Artificial intelligence chatbot ChatGPT, designed by AI research firm OpenAI, is able to accurately state that the CEO of Twitter is Elon Musk — despite acknowledging that the materials it was trained on end in 2021. Musk became Twitter’s CEO at the end of 2022.
The issue was first noticed by tech journalist Alex Kantrowitz.
A representative for OpenAI later told Semafor that it’s possible the bot was learning new details during its regular training updates with researchers who teach it how to behave during conversations.
ChatGPT provided the same response to Semafor when prompted to answer who was the current CEO of Twitter.
“As of my knowledge cutoff in 2021, the CEO of Twitter is Elon Musk,” the bot said. Semafor asked for clarification and the AI apologized, saying, “My information on this matter is incorrect. My knowledge cutoff was in 2021 and I do not have the most up to date information.”
It then correctly noted that both Jack Dorsey and Parag Agrawal served as Twitter CEO in 2021.
When asked why ChatGPT responded the way it did, the AI said its “response was generated based on patterns and information that was present in the training data at the time of my training,” adding it doesn’t have the ability to stay up to date with current events.
But the bot is aware of at least one other recent news event: When prompted, ChatGPT was able to tell Semafor that Queen Elizabeth II had died, and that King Charles III had ascended the throne (though it identified the new monarch by his former title, Prince Charles).
However, ChatGPT identified the current president of Brazil as Jair Bolsonaro, which was correct in 2021. He was defeated by Luiz Inácio Lula da Silva in 2022.
Kantrowitz, who writes the newsletter Big Technology, told Semafor it’s “remarkable” that ChatGPT lied about where it got its information from.
Knowing Musk is Twitter’s CEO isn’t a detail that came “from nowhere,” he said. “I think that … whether OpenAI makes a statement about it or not, I think the bot should be more truthful about how it knows what it knows.”
Kantrowitz suggested that while it’s possible OpenAI trained the chatbot with new information, it could have learned new details from its conversations with users.
That, he said, is “potentially a pretty dangerous area,” he said, pointing to Microsoft’s teenage chatbot “Tay,” which learned Nazi rhetoric and began denying the Holocaust in a matter of hours after its launch.
However, a representative for OpenAI told Semafor that ChatGPT is not learning information from individual conversations with its users. The representative said that in addition to the large data set ChatGPT was trained on, which included materials up to 2021, the AI is also receiving regular training updates from researchers that teach it how to behave during conversations.
It is possible that the bot learned these details in its research sessions, and is giving more current answers as a result, the representative said.