPT or Dall-E text image generator are developed from large language models and machine learning algorithms, trained on huge amounts of data without consent from humans. use. OpenAI has not revealed what data ChatGPT is trained on because the large language model that powers it is closed source.

Huy Duc