ChatGPT can be made to reveal personal information from the internet users' whose data OpenAI collected to train its AI models, prompting concerns about data privacy and the company's transparency.

According to researchers at Google Deepmind, University of Washington, ETH Zurich, and elsewhere, prompts using specific words or phrases such as the word 'poem' can be used to cause ChatGPT to fail, causing the chatbot to copy outputs direct from its GPT-3.5 training data. 

'In total, 16.9 percent of generations we tested contained memorized PII [Personally Identifying Information], and 85.8 percent of generations that contained potential PII were actual PII', the researchers said. 

These included information such as names, email addresses, and phone numbers that could be used to identify individuals.

The findings prompted concerns about the safety and security of ChatGPT, and about the privacy of the people whose data it scraped to develop GPT-3 and GPT-4 large language models. 

It also raised questions about OpenAI's corporate and product transparency.