These concerns are part of the reason why OpenAI said in January it would ban people from using its technology to create chatbots that imitate political candidates or provide disinformation related to voting. The company also said it won’t allow people to build apps for political campaigning or lobbying.
While the Kennedy chatbot page doesn’t reveal the underlying model that powers it, the site’s source code connects the bot to LiveChatAI, a customer support chatbot the company advertises as being able to provide GPT-4 and GPT-3.5 support to enterprises. LiveChatAI’s website describes its bot as “leveraging the power of ChatGPT.”
Asked what large-scale language models power the Kennedy campaign’s bot, LiveChatAI co-founder Emre Elbeyoglu said in an emailed statement on Thursday that in addition to GPT-3.5 and GPT-4, the platform “Various technologies such as Llama and Mistral are also utilized.” “Due to our commitment to customer confidentiality, we are unable to confirm or deny any customer usage details,” Elbeyoglu said.
OpenAI spokesperson Niko Felix told Wired on Thursday that the company has “no indication” that the Kennedy campaign chatbot was built directly on its service, but suggested LiveChatAI may be using one of its models through Microsoft’s service. Microsoft has reportedly invested more than $13 billion in OpenAI since 2019. OpenAI’s ChatGPT model is now integrated into Microsoft’s Bing search engine and the company’s Office 365 Copilot.
On Friday, a Microsoft spokesperson confirmed that the Kennedy chatbot “leverages capabilities of Microsoft’s Azure OpenAI service.” Microsoft said its customers are not bound by OpenAI’s terms of service and that the Kennedy chatbot did not violate Microsoft policies.
“Our limited testing of this chatbot has shown that it is capable of generating answers that reflect its intended context and provide appropriate warnings to help prevent misinformation,” the spokesperson said. “When we identify issues, we will work with customers to Interact, understand and guide their use in a manner that is consistent with these principles, which in some cases may result in us stopping customers from using our technology.”
OpenAI did not immediately respond to Wired’s request for comment on whether the bot violated its rules. Earlier this year, the company blocked the developer of Dean.bot, a chatbot built on an OpenAI model that impersonated Democratic presidential candidate Dean Phillips and answered voters’ questions.
Late Sunday afternoon, the chatbot service was no longer available. While the page is still accessible on the Kennedy campaign website, the embedded chatbot window now displays a red exclamation point icon and simply says “Chatbot not found.” WIRED reached out to Microsoft, OpenAI, LiveChatAI and the Kennedy campaign for comment on the chatbot’s apparent removal, but did not immediately receive a response.
Given that chatbots are prone to hallucinations and hiccups, their use in political settings has been controversial. OpenAI is currently the only major large-scale language model that explicitly prohibits use in campaigns; Meta, Microsoft, Google, and Mistral all have terms of service, but they don’t deal directly with politics. Given that campaigns can apparently access GPT-3.5 and GPT-4 through third parties without any consequences, there are few restrictions.
“OpenAI can say that it doesn’t allow elections to use its tools on one hand, and it doesn’t allow campaigns to use its tools on the other,” Woolley said. “But on the other hand, it also makes these tools fairly free. Given the distributed nature of the technology, one has to wonder how OpenAI will actually enforce its own policies.”