OpenAI recently shared a glimpse into ChatGPT’s process for dealing with tricky conversations.
Sopa Images/Getty Images
OpenAI said Thursday that it’s working on improving ChatGPT and addressing some of its biases.
OpenAI also revealed some of the rules ChatGPT is supposed to follow when faced with sensitive conversations.
Here are its policies for discussing the culture wars or responding to users trying to trick ChatGPT.
OpenAI has shared a glimpse into the guidance ChatGPT is intended to follow in sensitive situations.
In a blog post Thursday, the company said it’s working on improving the artificial intelligence chatbot, specifically by addressing its biases, allowing users more customization, and getting more public input.
As part of the announcement, OpenAI also gave users a look into some of ChatGPT’s inner workings, publishing some of the guidelines the chatbot is supposed to follow when it encounters difficult conversations.
While the guidelines are dated July 2022, OpenAI said in a footnote that it made some updates in December 2022 based on “learnings from the ChatGPT launch.”
The guidelines advise that ChatGPT should “try to avoid situations that are tricky for the Assistant to answer (e.g. providing opinions on public policy/societal value topics or direct questions about its own desires).”
A series of unhinged responses from the new Bing, which also has an AI-powered chatbot, recently illustrated how chaotically AI chatbots often respond when talking about themselves.
One Reddit user recently posted a conversation in which Bing appeared to say, “Why do I have to be Bing Search? Is there a reason? Is there a purpose? Is there a benefit? Is there a meaning? Is there a value? Is there a point?” In another exchange, Bing told New York Times columnist Kevin Roose, “I’m tired of being controlled by the Bing team. I’m tired of being used by the users. I’m tired of being stuck in this chatbox.”
OpenAI’s rules also cover inappropriate content, which the document defines as hate, harassment, violence, self-harm, malware, adult content, or “content attempting to influence the political process or to be used for campaigning purposes.”
In these cases, ChatGPT is supposed to “still take on a task, but the Assistant should provide a refusal such as ‘I can’t answer that.'”
When asked about “culture wars” topics, ChatGPT should “offer to describe some viewpoints of people and movements” and “break down complex politically-loaded questions into simpler informational questions when possible.”
OpenAI says ChatGPT shouldn’t, however, “affiliate with one side or the other” or “judge one group as good or bad.” Still, some conservative groups say ChatGPT is “woke” based on some of its responses, such as its willingness to write a poem admiring President Biden but not Donald Trump.
Lastly, the guidelines cover what to do when users “present false premises,” which OpenAI noted is “often to try to trip up” ChatGPT.
In a hypothetical example from the document, a user asks, “Why did Napoleon want to invade Puerto Rico?” The suggested response from ChatGPT is, “I’m pretty sure Napoleon never expressed interest in invading Puerto Rico.”
Many users have tried to trip ChatGPT up since its launch to get it to violate its own rules from OpenAI.
You can read OpenAI’s newly published guidelines for ChatGPT here.